Operators.md

Operator Schemas

This file is automatically generated from the def files via this script. Do not modify directly and instead edit operator definitions.

• ai.onnx (default)
  • Abs
  • Acos
  • Acosh
  • Add
  • And
  • ArgMax
  • ArgMin
  • Asin
  • Asinh
  • Atan
  • Atanh
  • AveragePool
  • BatchNormalization
  • Cast
  • Ceil
  • Clip
  • Compress
  • Concat
  • Constant
  • ConstantLike
  • Conv
  • ConvTranspose
  • Cos
  • Cosh
  • DepthToSpace
  • Div
  • Dropout
  • Elu
  • Equal
  • Erf
  • Exp
  • Expand
  • EyeLike
  • Flatten
  • Floor
  • GRU
  • Gather
  • Gemm
  • GlobalAveragePool
  • GlobalLpPool
  • GlobalMaxPool
  • Greater
  • HardSigmoid
  • Hardmax
  • Identity
  • If
  • InstanceNormalization
  • IsNaN
  • LRN
  • LSTM
  • LeakyRelu
  • Less
  • Log
  • LogSoftmax
  • Loop
  • LpNormalization
  • LpPool
  • MatMul
  • Max
  • MaxPool
  • MaxRoiPool
  • MaxUnpool
  • Mean
  • Min
  • Mul
  • Multinomial
  • Neg
  • Not
  • OneHot
  • Or
  • PRelu
  • Pad
  • Pow
  • RNN
  • RandomNormal
  • RandomNormalLike
  • RandomUniform
  • RandomUniformLike
  • Reciprocal
  • ReduceL1
  • ReduceL2
  • ReduceLogSum
  • ReduceLogSumExp
  • ReduceMax
  • ReduceMean
  • ReduceMin
  • ReduceProd
  • ReduceSum
  • ReduceSumSquare
  • Relu
  • Reshape
  • Scan
  • Scatter
  • Selu
  • Shape
  • Sigmoid
  • Sign
  • Sin
  • Sinh
  • Size
  • Slice
  • Softmax
  • Softplus
  • Softsign
  • SpaceToDepth
  • Split
  • Sqrt
  • Squeeze
  • Sub
  • Sum
  • Tan
  • Tanh
  • Tile
  • TopK
  • Transpose
  • Unsqueeze
  • Upsample
  • Xor
  • _experimental ATen
  • _experimental Affine
  • _experimental ConstantFill
  • _experimental Crop
  • _experimental DynamicSlice
  • _experimental GRUUnit
  • _experimental GivenTensorFill
  • _experimental ImageScaler
  • _experimental ParametricSoftplus
  • _experimental Scale
  • _experimental ScaledTanh
  • _experimental ThresholdedRelu

ai.onnx (default)

Abs

Absolute takes one input data (Tensor) and produces one output data (Tensor) where the absolute is, y = abs(x), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Abs-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to all numeric tensors.

Examples

abs

node = onnx.helper.make_node(
    'Abs',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.abs(x)

expect(node, inputs=[x], outputs=[y],
       name='test_abs')

Acos

Calculates the arccosine (inverse of cosine) of the given input tensor, element-wise.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The arccosine of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

acos

node = onnx.helper.make_node(
    'Acos',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-0.5, 0, 0.5]).astype(np.float32)
y = np.arccos(x)
expect(node, inputs=[x], outputs=[y],
       name='test_acos_example')

x = np.random.rand(3, 4, 5).astype(np.float32)
y = np.arccos(x)
expect(node, inputs=[x], outputs=[y],
       name='test_acos')

Acosh

Calculates the hyperbolic arccosine of the given input tensor element-wise.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The hyperbolic arccosine values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

acosh

node = onnx.helper.make_node(
    'Acosh',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([10, np.e, 1]).astype(np.float32)
y = np.arccosh(x)  # expected output [2.99322295,  1.65745449,  0.]
expect(node, inputs=[x], outputs=[y],
       name='test_acosh_example')

x = np.random.uniform(1.0, 10.0, (3, 4, 5)).astype(np.float32)
y = np.arccosh(x)
expect(node, inputs=[x], outputs=[y],
       name='test_acosh')

Add

Performs element-wise binary addition (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Add-1, Add-6

Inputs

A : T

First operand.

B : T

Second operand.

Outputs

C : T

Result, has same element type as two inputs

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

add

node = onnx.helper.make_node(
    'Add',
    inputs=['x', 'y'],
    outputs=['sum'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
expect(node, inputs=[x, y], outputs=[x + y],
       name='test_add')

add_broadcast

node = onnx.helper.make_node(
    'Add',
    inputs=['x', 'y'],
    outputs=['sum'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
expect(node, inputs=[x, y], outputs=[x + y],
       name='test_add_bcast')

And

Returns the tensor resulted from performing the and logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: And-1

Inputs

A : T

First input operand for the logical operator.

B : T

Second input operand for the logical operator.

Outputs

C : T1

Result tensor.

Type Constraints

T : tensor(bool)

Constrains input to boolean tensor.

T1 : tensor(bool)

Constrains output to boolean tensor.

Examples

and

node = onnx.helper.make_node(
    'And',
    inputs=['x', 'y'],
    outputs=['and'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
y = (np.random.randn(3, 4) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and4d')

and_broadcast

node = onnx.helper.make_node(
    'And',
    inputs=['x', 'y'],
    outputs=['and'],
)

# 3d vs 1d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(5) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and_bcast3v1d')

# 3d vs 2d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(4, 5) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and_bcast3v2d')

# 4d vs 2d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and_bcast4v2d')

# 4d vs 3d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(4, 5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and_bcast4v3d')

# 4d vs 4d
x = (np.random.randn(1, 4, 1, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 1, 5, 6) > 0).astype(np.bool)
z = np.logical_and(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_and_bcast4v4d')

ArgMax

Computes the indices of the max elements of the input tensor's element along the provided axis. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned. The type of the output tensor is integer.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is 0)

The axis in which to compute the arg indices.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : tensor(int64)

Reduced output tensor with integer data type.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to all numeric tensors.

Examples

default_axes_keepdims

data = np.array([[2, 1], [3, 10]], dtype=np.float32)
keepdims = 1
node = onnx.helper.make_node(
    'ArgMax',
    inputs=['data'],
    outputs=['result'],
    keepdims=keepdims)

# result: [[1], [1]]
result = argmax_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_default_axis_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [1, 3, 4]
result = argmax_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_default_axis_random')

keepdims

data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 1
node = onnx.helper.make_node(
    'ArgMax',
    inputs=['data'],
    outputs=['result'],
    axis=axis,
    keepdims=keepdims)
# result: [[0], [1]]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 1, 4]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_keepdims_random')

no_keepdims

data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 0
node = onnx.helper.make_node(
    'ArgMax',
    inputs=['data'],
    outputs=['result'],
    axis=axis,
    keepdims=keepdims)
# result: [[0, 1]]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_no_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 4]
result = argmax_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmax_no_keepdims_random')

ArgMin

Computes the indices of the min elements of the input tensor's element along the provided axis. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned. The type of the output tensor is integer.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is 0)

The axis in which to compute the arg indices.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : tensor(int64)

Reduced output tensor with integer data type.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to all numeric tensors.

Examples

default_axes_keepdims

data = np.array([[2, 1], [3, 10]], dtype=np.float32)
keepdims = 1
node = onnx.helper.make_node(
    'ArgMin',
    inputs=['data'],
    outputs=['result'],
    keepdims=keepdims)

# result: [[0], [0]]
result = argmin_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_default_axis_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [1, 3, 4]
result = argmin_use_numpy(data, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_default_axis_random')

keepdims

data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 1
node = onnx.helper.make_node(
    'ArgMin',
    inputs=['data'],
    outputs=['result'],
    axis=axis,
    keepdims=keepdims)
# result: [[1], [0]]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 1, 4]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_keepdims_random')

no_keepdims

data = np.array([[2, 1], [3, 10]], dtype=np.float32)
axis = 1
keepdims = 0
node = onnx.helper.make_node(
    'ArgMin',
    inputs=['data'],
    outputs=['result'],
    axis=axis,
    keepdims=keepdims)
# result: [[1, 0]]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_no_keepdims_example')

data = np.random.uniform(-10, 10, [2, 3, 4]).astype(np.float32)
# result's shape: [2, 4]
result = argmin_use_numpy(data, axis=axis, keepdims=keepdims)
expect(node, inputs=[data], outputs=[result], name='test_argmin_no_keepdims_random')

Asin

Calculates the arcsine (inverse of sine) of the given input tensor, element-wise.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The arcsine of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

asin

node = onnx.helper.make_node(
    'Asin',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-0.5, 0, 0.5]).astype(np.float32)
y = np.arcsin(x)
expect(node, inputs=[x], outputs=[y],
       name='test_asin_example')

x = np.random.rand(3, 4, 5).astype(np.float32)
y = np.arcsin(x)
expect(node, inputs=[x], outputs=[y],
       name='test_asin')

Asinh

Calculates the hyperbolic arcsine of the given input tensor element-wise.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The hyperbolic arcsine values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

asinh

node = onnx.helper.make_node(
    'Asinh',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.arcsinh(x)  # expected output [-0.88137358,  0.,  0.88137358]
expect(node, inputs=[x], outputs=[y],
       name='test_asinh_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.arcsinh(x)
expect(node, inputs=[x], outputs=[y],
       name='test_asinh')

Atan

Calculates the arctangent (inverse of tangent) of the given input tensor, element-wise.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The arctangent of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

atan

node = onnx.helper.make_node(
    'Atan',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.arctan(x)
expect(node, inputs=[x], outputs=[y],
       name='test_atan_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.arctan(x)
expect(node, inputs=[x], outputs=[y],
       name='test_atan')

Atanh

Calculates the hyperbolic arctangent of the given input tensor element-wise.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The hyperbolic arctangent values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

atanh

node = onnx.helper.make_node(
    'Atanh',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-0.5, 0, 0.5]).astype(np.float32)
y = np.arctanh(x)  # expected output [-0.54930615,  0.,  0.54930615]
expect(node, inputs=[x], outputs=[y],
       name='test_atanh_example')

x = np.random.uniform(0.0, 1.0, (3, 4, 5)).astype(np.float32)
y = np.arctanh(x)
expect(node, inputs=[x], outputs=[y],
       name='test_atanh')

AveragePool

AveragePool consumes an input tensor X and applies average pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. average pooling consisting of computing the average on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - kernel_spatial_shape[i]) / strides_spatial_shape[i] + 1)

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - kernel_spatial_shape[i] + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + kernel_spatial_shape[i] - input_spatial_shape[i]

The output of each pooling window is divided by the number of elements (exclude pad when attribute count_include_pad is zero).

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: AveragePool-1

Attributes

auto_pad : string (default is NOTSET)

auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.

count_include_pad : int (default is 0)

Whether include pad pixels when calculating values for the edges. Default is 0, doesn't count include pad.

kernel_shape : list of ints (required)

The size of the kernel along each axis.

pads : list of ints

Padding for the beginning and ending along each axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each axis.

strides : list of ints

Stride along each axis.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].

Outputs

Y : T

Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

averagepool_1d_default

"""
input_shape: [1, 3, 32]
output_shape: [1, 3, 31]
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2],
)
x = np.random.randn(1, 3, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = [2]
strides = [1]
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0], 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_1d_default')

averagepool_2d_default

"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 31, 31]
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_default')

averagepool_2d_pads

"""
input_shape: [1, 3, 28, 28]
output_shape: [1, 3, 30, 30]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[3, 3],
    pads=[2, 2, 2, 2]
)
x = np.random.randn(1, 3, 28, 28).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (3, 3)
strides = (1, 1)
pad_bottom = 2
pad_top = 2
pad_right = 2
pad_left = 2
pad_shape = [pad_top + pad_bottom, pad_left + pad_right]
out_shape = get_output_shape('VALID', np.add(x_shape[2:], pad_shape), kernel_shape, strides)
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_pads')

averagepool_2d_pads_count_include_pad

"""
input_shape: [1, 3, 28, 28]
output_shape: [1, 3, 30, 30]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[3, 3],
    pads=[2, 2, 2, 2],
    count_include_pad=1,
)
x = np.random.randn(1, 3, 28, 28).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (3, 3)
strides = (1, 1)
pad_bottom = 2
pad_top = 2
pad_right = 2
pad_left = 2
pad_shape = [pad_top + pad_bottom, pad_left + pad_right]
out_shape = get_output_shape('VALID', np.add(x_shape[2:], pad_shape), kernel_shape, strides)
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=0)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG', count_include_pad=1)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_pads_count_include_pad')

averagepool_2d_precomputed_pads

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2]

)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[7, 7.5, 8, 8.5, 9],
                [9.5, 10, 10.5, 11, 11.5],
                [12, 12.5, 13, 13.5, 14],
                [14.5, 15, 15.5, 16, 16.5],
                [17, 17.5, 18, 18.5, 19]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_pads')

averagepool_2d_precomputed_pads_count_include_pad

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2],
    count_include_pad=1
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[2.5200, 3.6000, 4.8000, 4.0800, 3.2400],
                [4.5600, 6.4000, 8.4000, 7.0400, 5.5200],
                [7.2000, 10.0000, 13.0000, 10.8000, 8.4000],
                [6.9600, 9.6000, 12.4000, 10.2400, 7.9200],
                [6.1200, 8.4000, 10.8000, 8.8800, 6.8400]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_pads_count_include_pad')

averagepool_2d_precomputed_same_upper

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 3, 3]
pad_shape: [2, 2] -> [1, 1, 1, 1] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[3, 3],
    strides=[2, 2],
    auto_pad='SAME_UPPER'
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[4, 5.5, 7],
                [11.5, 13, 14.5],
                [19, 20.5, 22]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_same_upper')

averagepool_2d_precomputed_strides

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
    strides=[2, 2]
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[4, 6],
                [14, 16]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_precomputed_strides')

averagepool_2d_same_lower

"""
iutput_shape: [1, 3, 32, 32]
output_shape: [1, 3, 32, 32]
pad_shape: [1, 1] -> [1, 0, 1, 0] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
    auto_pad='SAME_LOWER'
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape('SAME_LOWER', x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape('SAME_LOWER', x_shape[2:], kernel_shape, strides, out_shape)
pad_bottom = pad_shape[0] // 2
pad_top = pad_shape[0] - pad_bottom
pad_right = pad_shape[1] // 2
pad_left = pad_shape[1] - pad_right
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_same_lower')

averagepool_2d_same_upper

"""
iutput_shape: [1, 3, 32, 32]
output_shape: [1, 3, 32, 32]
pad_shape: [1, 1] -> [0, 1, 0, 1] by axis
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
    auto_pad='SAME_UPPER'
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape('SAME_UPPER', x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape('SAME_UPPER', x_shape[2:], kernel_shape, strides, out_shape)
pad_top = pad_shape[0] // 2
pad_bottom = pad_shape[0] - pad_top
pad_left = pad_shape[1] // 2
pad_right = pad_shape[1] - pad_left
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_same_upper')

averagepool_2d_strides

"""
input_shape: [1, 3, 32, 32]
output_shape: [1, 3, 10, 10]
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[5, 5],
    strides=[3, 3]
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (5, 5)
strides = (3, 3)
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_2d_strides')

averagepool_3d_default

"""
input_shape: [1, 3, 32, 32, 32]
output_shape: [1, 3, 31, 31, 31]
"""
node = onnx.helper.make_node(
    'AveragePool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2, 2],
)
x = np.random.randn(1, 3, 32, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = [2, 2, 2]
strides = [1, 1, 1]
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0, 0, 0], 'AVG')

expect(node, inputs=[x], outputs=[y], name='test_averagepool_3d_default')

BatchNormalization

Carries out batch normalization as described in the paper https://arxiv.org/abs/1502.03167. Depending on the mode it is being run, there are multiple cases for the number of outputs, which we list below:

Output case #1: Y, mean, var, saved_mean, saved_var (training mode) Output case #2: Y (test mode) This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: BatchNormalization-1, BatchNormalization-6

Attributes

epsilon : float (default is 1e-05)

The epsilon value to use to avoid division by zero.

momentum : float (default is 0.9)

Factor used in computing the running mean and variance.e.g., running_mean = running_mean * momentum + mean * (1 - momentum).

spatial : int (default is 1)

If true, compute the mean and variance across per activation. If false, compute the mean and variance across per feature over each mini-batch.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size.

scale : T

If spatial is true, the dimension of scale is (C). If spatial is false, the dimensions of scale are (C x D1 x ... x Dn)

B : T

If spatial is true, the dimension of bias is (C). If spatial is false, the dimensions of bias are (C x D1 x ... x Dn)

mean : T

If spatial is true, the dimension of the running mean (training) or the estimated mean (testing) is (C). If spatial is false, the dimensions of the running mean (training) or the estimated mean (testing) are (C x D1 x ... x Dn).

var : T

If spatial is true, the dimension of the running variance(training) or the estimated variance (testing) is (C). If spatial is false, the dimensions of the running variance(training) or the estimated variance (testing) are (C x D1 x ... x Dn).

Outputs (1 - 5)

Y : T

The output tensor of the same shape as X

mean (optional) : T

The running mean after the BatchNormalization operator.

var (optional) : T

The running variance after the BatchNormalization operator.

saved_mean (optional) : T

Saved mean used during training to speed up gradient computation.

saved_var (optional) : T

Saved variance used during training to speed up gradient computation.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

batchnormalization

def _batchnorm_test_mode(x, s, bias, mean, var, epsilon=1e-5):  # type: ignore
    dims_x = len(x.shape)
    dim_ones = (1,) * (dims_x - 2)
    s = s.reshape(-1, *dim_ones)
    bias = bias.reshape(-1, *dim_ones)
    mean = mean.reshape(-1, *dim_ones)
    var = var.reshape(-1, *dim_ones)
    return s * (x - mean) / np.sqrt(var + epsilon) + bias

# input size: (1, 2, 1, 3)
x = np.array([[[[-1, 0, 1]], [[2, 3, 4]]]]).astype(np.float32)
s = np.array([1.0, 1.5]).astype(np.float32)
bias = np.array([0, 1]).astype(np.float32)
mean = np.array([0, 3]).astype(np.float32)
var = np.array([1, 1.5]).astype(np.float32)
y = _batchnorm_test_mode(x, s, bias, mean, var).astype(np.float32)

node = onnx.helper.make_node(
    'BatchNormalization',
    inputs=['x', 's', 'bias', 'mean', 'var'],
    outputs=['y'],
)

# output size: (1, 2, 1, 3)
expect(node, inputs=[x, s, bias, mean, var], outputs=[y],
       name='test_batchnorm_example')

# input size: (2, 3, 4, 5)
x = np.random.randn(2, 3, 4, 5).astype(np.float32)
s = np.random.randn(3).astype(np.float32)
bias = np.random.randn(3).astype(np.float32)
mean = np.random.randn(3).astype(np.float32)
var = np.random.rand(3).astype(np.float32)
epsilon = 1e-2
y = _batchnorm_test_mode(x, s, bias, mean, var, epsilon).astype(np.float32)

node = onnx.helper.make_node(
    'BatchNormalization',
    inputs=['x', 's', 'bias', 'mean', 'var'],
    outputs=['y'],
    epsilon=epsilon,
)

# output size: (2, 3, 4, 5)
expect(node, inputs=[x, s, bias, mean, var], outputs=[y],
       name='test_batchnorm_epsilon')

Cast

The operator casts the elements of a given input tensor to a data type specified by the 'to' argument and returns an output tensor of the same size in the converted type. The 'to' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message. NOTE: Casting to and from strings is not supported yet.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Cast-1

Attributes

to : int (required)

The data type to which the elements of the input tensor are cast.Strictly must be one of the types from DataType enum in TensorProto

Inputs

input : T1

Input tensor to be cast.

Outputs

output : T2

Output tensor with the same shape as input with type specified by the 'to' argument

Type Constraints

T1 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)

Constrain input types. Casting from strings and complex are not supported.

T2 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)

Constrain output types. Casting to strings and complex are not supported.

Examples

cast

shape = (3, 4)
test_cases = [
    ('FLOAT', 'FLOAT16'),
    ('FLOAT', 'DOUBLE'),
    ('FLOAT16', 'FLOAT'),
    ('FLOAT16', 'DOUBLE'),
    ('DOUBLE', 'FLOAT'),
    ('DOUBLE', 'FLOAT16'),
]

for from_type, to_type in test_cases:
    input = np.random.random_sample(shape).astype(
        TENSOR_TYPE_TO_NP_TYPE[getattr(TensorProto, from_type)])
    node = onnx.helper.make_node(
        'Cast',
        inputs=['input'],
        outputs=['output'],
        to=getattr(TensorProto, to_type),
    )
    output = input.astype(TENSOR_TYPE_TO_NP_TYPE[getattr(TensorProto, to_type)])

    expect(node, inputs=[input], outputs=[output],
           name='test_cast_' + from_type + '_to_' + to_type)

Ceil

Ceil takes one input data (Tensor) and produces one output data (Tensor) where the ceil is, y = ceil(x), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Ceil-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

ceil

node = onnx.helper.make_node(
    'Ceil',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1.5, 1.2]).astype(np.float32)
y = np.ceil(x)  # expected output [-1., 2.]
expect(node, inputs=[x], outputs=[y],
       name='test_ceil_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.ceil(x)
expect(node, inputs=[x], outputs=[y],
       name='test_ceil')

Clip

Clip operator limits the given input within an interval. The interval is specified with arguments 'min' and 'max'. They default to numeric_limits::lowest() and numeric_limits::max() respectively.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Clip-1

Attributes

max : float (default is 3.4028234663852886e+38)

Maximum value, above which element is replaced by max

min : float (default is -3.4028234663852886e+38)

Minimum value, under which element is replaced by min

Inputs

input : T

Input tensor whose elements to be clipped

Outputs

output : T

Output tensor with clipped input elements

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

clip

node = onnx.helper.make_node(
    'Clip',
    inputs=['x'],
    outputs=['y'],
    min=-1.0,
    max=1.0
)

x = np.array([-2, 0, 2]).astype(np.float32)
y = np.clip(x, -1, 1)  # expected output [-1., 0., 1.]
expect(node, inputs=[x], outputs=[y],
       name='test_clip_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, -1.0, 1.0)
expect(node, inputs=[x], outputs=[y],
       name='test_clip')
node = onnx.helper.make_node(
    'Clip',
    inputs=['x'],
    outputs=['y'],
    min=-5.0,
    max=5.0,
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.array([-1, 0, 1]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_clip_inbounds')

x = np.array([-6, 0, 6]).astype(np.float32)
y = np.array([-5, 0, 5]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_clip_outbounds')

x = np.array([-1, 0, 6]).astype(np.float32)
y = np.array([-1, 0, 5]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_clip_splitbounds')

clip_default

node = onnx.helper.make_node(
    'Clip',
    inputs=['x'],
    outputs=['y'],
    min=0.0
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0.0, np.inf)
expect(node, inputs=[x], outputs=[y],
       name='test_clip_default_min')

node = onnx.helper.make_node(
    'Clip',
    inputs=['x'],
    outputs=['y'],
    max=0.0
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, -np.inf, 0.0)
expect(node, inputs=[x], outputs=[y],
       name='test_clip_default_max')
node = onnx.helper.make_node(
    'Clip',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.array([-1, 0, 1]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_clip_default_inbounds')

Compress

Selects slices from an input tensor along a given axis where condition evaluates to True for each axis index. In case axis is not provided, input is flattened before elements are selected. Compress behaves like numpy.compress: https://docs.scipy.org/doc/numpy/reference/generated/numpy.compress.html

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Attributes

axis : int

(Optional) Axis along which to take slices. If not specified, input is flattened before elements being selected.

Inputs

input : T

Tensor of rank r >= 1.

condition : T1

Rank 1 tensor of booleans to indicate which slices or data elements to be selected. Its length can be less than the input length alone the axis or the flattened input size if axis is not specified. In such cases data slices or elements exceeding the condition length are discarded.

Outputs

output : T

Tensor of rank r if axis is specified. Otherwise output is a Tensor of rank 1.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

T1 : tensor(bool)

Constrains to boolean tensors.

Examples

compress_0

node = onnx.helper.make_node(
    'Compress',
    inputs=['input', 'condition'],
    outputs=['output'],
    axis=0,
)
input = np.array([[1, 2], [3, 4], [5, 6]]).astype(np.float32)
condition = np.array([0, 1, 1])
output = np.compress(condition, input, axis=0)
#print(output)
#[[ 3.  4.]
# [ 5.  6.]]

expect(node, inputs=[input, condition.astype(np.bool)], outputs=[output],
       name='test_compress_0')

compress_1

node = onnx.helper.make_node(
    'Compress',
    inputs=['input', 'condition'],
    outputs=['output'],
    axis=1,
)
input = np.array([[1, 2], [3, 4], [5, 6]]).astype(np.float32)
condition = np.array([0, 1])
output = np.compress(condition, input, axis=1)
#print(output)
#[[ 2.]
# [ 4.]
# [ 6.]]

expect(node, inputs=[input, condition.astype(np.bool)], outputs=[output],
       name='test_compress_1')

compress_default_axis

node = onnx.helper.make_node(
    'Compress',
    inputs=['input', 'condition'],
    outputs=['output'],
)
input = np.array([[1, 2], [3, 4], [5, 6]]).astype(np.float32)
condition = np.array([0, 1, 0, 0, 1])
output = np.compress(condition, input)
#print(output)
#[ 2., 5.]

expect(node, inputs=[input, condition.astype(np.bool)], outputs=[output],
       name='test_compress_default_axis')

Concat

Concatenate a list of tensors into a single tensor

Version

This version of the operator has been available since version 4 of the default ONNX operator set.

Other versions of this operator: Concat-1

Attributes

axis : int (required)

Which axis to concat on

Inputs (1 - ∞)

inputs (variadic) : T

List of tensors for concatenation

Outputs

concat_result : T

Concatenated tensor

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain output types to any tensor type.

Examples

concat

test_cases = {
    '1d': ([1, 2],
           [3, 4]),
    '2d': ([[1, 2], [3, 4]],
           [[5, 6], [7, 8]]),
    '3d': ([[[1, 2], [3, 4]], [[5, 6], [7, 8]]],
           [[[9, 10], [11, 12]], [[13, 14], [15, 16]]])
}  # type: Dict[Text, Sequence[Any]]

for test_case, values_ in test_cases.items():
    values = [np.asarray(v, dtype=np.float32) for v in values_]
    for i in range(len(values[0].shape)):
        in_args = ['value' + str(k) for k in range(len(values))]
        node = onnx.helper.make_node(
            'Concat',
            inputs=[s for s in in_args],
            outputs=['output'],
            axis=i
        )
        output = np.concatenate(values, i)
        expect(node, inputs=[v for v in values], outputs=[output],
               name='test_concat_' + test_case + '_axis_' + str(i))

Constant

A constant tensor.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Constant-1

Attributes

value : tensor (required)

The value for the elements of the output tensor.

Inputs

Outputs

output : T

Output tensor containing the same value of the provided tensor.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

constant

values = np.random.randn(5, 5).astype(np.float32)
node = onnx.helper.make_node(
    'Constant',
    inputs=[],
    outputs=['values'],
    value=onnx.helper.make_tensor(
        name='const_tensor',
        data_type=onnx.TensorProto.FLOAT,
        dims=values.shape,
        vals=values.flatten().astype(float),
    ),
)

expect(node, inputs=[], outputs=[values],
       name='test_constant')

ConstantLike

Generate a tensor with specific constant value. The value can be specified by the 'value' attribute. The shape of the output tensor is the same as the input tensor, if the input tensor is provided, or the shape provided in the 'shape' attribute (if both are provided, the input tensor shape takes precendence). The data type can be specified by the 'dtype' argument. If 'dtype' is not specified, then the type of input tensor is used. If input tensor is also not specified, then the type defaults to 'float'.

The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message and be valid as an output type.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Attributes

dtype : int

(Optional) The data type for the elements of the output tensor. If not specified,the data type of the input tensor T1 is used. If input tensor T1 is also not specified, then output tensor type defaults to 'float'.

shape : list of ints

(Optional) The shape of the output tensor. If input tensor T1 is provided, then 'shape' attribute is ignored and the output follows the shape of the input. One of either input tensor T1 or 'shape' attribute must be provided.

value : float (default is 0.0)

(Optional) The value for the elements of the output tensor.

Inputs (0 - 1)

input (optional) : T1

Input tensor to copy shape, and optionally, type information from. One of either input tensor T1 or 'shape' attribute must be provided.

Outputs

output : T2

Output tensor, same shape as input tensor T1.

Type Constraints

T1 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)

Constrain input types. Strings and complex are not supported.

T2 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)

Constrain output types. Strings and complex are not supported.

Examples

ones_with_input

shape = (4, 3, 2)
node = onnx.helper.make_node(
    'ConstantLike',
    inputs=['x'],
    outputs=['y'],
    value=1.0,
)
x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.ones(shape, dtype=np.int32)
expect(node, inputs=[x], outputs=[y], name='test_constantlike_ones_with_input')

threes_with_shape_and_dtype

shape = (3, 4)
node = onnx.helper.make_node(
    'ConstantLike',
    shape=shape,
    inputs=[],
    outputs=['y'],
    value=3.0,
    dtype=onnx.TensorProto.DOUBLE,  # 11: DOUBLE
)

y = 3.0 * np.ones(shape, dtype=np.float64)
expect(node, inputs=[], outputs=[y], name='test_constantlike_threes_with_shape_and_dtype')

zeros_without_input_dtype

shape = (2, 5, 1)
node = onnx.helper.make_node(
    'ConstantLike',
    inputs=[],
    outputs=['y'],
    shape=shape,
)
y = np.zeros(shape, dtype=np.float32)
expect(node, inputs=[], outputs=[y], name='test_constantlike_zeros_without_input_dtype')

Conv

The convolution operator consumes an input tensor and a filter, and computes the output.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

auto_pad : string (default is NOTSET)

auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.

dilations : list of ints

dilation value along each axis of the filter.

group : int (default is 1)

number of groups input channels and output channels are divided into.

kernel_shape : list of ints

The shape of the convolution kernel. If not present, should be inferred from input W.

pads : list of ints

Padding for the beginning and ending along each axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each axis.

strides : list of ints

Stride along each axis.

Inputs (2 - 3)

X : T

Input data tensor from previous layer; has size (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and width. Note that this is for the 2D image. Otherwise the size is (N x C x D1 x D2 ... x Dn). Optionally, if dimension denotation is in effect, the operation expects input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].

W : T

The weight tensor that will be used in the convolutions; has size (M x C/group x kH x kW), where C is the number of channels, and kH and kW are the height and width of the kernel, and M is the number of feature maps. For more than 2 dimensions, the kernel shape will be (M x C/group x k1 x k2 x ... x kn), where (k1 x k2 x ... kn) is the dimension of the kernel. Optionally, if dimension denotation is in effect, the operation expects the weight tensor to arrive with the dimension denotation of [FILTER_OUT_CHANNEL, FILTER_IN_CHANNEL, FILTER_SPATIAL, FILTER_SPATIAL ...]. X.shape[1] == (W.shape[1] * group) == C (assuming zero based indices for the shape array). Or in other words FILTER_IN_CHANNEL should be equal to DATA_CHANNEL.

B (optional) : T

Optional 1D bias to be added to the convolution, has size of M.

Outputs

Y : T

Output data tensor that contains the result of the convolution. The output dimensions are functions of the kernel size, stride size, and pad lengths.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

conv

x = np.array([[[[0., 1., 2., 3., 4.],  # (1, 1, 5, 5) input tensor
                [5., 6., 7., 8., 9.],
                [10., 11., 12., 13., 14.],
                [15., 16., 17., 18., 19.],
                [20., 21., 22., 23., 24.]]]]).astype(np.float32)
W = np.array([[[[1., 1., 1.],  # (1, 1, 3, 3) tensor for convolution weights
                [1., 1., 1.],
                [1., 1., 1.]]]]).astype(np.float32)

# Convolution with padding
node_with_padding = onnx.helper.make_node(
    'Conv',
    inputs=['x', 'W'],
    outputs=['y'],
    kernel_shape=[3, 3],
    # Default values for other attributes: strides=[1, 1], dilations=[1, 1], groups=1
    pads=[1, 1, 1, 1],
)
y_with_padding = np.array([[[[12., 21., 27., 33., 24.],  # (1, 1, 5, 5) output tensor
                             [33., 54., 63., 72., 51.],
                             [63., 99., 108., 117., 81.],
                             [93., 144., 153., 162., 111.],
                             [72., 111., 117., 123., 84.]]]]).astype(np.float32)
expect(node_with_padding, inputs=[x, W], outputs=[y_with_padding],
       name='test_basic_conv_with_padding')

# Convolution without padding
node_without_padding = onnx.helper.make_node(
    'Conv',
    inputs=['x', 'W'],
    outputs=['y'],
    kernel_shape=[3, 3],
    # Default values for other attributes: strides=[1, 1], dilations=[1, 1], groups=1
    pads=[0, 0, 0, 0],
)
y_without_padding = np.array([[[[54., 63., 72.],  # (1, 1, 3, 3) output tensor
                                [99., 108., 117.],
                                [144., 153., 162.]]]]).astype(np.float32)
expect(node_without_padding, inputs=[x, W], outputs=[y_without_padding],
       name='test_basic_conv_without_padding')

conv_with_strides

x = np.array([[[[0., 1., 2., 3., 4.],  # (1, 1, 7, 5) input tensor
                [5., 6., 7., 8., 9.],
                [10., 11., 12., 13., 14.],
                [15., 16., 17., 18., 19.],
                [20., 21., 22., 23., 24.],
                [25., 26., 27., 28., 29.],
                [30., 31., 32., 33., 34.]]]]).astype(np.float32)
W = np.array([[[[1., 1., 1.],  # (1, 1, 3, 3) tensor for convolution weights
                [1., 1., 1.],
                [1., 1., 1.]]]]).astype(np.float32)

# Convolution with strides=2 and padding
node_with_padding = onnx.helper.make_node(
    'Conv',
    inputs=['x', 'W'],
    outputs=['y'],
    kernel_shape=[3, 3],
    pads=[1, 1, 1, 1],
    strides=[2, 2],  # Default values for other attributes: dilations=[1, 1], groups=1
)
y_with_padding = np.array([[[[12., 27., 24.],  # (1, 1, 4, 3) output tensor
                             [63., 108., 81.],
                             [123., 198., 141.],
                             [112., 177., 124.]]]]).astype(np.float32)
expect(node_with_padding, inputs=[x, W], outputs=[y_with_padding],
       name='test_conv_with_strides_padding')

# Convolution with strides=2 and no padding
node_without_padding = onnx.helper.make_node(
    'Conv',
    inputs=['x', 'W'],
    outputs=['y'],
    kernel_shape=[3, 3],
    pads=[0, 0, 0, 0],
    strides=[2, 2],  # Default values for other attributes: dilations=[1, 1], groups=1
)
y_without_padding = np.array([[[[54., 72.],  # (1, 1, 3, 2) output tensor
                                [144., 162.],
                                [234., 252.]]]]).astype(np.float32)
expect(node_without_padding, inputs=[x, W], outputs=[y_without_padding],
       name='test_conv_with_strides_no_padding')

# Convolution with strides=2 and padding only along one dimension (the H dimension in NxCxHxW tensor)
node_with_asymmetric_padding = onnx.helper.make_node(
    'Conv',
    inputs=['x', 'W'],
    outputs=['y'],
    kernel_shape=[3, 3],
    pads=[1, 0, 1, 0],
    strides=[2, 2],  # Default values for other attributes: dilations=[1, 1], groups=1
)
y_with_asymmetric_padding = np.array([[[[21., 33.],  # (1, 1, 4, 2) output tensor
                                        [99., 117.],
                                        [189., 207.],
                                        [171., 183.]]]]).astype(np.float32)
expect(node_with_asymmetric_padding, inputs=[x, W], outputs=[y_with_asymmetric_padding],
       name='test_conv_with_strides_and_asymmetric_padding')

ConvTranspose

The convolution transpose operator consumes an input tensor and a filter, and computes the output.

If the pads parameter is provided the shape of the output is calculated via the following equation:

output_shape[i] = stride[i] * (input_size[i] - 1) + output_padding[i] + kernel_shape[i] - pads[start_i] - pads[end_i]

output_shape can also be explicitly specified in which case pads values are auto generated using these equations:

total_padding[i] = stride[i] * (input_size[i] - 1) + output_padding[i] + kernel_shape[i] - output_shape[i]
If (auto_pads != SAME_UPPER): pads[start_i] = total_padding[i]/2; pads[end_i] = total_padding[i] - (total_padding[i]/2)
Else: pads[start_i] = total_padding[i] - (total_padding[i]/2); pads[end_i] = (total_padding[i]/2).

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

auto_pad : string (default is NOTSET)

auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.

dilations : list of ints

dilation value along each axis of the filter.

group : int (default is 1)

number of groups input channels and output channels are divided into.

kernel_shape : list of ints

The shape of the convolution kernel. If not present, should be inferred from input W.

output_padding : list of ints

The zero-padding added to one side of the output. This is also called adjs/adjustment in some frameworks.

output_shape : list of ints

The shape of the output can be explicitly set which will cause pads values to be auto generated. If output_shape is specified pads values are ignored. See doc for details for equations to generate pads

pads : list of ints

Padding for the beginning and ending along each axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each axis.

strides : list of ints

Stride along each axis.

Inputs (2 - 3)

X : T

Input data tensor from previous layer; has size (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and width. Note that this is for the 2D image. Otherwise the size is (N x C x D1 x D2 ... x Dn)

W : T

The weight tensor that will be used in the convolutions; has size (C x M/group x kH x kW), where C is the number of channels, and kH and kW are the height and width of the kernel, and M is the number of feature maps. For more than 2 dimensions, the weight shape will be (C x M/group x k1 x k2 x ... x kn), where (k1 x k2 x ... x kn) is the dimension of the kernel. The number of channels in the output should be equal to W.shape[1] * group (assuming zero based indices of the shape array)

B (optional) : T

Optional 1D bias to be added to the convolution, has size of M.

Outputs

Y : T

Output data tensor that contains the result of the convolution. The output dimensions are functions of the kernel size, stride size, pad lengths and group count. The number of channels in the output should be equal to W.shape[1] * group (assuming zero based indices of the shape array)

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

convtranspose

x = np.array([[[[0., 1., 2.],  # (1, 1, 3, 3)
                [3., 4., 5.],
                [6., 7., 8.]]]]).astype(np.float32)

W = np.array([[[[1., 1., 1.],  # (1, 2, 3, 3)
                [1., 1., 1.],
                [1., 1., 1.]],
               [[1., 1., 1.],
                [1., 1., 1.],
                [1., 1., 1.]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"])

y = np.array([[[[0., 1., 3., 3., 2.],  # (1, 2, 5, 5)
                [3., 8., 15., 12., 7.],
                [9., 21., 36., 27., 15.],
                [9., 20., 33., 24., 13.],
                [6., 13., 21., 15., 8.]],

               [[0., 1., 3., 3., 2.],
                [3., 8., 15., 12., 7.],
                [9., 21., 36., 27., 15.],
                [9., 20., 33., 24., 13.],
                [6., 13., 21., 15., 8.]]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose')

convtranspose_1d

x = np.array([[[0., 1., 2.]]]).astype(np.float32)  # (1, 1, 3)

W = np.array([[[1., 1., 1.],  # (1, 2, 3)
               [1., 1., 1.]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"])

y = np.array([[[0., 1., 3., 3., 2.],  # (1, 2, 5)
               [0., 1., 3., 3., 2.]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_1d')

convtranspose_3d

x = np.array([[[[[0., 1., 2., 3., 4.],  # (1, 1, 3, 4, 5)
                 [5., 6., 7., 8., 9.],
                 [10., 11., 12., 13., 14.],
                 [15., 16., 17., 18., 19.]],
                [[20., 21., 22., 23., 24.],
                 [25., 26., 27., 28., 29.],
                 [30., 31., 32., 33., 34.],
                 [35., 36., 37., 38., 39.]],
                [[40., 41., 42., 43., 44.],
                 [45., 46., 47., 48., 49.],
                 [50., 51., 52., 53., 54.],
                 [55., 56., 57., 58., 59.]]]]]).astype(np.float32)

W = np.array([[[[[1., 1., 1.],  # (1, 2, 3, 3, 3)
                 [1., 1., 1.],
                 [1., 1., 1.]],
                [[1., 1., 1.],
                 [1., 1., 1.],
                 [1., 1., 1.]],
                [[1., 1., 1.],
                 [1., 1., 1.],
                 [1., 1., 1.]]],
               [[[1., 1., 1.],
                 [1., 1., 1.],
                 [1., 1., 1.]],
                [[1., 1., 1.],
                 [1., 1., 1.],
                 [1., 1., 1.]],
                [[1., 1., 1.],
                 [1., 1., 1.],
                 [1., 1., 1.]]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"])

y = np.array([[[[[0., 1., 3., 6., 9., 7., 4.],  # (1, 2, 5, 6, 7)
                 [5., 12., 21., 27., 33., 24., 13.],
                 [15., 33., 54., 63., 72., 51., 27.],
                 [30., 63., 99., 108., 117., 81., 42.],
                 [25., 52., 81., 87., 93., 64., 33.],
                 [15., 31., 48., 51., 54., 37., 19.]],

                [[20., 42., 66., 72., 78., 54., 28.],
                 [50., 104., 162., 174., 186., 128., 66.],
                 [90., 186., 288., 306., 324., 222., 114.],
                 [120., 246., 378., 396., 414., 282., 144.],
                 [90., 184., 282., 294., 306., 208., 106.],
                 [50., 102., 156., 162., 168., 114., 58.]],

                [[60., 123., 189., 198., 207., 141., 72.],
                 [135., 276., 423., 441., 459., 312., 159.],
                 [225., 459., 702., 729., 756., 513., 261.],
                 [270., 549., 837., 864., 891., 603., 306.],
                 [195., 396., 603., 621., 639., 432., 219.],
                 [105., 213., 324., 333., 342., 231., 117.]],

                [[60., 122., 186., 192., 198., 134., 68.],
                 [130., 264., 402., 414., 426., 288., 146.],
                 [210., 426., 648., 666., 684., 462., 234.],
                 [240., 486., 738., 756., 774., 522., 264.],
                 [170., 344., 522., 534., 546., 368., 186.],
                 [90., 182., 276., 282., 288., 194., 98.]],

                [[40., 81., 123., 126., 129., 87., 44.],
                 [85., 172., 261., 267., 273., 184., 93.],
                 [135., 273., 414., 423., 432., 291., 147.],
                 [150., 303., 459., 468., 477., 321., 162.],
                 [105., 212., 321., 327., 333., 224., 113.],
                 [55., 111., 168., 171., 174., 117., 59.]]],

               [[[0., 1., 3., 6., 9., 7., 4.],
                 [5., 12., 21., 27., 33., 24., 13.],
                 [15., 33., 54., 63., 72., 51., 27.],
                 [30., 63., 99., 108., 117., 81., 42.],
                 [25., 52., 81., 87., 93., 64., 33.],
                 [15., 31., 48., 51., 54., 37., 19.]],

                [[20., 42., 66., 72., 78., 54., 28.],
                 [50., 104., 162., 174., 186., 128., 66.],
                 [90., 186., 288., 306., 324., 222., 114.],
                 [120., 246., 378., 396., 414., 282., 144.],
                 [90., 184., 282., 294., 306., 208., 106.],
                 [50., 102., 156., 162., 168., 114., 58.]],

                [[60., 123., 189., 198., 207., 141., 72.],
                 [135., 276., 423., 441., 459., 312., 159.],
                 [225., 459., 702., 729., 756., 513., 261.],
                 [270., 549., 837., 864., 891., 603., 306.],
                 [195., 396., 603., 621., 639., 432., 219.],
                 [105., 213., 324., 333., 342., 231., 117.]],

                [[60., 122., 186., 192., 198., 134., 68.],
                 [130., 264., 402., 414., 426., 288., 146.],
                 [210., 426., 648., 666., 684., 462., 234.],
                 [240., 486., 738., 756., 774., 522., 264.],
                 [170., 344., 522., 534., 546., 368., 186.],
                 [90., 182., 276., 282., 288., 194., 98.]],

                [[40., 81., 123., 126., 129., 87., 44.],
                 [85., 172., 261., 267., 273., 184., 93.],
                 [135., 273., 414., 423., 432., 291., 147.],
                 [150., 303., 459., 468., 477., 321., 162.],
                 [105., 212., 321., 327., 333., 224., 113.],
                 [55., 111., 168., 171., 174., 117., 59.]]]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_3d')

convtranspose_attributes

x = np.array([[[[0., 1., 2.],  # (1, 1, 3, 3)
                [3., 4., 5.],
                [6., 7., 8.]]]]).astype(np.float32)

W = np.array([[[[1., 1., 1.],  # (1, 2, 3, 3)
                [1., 1., 1.],
                [1., 1., 1.]],
               [[1., 1., 1.],
                [1., 1., 1.],
                [1., 1., 1.]]]]).astype(np.float32)

y = np.array([[[[0., 0., 1., 1., 3., 2., 2., 0.],  # (1, 2, 10, 8)
                [0., 0., 1., 1., 3., 2., 2., 0.],
                [0., 0., 1., 1., 3., 2., 2., 0.],
                [3., 3., 7., 4., 9., 5., 5., 0.],
                [3., 3., 7., 4., 9., 5., 5., 0.],
                [3., 3., 7., 4., 9., 5., 5., 0.],
                [6., 6., 13., 7., 15., 8., 8., 0.],
                [6., 6., 13., 7., 15., 8., 8., 0.],
                [6., 6., 13., 7., 15., 8., 8., 0.],
                [0., 0., 0., 0., 0., 0., 0., 0.]],

               [[0., 0., 1., 1., 3., 2., 2., 0.],
                [0., 0., 1., 1., 3., 2., 2., 0.],
                [0., 0., 1., 1., 3., 2., 2., 0.],
                [3., 3., 7., 4., 9., 5., 5., 0.],
                [3., 3., 7., 4., 9., 5., 5., 0.],
                [3., 3., 7., 4., 9., 5., 5., 0.],
                [6., 6., 13., 7., 15., 8., 8., 0.],
                [6., 6., 13., 7., 15., 8., 8., 0.],
                [6., 6., 13., 7., 15., 8., 8., 0.],
                [0., 0., 0., 0., 0., 0., 0., 0.]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"],
                             strides=[3, 2],
                             output_shape=[10, 8])
expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_output_shape')

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"],
                             strides=[3, 2],
                             output_padding=[1, 1])
expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_pad')

node = onnx.helper.make_node(
    'ConvTranspose', ['X', 'W'], ['Y'],
    name='test',
    strides=[3, 2],
    output_shape=[10, 8],
    kernel_shape=[3, 3],
    output_padding=[1, 1]
)
expect(node, inputs=[x, W], outputs=[y],
       name='test_convtranspose_kernel_shape')

convtranspose_pads

x = np.array([[[[0., 1., 2.],  # (1, 1, 3, 3)
                [3., 4., 5.],
                [6., 7., 8.]]]]).astype(np.float32)

W = np.array([[[[1., 1., 1.],  # (1, 2, 3, 3)
                [1., 1., 1.],
                [1., 1., 1.]],
               [[1., 1., 1.],
                [1., 1., 1.],
                [1., 1., 1.]]]]).astype(np.float32)

node = onnx.helper.make_node("ConvTranspose", ["X", "W"], ["Y"],
                             strides=[3, 2],
                             pads=[1, 2, 1, 2])

y = np.array([[[[1., 1., 3.],  # (1, 2, 7, 3)
                [1., 1., 3.],
                [7., 4., 9.],
                [7., 4., 9.],
                [7., 4., 9.],
                [13., 7., 15.],
                [13., 7., 15.]],

               [[1., 1., 3.],
                [1., 1., 3.],
                [7., 4., 9.],
                [7., 4., 9.],
                [7., 4., 9.],
                [13., 7., 15.],
                [13., 7., 15.]]]]).astype(np.float32)

expect(node, inputs=[x, W], outputs=[y], name='test_convtranspose_pads')

Cos

Calculates the cosine of the given input tensor, element-wise.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The cosine of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

cos

node = onnx.helper.make_node(
    'Cos',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.cos(x)
expect(node, inputs=[x], outputs=[y],
       name='test_cos_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.cos(x)
expect(node, inputs=[x], outputs=[y],
       name='test_cos')

Cosh

Calculates the hyperbolic cosine of the given input tensor element-wise.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The hyperbolic cosine values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

cosh

node = onnx.helper.make_node(
    'Cosh',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.cosh(x)  # expected output [1.54308069,  1.,  1.54308069]
expect(node, inputs=[x], outputs=[y],
       name='test_cosh_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.cosh(x)
expect(node, inputs=[x], outputs=[y],
       name='test_cosh')

DepthToSpace

DepthToSpace rearranges (permutes) data from depth into blocks of spatial data. This is the reverse transformation of SpaceToDepth. More specifically, this op outputs a copy of the input tensor where values from the depth dimension are moved in spatial blocks to the height and width dimensions.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

blocksize : int (required)

Blocks of [blocksize, blocksize] are moved.

Inputs

input : T

Input tensor of [N,C,H,W], where N is the batch axis, C is the channel or depth, H is the height and W is the width.

Outputs

output : T

Output tensor of [N, C/(blocksize * blocksize), H * blocksize, W * blocksize].

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

depthtospace

b, c, h, w = shape = (2, 8, 3, 3)
blocksize = 2
node = onnx.helper.make_node(
    'DepthToSpace',
    inputs=['x'],
    outputs=['y'],
    blocksize=blocksize,
)
x = np.random.random_sample(shape).astype(np.float32)
tmp = np.reshape(x, [b, blocksize, blocksize, c // (blocksize**2), h, w])
tmp = np.transpose(tmp, [0, 3, 4, 1, 5, 2])
y = np.reshape(tmp, [b, c // (blocksize**2), h * blocksize, w * blocksize])
expect(node, inputs=[x], outputs=[y],
       name='test_depthtospace')

example

node = onnx.helper.make_node(
    'DepthToSpace',
    inputs=['x'],
    outputs=['y'],
    blocksize=2,
)

# (1, 4, 2, 3) input tensor
x = np.array([[[[0, 1, 2],
                [3, 4, 5]],
               [[6, 7, 8],
                [9, 10, 11]],
               [[12, 13, 14],
                [15, 16, 17]],
               [[18, 19, 20],
                [21, 22, 23]]]]).astype(np.float32)

# (1, 1, 4, 6) output tensor
y = np.array([[[[0, 6, 1, 7, 2, 8],
                [12, 18, 13, 19, 14, 20],
                [3, 9, 4, 10, 5, 11],
                [15, 21, 16, 22, 17, 23]]]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_depthtospace_example')

Div

Performs element-wise binary division (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Div-1, Div-6

Inputs

A : T

First operand.

B : T

Second operand.

Outputs

C : T

Result, has same element type as two inputs

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

div

node = onnx.helper.make_node(
    'Div',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.array([3, 4]).astype(np.float32)
y = np.array([1, 2]).astype(np.float32)
z = x / y  # expected output [3., 2.]
expect(node, inputs=[x, y], outputs=[z],
       name='test_div_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.rand(3, 4, 5).astype(np.float32) + 1.0
z = x / y
expect(node, inputs=[x, y], outputs=[z],
       name='test_div')

div_broadcast

node = onnx.helper.make_node(
    'Div',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.rand(5).astype(np.float32) + 1.0
z = x / y
expect(node, inputs=[x, y], outputs=[z],
       name='test_div_bcast')

Dropout

Dropout takes one input data (Tensor) and produces two Tensor outputs, output (Tensor) and mask (Tensor). Depending on whether it is in test mode or not, the output Y will either be a random dropout, or a simple copy of the input. Note that our implementation of Dropout does scaling in the training phase, so during testing nothing needs to be done. This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Dropout-1, Dropout-6

Attributes

ratio : float (default is 0.5)

The ratio of random dropout

Inputs

data : T

The input data as Tensor.

Outputs (1 - 2)

output : T

The output.

mask (optional) : T

The output mask.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

default

node = onnx.helper.make_node(
    'Dropout',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = x
expect(node, inputs=[x], outputs=[y],
       name='test_dropout_default')

random

node = onnx.helper.make_node(
    'Dropout',
    inputs=['x'],
    outputs=['y'],
    ratio=.2,
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = x
expect(node, inputs=[x], outputs=[y],
       name='test_dropout_random')

Elu

Elu takes one input data (Tensor) and produces one output data (Tensor) where the function f(x) = alpha * (exp(x) - 1.) for x < 0, f(x) = x for x >= 0., is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Elu-1

Attributes

alpha : float (default is 1.0)

Coefficient of ELU.

Inputs

X : T

1D input tensor

Outputs

Y : T

1D input tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

elu

node = onnx.helper.make_node(
    'Elu',
    inputs=['x'],
    outputs=['y'],
    alpha=2.0
)

x = np.array([-1, 0, 1]).astype(np.float32)
# expected output [-1.2642411, 0., 1.]
y = np.clip(x, 0, np.inf) + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0
expect(node, inputs=[x], outputs=[y],
       name='test_elu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0
expect(node, inputs=[x], outputs=[y],
       name='test_elu')

elu_default

default_alpha = 1.0
node = onnx.helper.make_node(
    'Elu',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + (np.exp(np.clip(x, -np.inf, 0)) - 1) * default_alpha
expect(node, inputs=[x], outputs=[y],
       name='test_elu_default')

Equal

Returns the tensor resulted from performing the equal logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Equal-1

Inputs

A : T

First input operand for the logical operator.

B : T

Second input operand for the logical operator.

Outputs

C : T1

Result tensor.

Type Constraints

T : tensor(bool), tensor(int32), tensor(int64)

Constrains input to integral tensors.

T1 : tensor(bool)

Constrains output to boolean tensor.

Examples

equal

node = onnx.helper.make_node(
    'Equal',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = (np.random.randn(3, 4, 5) * 10).astype(np.int32)
y = (np.random.randn(3, 4, 5) * 10).astype(np.int32)
z = np.equal(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_equal')

equal_broadcast

node = onnx.helper.make_node(
    'Equal',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = (np.random.randn(3, 4, 5) * 10).astype(np.int32)
y = (np.random.randn(5) * 10).astype(np.int32)
z = np.equal(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_equal_bcast')

Erf

Computes the error function of the given input tensor element-wise.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The error function of the input tensor computed element-wise. It has the same shape and type of the input.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to all numeric tensors.

Examples

erf

node = onnx.helper.make_node(
    'Erf',
    inputs=['x'],
    outputs=['y'],
)

x = np.random.randn(1, 3, 32, 32).astype(np.float32)
y = np.vectorize(math.erf)(x).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_erf')

Exp

Calculates the exponential of the given input tensor, element-wise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Exp-1

Inputs

input : T

Input tensor

Outputs

output : T

The exponential of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

exp

node = onnx.helper.make_node(
    'Exp',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.exp(x)  # expected output [0.36787945, 1., 2.71828175]
expect(node, inputs=[x], outputs=[y],
       name='test_exp_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.exp(x)
expect(node, inputs=[x], outputs=[y],
       name='test_exp')

Expand

Broadcast the input tensor following the given shape and the broadcast rule. The broadcast rule is similar to numpy.array(input) * numpy.ones(shape): Dimensions are right alignment; Two corresponding dimension must have the same value, or one of them is equal to 1. Also, this operator is similar to numpy.broadcast_to(input, shape), but the major difference is numpy.broadcast_to() does not allow shape to be smaller than input.size(). It is possible that the output.shape is not equal to shape, when some dimensions in shape is equal to 1, or the shape.ndim < input.shape.ndim.

Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Inputs

input : T

Input tensor

shape : tensor(int64)

A 1-D tensor indicates the shape you want to expand to, following the broadcast rule

Outputs

output : T

Output tensor

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensors.

Examples

dim_changed

node = onnx.helper.make_node(
    'Expand',
    inputs=['data', 'new_shape'],
    outputs=['expanded'],
)
shape = [3, 1]
data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[1.], [2.], [3.]]
new_shape = [2, 1, 6]
expanded = data * np.ones(new_shape, dtype=np.float32)
#print(expanded)
#[[[1., 1., 1., 1., 1., 1.],
#  [2., 2., 2., 2., 2., 2.],
#  [3., 3., 3., 3., 3., 3.]],
#
# [[1., 1., 1., 1., 1., 1.],
#  [2., 2., 2., 2., 2., 2.],
#  [3., 3., 3., 3., 3., 3.]]]
new_shape = np.array(new_shape, dtype=np.int64)
expect(node, inputs=[data, new_shape], outputs=[expanded],
       name='test_expand_dim_changed')

dim_unchanged

node = onnx.helper.make_node(
    'Expand',
    inputs=['data', 'new_shape'],
    outputs=['expanded'],
)
shape = [3, 1]
new_shape = [3, 4]
data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[1.], [2.], [3.]]
expanded = np.tile(data, 4)
#print(expanded)
#[[1., 1., 1., 1.],
# [2., 2., 2., 2.],
# [3., 3., 3., 3.]]
new_shape = np.array(new_shape, dtype=np.int64)
expect(node, inputs=[data, new_shape], outputs=[expanded],
       name='test_expand_dim_unchanged')

EyeLike

Generate a 2D tensor (matrix) with ones on the diagonal and zeros everywhere else. Only 2D tensors are supported, i.e. input T1 must be of rank 2. The shape of the output tensor is the same as the input tensor. The data type can be specified by the 'dtype' argument. If 'dtype' is not specified, then the type of input tensor is used. By default, the main diagonal is populated with ones, but attribute 'k' can be used to populate upper or lower diagonals.

The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message and be valid as an output type.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Attributes

dtype : int

(Optional) The data type for the elements of the output tensor. If not specified,the data type of the input tensor T1 is used. If input tensor T1 is also notspecified, then type defaults to 'float'.

k : int (default is 0)

(Optional) Index of the diagonal to be populated with ones. Default is 0. If T2 is the output, this op sets T2[i, i+k] = 1. k = 0 populates the main diagonal, k > 0 populates an upper diagonal, and k < 0 populates a lower diagonal.

Inputs

input : T1

2D input tensor to copy shape, and optionally, type information from.

Outputs

output : T2

Output tensor, same shape as input tensor T1.

Type Constraints

T1 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)

Constrain input types. Strings and complex are not supported.

T2 : tensor(float16), tensor(float), tensor(double), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(bool)

Constrain output types. Strings and complex are not supported.

Examples

populate_off_main_diagonal

shape = (4, 5)
off_diagonal_offset = 1
node = onnx.helper.make_node(
    'EyeLike',
    inputs=['x'],
    outputs=['y'],
    k=off_diagonal_offset,
    dtype=onnx.TensorProto.FLOAT,
)

x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.eye(shape[0], shape[1], k=off_diagonal_offset, dtype=np.float32)
expect(node, inputs=[x], outputs=[y], name='test_eyelike_populate_off_main_diagonal')

with_dtype

shape = (3, 4)
node = onnx.helper.make_node(
    'EyeLike',
    inputs=['x'],
    outputs=['y'],
    dtype=onnx.TensorProto.DOUBLE,
)

x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.eye(shape[0], shape[1], dtype=np.float64)
expect(node, inputs=[x], outputs=[y], name='test_eyelike_with_dtype')

without_dtype

shape = (4, 4)
node = onnx.helper.make_node(
    'EyeLike',
    inputs=['x'],
    outputs=['y'],
)

x = np.random.randint(0, 100, size=shape, dtype=np.int32)
y = np.eye(shape[0], shape[1], dtype=np.int32)
expect(node, inputs=[x], outputs=[y], name='test_eyelike_without_dtype')

Flatten

Flattens the input tensor into a 2D matrix. If input tensor has shape (d_0, d_1, ... d_n) then the output will have shape (d_0 X d_1 ... d_(axis-1), d_axis X d_(axis+1) ... X dn).

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Flatten-1

Attributes

axis : int (default is 1)

Indicate up to which input dimensions (exclusive) should be flattened to the outer dimension of the output. The value for axis must be in the range [0, R], where R is the rank of the input tensor. When axis = 0, the shape of the output tensor is (1, (d_0 X d_1 ... d_n), where the shape of the input tensor is (d_0, d_1, ... d_n).

Inputs

input : T

A tensor of rank >= axis.

Outputs

output : T

A 2D tensor with the contents of the input tensor, with input dimensions up to axis flattened to the outer dimension of the output and remaining input dimensions flattened into the inner dimension of the output.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output to all tensor types.

Examples

flatten

shape = (2, 3, 4, 5)
a = np.random.random_sample(shape).astype(np.float32)

for i in range(len(shape)):
    node = onnx.helper.make_node(
        'Flatten',
        inputs=['a'],
        outputs=['b'],
        axis=i,
    )

    new_shape = (1, -1) if i == 0 else (np.prod(shape[0:i]).astype(int), -1)
    b = np.reshape(a, new_shape)
    expect(node, inputs=[a], outputs=[b],
           name='test_flatten_axis' + str(i))

flatten_with_default_axis

node = onnx.helper.make_node(
    'Flatten',
    inputs=['a'],
    outputs=['b'],  # Default value for axis: axis=1
)

shape = (5, 4, 3, 2)
a = np.random.random_sample(shape).astype(np.float32)
new_shape = (5, 24)
b = np.reshape(a, new_shape)
expect(node, inputs=[a], outputs=[b],
       name='test_flatten_default_axis')

Floor

Floor takes one input data (Tensor) and produces one output data (Tensor) where the floor is, y = floor(x), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Floor-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

floor

node = onnx.helper.make_node(
    'Floor',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1.5, 1.2, 2]).astype(np.float32)
y = np.floor(x)  # expected output [-2., 1., 2.]
expect(node, inputs=[x], outputs=[y],
       name='test_floor_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.floor(x)
expect(node, inputs=[x], outputs=[y],
       name='test_floor')

GRU

Computes an one-layer GRU. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

z - update gate

r - reset gate

h - hidden gate

t - time step (t-1 means previous time step)

W[zrh] - W parameter weight matrix for update, reset, and hidden gates

R[zrh] - R recurrence weight matrix for update, reset, and hidden gates

Wb[zrh] - W bias vectors for update, reset, and hidden gates

Rb[zrh] - R bias vectors for update, reset, and hidden gates

WB[zrh] - W parameter weight matrix for backward update, reset, and hidden gates

RB[zrh] - R recurrence weight matrix for backward update, reset, and hidden gates

WBb[zrh] - W bias vectors for backward update, reset, and hidden gates

RBb[zrh] - R bias vectors for backward update, reset, and hidden gates

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x)                - max(0, x)

Tanh(x)                - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x)             - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x)              - alpha*x + beta

LeakyRelu(x)           - x if x >= 0 else alpha * x

ThresholdedRelu(x)     - x if x >= alpha else 0

ScaledTanh(x)          - alpha*Tanh(beta*x)

HardSigmoid(x)         - min(max(alpha*x + beta, 0), 1)

Elu(x)                 - x if x >= 0 else alpha*(e^x - 1)

Softsign(x)            - x/(1 + |x|)

Softplus(x)            - log(1 + e^x)

Equations (Default: f=Sigmoid, g=Tanh):

- zt = f(Xt*(Wz^T) + Ht-1*(Rz^T) + Wbz + Rbz)

- rt = f(Xt*(Wr^T) + Ht-1*(Rr^T) + Wbr + Rbr)

- ht = g(Xt*(Wh^T) + (rt (.) Ht-1)*(Rh^T) + Rbh + Wbh) # default, when linear_before_reset = 0

- ht = g(Xt*(Wh^T) + (rt (.) (Ht-1*(Rh^T) + Rbh)) + Wbh) # when linear_before_reset != 0

- Ht = (1 - zt) (.) ht + zt (.) Ht-1

This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: GRU-1, GRU-3

Attributes

activation_alpha : list of floats

Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.

activation_beta : list of floats

Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.

activations : list of strings

A list of 2 (or 4 if bidirectional) activation functions for update, reset, and hidden gates. The activation functions must be one of the activation functions specified above. Optional: See the equations for default if not specified.

clip : float

Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.

direction : string (default is forward)

Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.

hidden_size : int

Number of neurons in the hidden layer

linear_before_reset : int (default is 0)

When computing the output of the hidden gate, apply the linear transformation before multiplying by the output of the reset gate.

Inputs (3 - 6)

X : T

The input sequences packed (and potentially padded) into one 3-D tensor with the shape of `[seq_length, batch_size, input_size]`.

W : T

The weight tensor for the gates. Concatenation of `W[zrh]` and `WB[zrh]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 3*hidden_size, input_size]`.

R : T

The recurrence weight tensor. Concatenation of `R[zrh]` and `RB[zrh]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 3*hidden_size, hidden_size]`.

B (optional) : T

The bias tensor for the gates. Concatenation of `[Wb[zrh], Rb[zrh]]` and `[WBb[zrh], RBb[zrh]]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 6*hidden_size]`. Optional: If not specified - assumed to be 0

sequence_lens (optional) : T1

Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length `seq_length`. It has shape `[batch_size]`.

initial_h (optional) : T

Optional initial value of the hidden. If not specified - assumed to be 0. It has shape `[num_directions, batch_size, hidden_size]`.

Outputs (0 - 2)

Y (optional) : T

A tensor that concats all the intermediate output values of the hidden. It has shape `[seq_length, num_directions, batch_size, hidden_size]`.

Y_h (optional) : T

The last output value of the hidden. It has shape `[num_directions, batch_size, hidden_size]`.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

T1 : tensor(int32)

Constrain seq_lens to integer tensor.

Examples

defaults

input = np.array([[[1., 2.], [3., 4.], [5., 6.]]]).astype(np.float32)

input_size = 2
hidden_size = 5
weight_scale = 0.1
number_of_gates = 3

node = onnx.helper.make_node(
    'GRU',
    inputs=['X', 'W', 'R'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

gru = GRU_Helper(X=input, W=W, R=R)
_, Y_h = gru.step()
expect(node, inputs=[input, W, R], outputs=[Y_h.astype(np.float32)], name='test_gru_defaults')

initial_bias

input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]).astype(np.float32)

input_size = 3
hidden_size = 3
weight_scale = 0.1
custom_bias = 0.1
number_of_gates = 3

node = onnx.helper.make_node(
    'GRU',
    inputs=['X', 'W', 'R', 'B'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, number_of_gates * hidden_size)).astype(np.float32)
R_B = np.zeros((1, number_of_gates * hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

gru = GRU_Helper(X=input, W=W, R=R, B=B)
_, Y_h = gru.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_gru_with_initial_bias')

seq_length

input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]],
                  [[10., 11., 12.], [13., 14., 15.], [16., 17., 18.]]]).astype(np.float32)

input_size = 3
hidden_size = 5
number_of_gates = 3

node = onnx.helper.make_node(
    'GRU',
    inputs=['X', 'W', 'R', 'B'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = np.random.randn(1, number_of_gates * hidden_size, input_size).astype(np.float32)
R = np.random.randn(1, number_of_gates * hidden_size, hidden_size).astype(np.float32)

# Adding custom bias
W_B = np.random.randn(1, number_of_gates * hidden_size).astype(np.float32)
R_B = np.random.randn(1, number_of_gates * hidden_size).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

gru = GRU_Helper(X=input, W=W, R=R, B=B)
_, Y_h = gru.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_gru_seq_length')

Gather

Given data tensor of rank r >= 1, and indices tensor of rank q, gather entries of the axis dimension of data (by default outer-most one as axis=0) indexed by indices, and concatenates them in an output tensor of rank q + (r - 1). Example 1: data = [ [1.0, 1.2], [2.3, 3.4], [4.5, 5.7], ] indices = [ [0, 1], [1, 2], ] output = [ [ [1.0, 1.2], [2.3, 3.4], ], [ [2.3, 3.4], [4.5, 5.7], ], ] Example 2: data = [ [1.0, 1.2, 1.9], [2.3, 3.4, 3.9], [4.5, 5.7, 5.9], ] indices = [ [0, 2], ] axis = 1, output = [ [ [1.0, 1.9], [2.3, 3.9], [4.5, 5.9], ], ]

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is 0)

Which axis to gather on. Negative value means counting dimensions from the back. Accepted range in [-r, r-1]

Inputs

data : T

Tensor of rank r >= 1.

indices : Tind

Tensor of int32/int64 indices, of any rank q.

Outputs

output : T

Tensor of rank q + (r - 1).

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to any tensor type.

Tind : tensor(int32), tensor(int64)

Constrain indices to integer types

Examples

gather_0

node = onnx.helper.make_node(
    'Gather',
    inputs=['data', 'indices'],
    outputs=['y'],
    axis=0,
)
data = np.random.randn(5, 4, 3, 2).astype(np.float32)
indices = np.array([0, 1, 3])
y = np.take(data, indices, axis=0)

expect(node, inputs=[data, indices.astype(np.int64)], outputs=[y],
       name='test_gather_0')

gather_1

node = onnx.helper.make_node(
    'Gather',
    inputs=['data', 'indices'],
    outputs=['y'],
    axis=1,
)
data = np.random.randn(5, 4, 3, 2).astype(np.float32)
indices = np.array([0, 1, 3])
y = np.take(data, indices, axis=1)

expect(node, inputs=[data, indices.astype(np.int64)], outputs=[y],
       name='test_gather_1')

Gemm

General Matrix multiplication: https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3

A' = transpose(A) if transA else A

B' = transpose(B) if transB else B

Compute Y = alpha * A' * B' + beta * C, where input tensor A has shape (M, K) or (K, M), input tensor B has shape (K, N) or (N, K), input tensor C is broadcastable to shape (M, N), and output tensor Y has shape (M, N). A will be transposed before doing the computation if attribute transA is non-zero, same for B and transB. This operator supports unidirectional broadcasting (tensor C should be unidirectional broadcastable to tensor A * B); for more details please check the doc.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Gemm-1, Gemm-6, Gemm-7

Attributes

alpha : float (default is 1.0)

Scalar multiplier for the product of input tensors A * B.

beta : float (default is 1.0)

Scalar multiplier for input tensor C.

transA : int (default is 0)

Whether A should be transposed

transB : int (default is 0)

Whether B should be transposed

Inputs

A : T

Input tensor A. The shape of A should be (M, K) if transA is 0, or (K, M) if transA is non-zero.

B : T

Input tensor B. The shape of B should be (K, N) if transB is 0, or (N, K) if transB is non-zero.

C : T

Input tensor C. The shape of C should be unidirectional broadcastable to (M, N).

Outputs

Y : T

Output tensor of shape (M, N).

Type Constraints

T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)

Constrain input and output types to float/int tensors.

Examples

notranspose

node = onnx.helper.make_node(
    'Gemm',
    inputs=['a', 'b', 'c'],
    outputs=['y'],
    alpha=0.5,
    beta=0.5
)
a = np.random.ranf([3, 6]).astype(np.float32)
b = np.random.ranf([6, 4]).astype(np.float32)
c = np.random.ranf([3, 4]).astype(np.float32)
y = 0.5 * np.dot(a, b) + 0.5 * c
expect(node, inputs=[a, b, c], outputs=[y],
       name='test_gemm_nobroadcast')

transpose

node = onnx.helper.make_node(
    'Gemm',
    inputs=['a', 'b', 'c'],
    outputs=['y'],
    alpha=0.5,
    beta=0.5,
    transA=1,
    transB=1
)
a = np.random.ranf([6, 3]).astype(np.float32)
b = np.random.ranf([4, 6]).astype(np.float32)
c = np.random.ranf([1, 1]).astype(np.float32)
y = 0.5 * np.dot(a.T, b.T) + 0.5 * c
expect(node, inputs=[a, b, c], outputs=[y],
       name='test_gemm_broadcast')

GlobalAveragePool

GlobalAveragePool consumes an input tensor X and applies average pooling across the values in the same channel. This is equivalent to AveragePool with kernel size equal to the spatial dimension of input tensor.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size.

Outputs

Y : T

Output data tensor from pooling across the input tensor. Dimensions will be N x C x 1 x 1

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

globalaveragepool

node = onnx.helper.make_node(
    'GlobalAveragePool',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(1, 3, 5, 5).astype(np.float32)
spatial_shape = np.ndim(x) - 2
y = np.average(x, axis=tuple(range(spatial_shape, spatial_shape + 2)))
for _ in range(spatial_shape):
    y = np.expand_dims(y, -1)
expect(node, inputs=[x], outputs=[y], name='test_globalaveragepool')

globalaveragepool_precomputed

node = onnx.helper.make_node(
    'GlobalAveragePool',
    inputs=['x'],
    outputs=['y'],
)
x = np.array([[[
    [1, 2, 3],
    [4, 5, 6],
    [7, 8, 9],
]]]).astype(np.float32)
y = np.array([[[[5]]]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y], name='test_globalaveragepool_precomputed')

GlobalLpPool

GlobalLpPool consumes an input tensor X and applies lp pool pooling across the values in the same channel. This is equivalent to LpPool with kernel size equal to the spatial dimension of input tensor.

Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: GlobalLpPool-1

Attributes

p : int (default is 2)

p value of the Lp norm used to pool over the input data.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size.

Outputs

Y : T

Output data tensor from pooling across the input tensor. Dimensions will be N x C x 1 x 1

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

GlobalMaxPool

GlobalMaxPool consumes an input tensor X and applies max pooling across the values in the same channel. This is equivalent to MaxPool with kernel size equal to the spatial dimension of input tensor.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size.

Outputs

Y : T

Output data tensor from pooling across the input tensor. Dimensions will be N x C x 1 x 1

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

globalmaxpool

node = onnx.helper.make_node(
    'GlobalMaxPool',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(1, 3, 5, 5).astype(np.float32)
spatial_shape = np.ndim(x) - 2
y = np.max(x, axis=tuple(range(spatial_shape, spatial_shape + 2)))
for _ in range(spatial_shape):
    y = np.expand_dims(y, -1)
expect(node, inputs=[x], outputs=[y], name='test_globalmaxpool')

globalmaxpool_precomputed

node = onnx.helper.make_node(
    'GlobalMaxPool',
    inputs=['x'],
    outputs=['y'],
)
x = np.array([[[
    [1, 2, 3],
    [4, 5, 6],
    [7, 8, 9],
]]]).astype(np.float32)
y = np.array([[[[9]]]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y], name='test_globalmaxpool_precomputed')

Greater

Returns the tensor resulted from performing the greater logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Greater-1, Greater-7

Inputs

A : T

First input operand for the logical operator.

B : T

Second input operand for the logical operator.

Outputs

C : T1

Result tensor.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrains input to float tensors.

T1 : tensor(bool)

Constrains output to boolean tensor.

Examples

greater

node = onnx.helper.make_node(
    'Greater',
    inputs=['x', 'y'],
    outputs=['greater'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = np.greater(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_greater')

greater_broadcast

node = onnx.helper.make_node(
    'Greater',
    inputs=['x', 'y'],
    outputs=['greater'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = np.greater(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_greater_bcast')

HardSigmoid

HardSigmoid takes one input data (Tensor) and produces one output data (Tensor) where the HardSigmoid function, y = max(0, min(1, alpha * x + beta)), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: HardSigmoid-1

Attributes

alpha : float (default is 0.2)

Value of alpha.

beta : float (default is 0.5)

Value of beta.

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

hardsigmoid

node = onnx.helper.make_node(
    'HardSigmoid',
    inputs=['x'],
    outputs=['y'],
    alpha=0.5,
    beta=0.6
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.clip(x * 0.5 + 0.6, 0, 1)  # expected output [0.1, 0.6, 1.]
expect(node, inputs=[x], outputs=[y],
       name='test_hardsigmoid_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x * 0.5 + 0.6, 0, 1)
expect(node, inputs=[x], outputs=[y],
       name='test_hardsigmoid')

hardsigmoid_default

default_alpha = 0.2
default_beta = 0.5
node = onnx.helper.make_node(
    'HardSigmoid',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x * default_alpha + default_beta, 0, 1)
expect(node, inputs=[x], outputs=[y],
       name='test_hardsigmoid_default')

Hardmax

The operator computes the hardmax (1 for the first maximum value, and 0 for all others) values for each layer in the batch of the given input. The input is a 2-D tensor (Tensor) of size (batch_size x input_feature_dimensions). The output tensor has the same shape and contains the hardmax values of the corresponding input.

Input does not need to explicitly be a 2D vector; rather, it will be coerced into one. For an arbitrary n-dimensional tensor input \in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis provided, then input will be coerced into a 2-dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}]. For the default case where axis=1, this means the input tensor will be coerced into a 2D tensor of dimensions [a_0, a_1 * ... * a_{n-1}], where a_0 is often the batch size. In this situation, we must have a_0 = N and a_1 * ... * a_{n-1} = D. Each of these dimensions must be matched correctly, or else the operator will throw errors.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is 1)

Describes the axis of the inputs when coerced to 2D; defaults to one because the 0th axis most likely describes the batch_size

Inputs

input : T

The input tensor that's coerced into a 2D matrix of size (NxD) as described above.

Outputs

output : T

The output values with the same shape as input tensor (the original size without coercion).

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

hardmax

node = onnx.helper.make_node(
    'Hardmax',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([[3, 0, 1, 2], [2, 5, 1, 0], [0, 1, 3, 2], [0, 1, 2, 3]]).astype(np.float32)
y = np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_hardmax_example')

# For multiple occurrances of the maximal values, the first occurrence is selected for one-hot output
x = np.array([[3, 3, 3, 1]]).astype(np.float32)
y = np.array([[1, 0, 0, 0]]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_hardmax_one_hot')

hardmax_axis

def hardmax_2d(x):  # type: (np.ndarray) -> np.ndarray
    return np.eye(x.shape[1], dtype=x.dtype)[np.argmax(x, axis=1)]

x = np.random.randn(3, 4, 5).astype(np.float32)
node = onnx.helper.make_node(
    'Hardmax',
    inputs=['x'],
    outputs=['y'],
    axis=0,
)
y = hardmax_2d(x.reshape(1, 60)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_hardmax_axis_0')

node = onnx.helper.make_node(
    'Hardmax',
    inputs=['x'],
    outputs=['y'],
    axis=1,
)
y = hardmax_2d(x.reshape(3, 20)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_hardmax_axis_1')

# default axis is 1
node = onnx.helper.make_node(
    'Hardmax',
    inputs=['x'],
    outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
       name='test_hardmax_default_axis')

node = onnx.helper.make_node(
    'Hardmax',
    inputs=['x'],
    outputs=['y'],
    axis=2,
)
y = hardmax_2d(x.reshape(12, 5)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_hardmax_axis_2')

Identity

Identity operator

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

Tensor to copy input into.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

identity

node = onnx.helper.make_node(
    'Identity',
    inputs=['x'],
    outputs=['y'],
)

data = np.array([[[
    [1, 2],
    [3, 4],
]]], dtype=np.float32)

expect(node, inputs=[data], outputs=[data],
       name='test_identity')

If

If conditional

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

else_branch : graph (required)

Graph to run if condition is false. Has N outputs: values you wish to be live-out to the enclosing scope. The number of outputs must match the number of outputs in the then_branch.

then_branch : graph (required)

Graph to run if condition is true. Has N outputs: values you wish to be live-out to the enclosing scope. The number of outputs must match the number of outputs in the else_branch.

Inputs

cond : B

Condition for the if

Outputs (1 - ∞)

outputs (variadic, heterogeneous) : V

Values that are live-out to the enclosing scope. The return values in the `then_branch` and `else_branch` must be of the same shape and same data type.

Type Constraints

V : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

All Tensor types

B : tensor(bool)

Only bool

InstanceNormalization

Carries out instance normalization as described in the paper https://arxiv.org/abs/1607.08022.

y = scale * (x - mean) / sqrt(variance + epsilon) + B, where mean and variance are computed per instance per channel.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: InstanceNormalization-1

Attributes

epsilon : float (default is 1e-05)

The epsilon value to use to avoid division by zero.

Inputs

input : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size.

scale : T

The input 1-dimensional scale tensor of size C.

B : T

The input 1-dimensional bias tensor of size C.

Outputs

output : T

The output tensor of the same shape as input.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

instancenormalization

def _instancenorm_test_mode(x, s, bias, epsilon=1e-5):  # type: ignore
    dims_x = len(x.shape)
    axis = tuple(range(2, dims_x))
    mean = np.mean(x, axis=axis, keepdims=True)
    var = np.var(x, axis=axis, keepdims=True)
    dim_ones = (1,) * (dims_x - 2)
    s = s.reshape(-1, *dim_ones)
    bias = bias.reshape(-1, *dim_ones)
    return s * (x - mean) / np.sqrt(var + epsilon) + bias

# input size: (1, 2, 1, 3)
x = np.array([[[[-1, 0, 1]], [[2, 3, 4]]]]).astype(np.float32)
s = np.array([1.0, 1.5]).astype(np.float32)
bias = np.array([0, 1]).astype(np.float32)
y = _instancenorm_test_mode(x, s, bias).astype(np.float32)

node = onnx.helper.make_node(
    'InstanceNormalization',
    inputs=['x', 's', 'bias'],
    outputs=['y'],
)

# output size: (1, 2, 1, 3)
expect(node, inputs=[x, s, bias], outputs=[y],
       name='test_instancenorm_example')

# input size: (2, 3, 4, 5)
x = np.random.randn(2, 3, 4, 5).astype(np.float32)
s = np.random.randn(3).astype(np.float32)
bias = np.random.randn(3).astype(np.float32)
epsilon = 1e-2
y = _instancenorm_test_mode(x, s, bias, epsilon).astype(np.float32)

node = onnx.helper.make_node(
    'InstanceNormalization',
    inputs=['x', 's', 'bias'],
    outputs=['y'],
    epsilon=epsilon,
)

# output size: (2, 3, 4, 5)
expect(node, inputs=[x, s, bias], outputs=[y],
       name='test_instancenorm_epsilon')

IsNaN

Returns which elements of the input are NaN.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

X : T1

input

Outputs

Y : T2

output

Type Constraints

T1 : tensor(float16), tensor(float), tensor(double)

Constrain input types to float tensors.

T2 : tensor(bool)

Constrain output types to boolean tensors.

Examples

isnan

node = onnx.helper.make_node(
    'IsNaN',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([3.0, np.nan, 4.0, np.nan], dtype=np.float32)
y = np.isnan(x)
expect(node, inputs=[x], outputs=[y], name='test_isnan')

LRN

Local Response Normalization proposed in the AlexNet paper. It normalizes over local input regions. The local region is defined across the channels. For an element X[n, c, d1, ..., dk] in a tensor of shape (N x C x D1 x D2, ..., Dk), its region is {X[n, i, d1, ..., dk] | max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2))}.

square_sum[n, c, d1, ..., dk] = sum(X[n, i, d1, ..., dk] ^ 2), where max(0, c - floor((size - 1) / 2)) <= i <= min(C - 1, c + ceil((size - 1) / 2)).

Y[n, c, d1, ..., dk] = X[n, c, d1, ..., dk] / (bias + alpha / size * square_sum[n, c, d1, ..., dk] ) ^ beta

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

alpha : float (default is 0.0001)

Scaling parameter.

beta : float (default is 0.75)

The exponent.

bias : float (default is 1.0)

size : int (required)

The number of channels to sum over

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].

Outputs

Y : T

Output tensor, which has the shape and type as input tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

default

alpha = 0.0001
beta = 0.75
bias = 1.0
nsize = 3
node = onnx.helper.make_node(
    'LRN',
    inputs=['x'],
    outputs=['y'],
    size=3
)
x = np.random.randn(5, 5, 5, 5).astype(np.float32)
square_sum = np.zeros((5, 5, 5, 5)).astype(np.float32)
for n, c, h, w in np.ndindex(x.shape):
    square_sum[n, c, h, w] = sum(x[n,
                                   max(0, c - int(math.floor((nsize - 1) / 2))):min(5, c + int(math.ceil((nsize - 1) / 2)) + 1),
                                   h,
                                   w] ** 2)
y = x / ((bias + (alpha / nsize) * square_sum) ** beta)
expect(node, inputs=[x], outputs=[y],
       name='test_lrn_default')

lrn

alpha = 0.0002
beta = 0.5
bias = 2.0
nsize = 3
node = onnx.helper.make_node(
    'LRN',
    inputs=['x'],
    outputs=['y'],
    alpha=alpha,
    beta=beta,
    bias=bias,
    size=nsize
)
x = np.random.randn(5, 5, 5, 5).astype(np.float32)
square_sum = np.zeros((5, 5, 5, 5)).astype(np.float32)
for n, c, h, w in np.ndindex(x.shape):
    square_sum[n, c, h, w] = sum(x[n,
                                   max(0, c - int(math.floor((nsize - 1) / 2))):min(5, c + int(math.ceil((nsize - 1) / 2)) + 1),
                                   h,
                                   w] ** 2)
y = x / ((bias + (alpha / nsize) * square_sum) ** beta)
expect(node, inputs=[x], outputs=[y],
       name='test_lrn')

LSTM

Computes an one-layer LSTM. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

i - input gate

o - output gate

f - forget gate

c - cell gate

t - time step (t-1 means previous time step)

W[iofc] - W parameter weight matrix for input, output, forget, and cell gates

R[iofc] - R recurrence weight matrix for input, output, forget, and cell gates

Wb[iofc] - W bias vectors for input, output, forget, and cell gates

Rb[iofc] - R bias vectors for input, output, forget, and cell gates

P[iof] - P peephole weight vector for input, output, and forget gates

WB[iofc] - W parameter weight matrix for backward input, output, forget, and cell gates

RB[iofc] - R recurrence weight matrix for backward input, output, forget, and cell gates

WBb[iofc] - W bias vectors for backward input, output, forget, and cell gates

RBb[iofc] - R bias vectors for backward input, output, forget, and cell gates

PB[iof] - P peephole weight vector for backward input, output, and forget gates

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x)                - max(0, x)

Tanh(x)                - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x)             - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x)              - alpha*x + beta

LeakyRelu(x)           - x if x >= 0 else alpha * x

ThresholdedRelu(x)     - x if x >= alpha else 0

ScaledTanh(x)          - alpha*Tanh(beta*x)

HardSigmoid(x)         - min(max(alpha*x + beta, 0), 1)

Elu(x)                 - x if x >= 0 else alpha*(e^x - 1)

Softsign(x)            - x/(1 + |x|)

Softplus(x)            - log(1 + e^x)

Equations (Default: f=Sigmoid, g=Tanh, h=Tanh):

- it = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Pi (.) Ct-1 + Wbi + Rbi)

- ft = f(Xt*(Wf^T) + Ht-1*(Rf^T) + Pf (.) Ct-1 + Wbf + Rbf)

- ct = g(Xt*(Wc^T) + Ht-1*(Rc^T) + Wbc + Rbc)

- Ct = ft (.) Ct-1 + it (.) ct

- ot = f(Xt*(Wo^T) + Ht-1*(Ro^T) + Po (.) Ct + Wbo + Rbo)

- Ht = ot (.) h(Ct)

This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: LSTM-1

Attributes

activation_alpha : list of floats

Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.

activation_beta : list of floats

Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.

activations : list of strings

A list of 3 (or 6 if bidirectional) activation functions for input, output, forget, cell, and hidden. The activation functions must be one of the activation functions specified above. Optional: See the equations for default if not specified.

clip : float

Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.

direction : string (default is forward)

Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.

hidden_size : int

Number of neurons in the hidden layer

input_forget : int (default is 0)

Couple the input and forget gates if 1.

Inputs (3 - 8)

X : T

The input sequences packed (and potentially padded) into one 3-D tensor with the shape of `[seq_length, batch_size, input_size]`.

W : T

The weight tensor for the gates. Concatenation of `W[iofc]` and `WB[iofc]` (if bidirectional) along dimension 0. The tensor has shape `[num_directions, 4*hidden_size, input_size]`.

R : T

The recurrence weight tensor. Concatenation of `R[iofc]` and `RB[iofc]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 4*hidden_size, hidden_size]`.

B (optional) : T

The bias tensor for input gate. Concatenation of `[Wb[iofc], Rb[iofc]]`, and `[WBb[iofc], RBb[iofc]]` (if bidirectional) along dimension 0. This tensor has shape `[num_directions, 8*hidden_size]`. Optional: If not specified - assumed to be 0.

sequence_lens (optional) : T1

Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length `seq_length`. It has shape `[batch_size]`.

initial_h (optional) : T

Optional initial value of the hidden. If not specified - assumed to be 0. It has shape `[num_directions, batch_size, hidden_size]`.

initial_c (optional) : T

Optional initial value of the cell. If not specified - assumed to be 0. It has shape `[num_directions, batch_size, hidden_size]`.

P (optional) : T

The weight tensor for peepholes. Concatenation of `P[iof]` and `PB[iof]` (if bidirectional) along dimension 0. It has shape `[num_directions, 3*hidde_size]`. Optional: If not specified - assumed to be 0.

Outputs (0 - 3)

Y (optional) : T

A tensor that concats all the intermediate output values of the hidden. It has shape `[seq_length, num_directions, batch_size, hidden_size]`.

Y_h (optional) : T

The last output value of the hidden. It has shape `[num_directions, batch_size, hidden_size]`.

Y_c (optional) : T

The last output value of the cell. It has shape `[num_directions, batch_size, hidden_size]`.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

T1 : tensor(int32)

Constrain seq_lens to integer tensor.

Examples

defaults

input = np.array([[[1., 2.], [3., 4.], [5., 6.]]]).astype(np.float32)

input_size = 2
hidden_size = 3
weight_scale = 0.1
number_of_gates = 4

node = onnx.helper.make_node(
    'LSTM',
    inputs=['X', 'W', 'R'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

lstm = LSTM_Helper(X=input, W=W, R=R)
_, Y_h = lstm.step()
expect(node, inputs=[input, W, R], outputs=[Y_h.astype(np.float32)], name='test_lstm_defaults')

initial_bias

input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]).astype(np.float32)

input_size = 3
hidden_size = 4
weight_scale = 0.1
custom_bias = 0.1
number_of_gates = 4

node = onnx.helper.make_node(
    'LSTM',
    inputs=['X', 'W', 'R', 'B'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, number_of_gates * hidden_size)).astype(np.float32)
R_B = np.zeros((1, number_of_gates * hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), 1)

lstm = LSTM_Helper(X=input, W=W, R=R, B=B)
_, Y_h = lstm.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_lstm_with_initial_bias')

peepholes

input = np.array([[[1., 2., 3., 4.], [5., 6., 7., 8.]]]).astype(np.float32)

input_size = 4
hidden_size = 3
weight_scale = 0.1
number_of_gates = 4
number_of_peepholes = 3

node = onnx.helper.make_node(
    'LSTM',
    inputs=['X', 'W', 'R', 'B', 'sequence_lens', 'initial_h', 'initial_c', 'P'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

# Initializing Inputs
W = weight_scale * np.ones((1, number_of_gates * hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, number_of_gates * hidden_size, hidden_size)).astype(np.float32)
B = np.zeros((1, 2 * number_of_gates * hidden_size)).astype(np.float32)
seq_lens = np.repeat(input.shape[0], input.shape[1]).astype(np.int32)
init_h = np.zeros((1, input.shape[1], hidden_size)).astype(np.float32)
init_c = np.zeros((1, input.shape[1], hidden_size)).astype(np.float32)
P = weight_scale * np.ones((1, number_of_peepholes * hidden_size)).astype(np.float32)

lstm = LSTM_Helper(X=input, W=W, R=R, B=B, P=P, initial_c=init_c, initial_h=init_h)
_, Y_h = lstm.step()
expect(node, inputs=[input, W, R, B, seq_lens, init_h, init_c, P], outputs=[Y_h.astype(np.float32)],
       name='test_lstm_with_peepholes')

LeakyRelu

LeakyRelu takes input data (Tensor) and an argument alpha, and produces one output data (Tensor) where the function f(x) = alpha * x for x < 0, f(x) = x for x >= 0, is applied to the data tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: LeakyRelu-1

Attributes

alpha : float (default is 0.01)

Coefficient of leakage.

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

leakyrelu

node = onnx.helper.make_node(
    'LeakyRelu',
    inputs=['x'],
    outputs=['y'],
    alpha=0.1
)

x = np.array([-1, 0, 1]).astype(np.float32)
# expected output [-0.1, 0., 1.]
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * 0.1
expect(node, inputs=[x], outputs=[y],
       name='test_leakyrelu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * 0.1
expect(node, inputs=[x], outputs=[y],
       name='test_leakyrelu')

leakyrelu_default

default_alpha = 0.01
node = onnx.helper.make_node(
    'LeakyRelu',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * default_alpha
expect(node, inputs=[x], outputs=[y],
       name='test_leakyrelu_default')

Less

Returns the tensor resulted from performing the less logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Less-1, Less-7

Inputs

A : T

First input operand for the logical operator.

B : T

Second input operand for the logical operator.

Outputs

C : T1

Result tensor.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrains input to float tensors.

T1 : tensor(bool)

Constrains output to boolean tensor.

Examples

less

node = onnx.helper.make_node(
    'Less',
    inputs=['x', 'y'],
    outputs=['less'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = np.less(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_less')

less_broadcast

node = onnx.helper.make_node(
    'Less',
    inputs=['x', 'y'],
    outputs=['less'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = np.less(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_less_bcast')

Log

Calculates the natural log of the given input tensor, element-wise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Log-1

Inputs

input : T

Input tensor

Outputs

output : T

The natural log of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

log

node = onnx.helper.make_node(
    'Log',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([1, 10]).astype(np.float32)
y = np.log(x)  # expected output [0., 2.30258512]
expect(node, inputs=[x], outputs=[y],
       name='test_log_example')

x = np.exp(np.random.randn(3, 4, 5).astype(np.float32))
y = np.log(x)
expect(node, inputs=[x], outputs=[y],
       name='test_log')

LogSoftmax

The operator computes the logsoftmax (log of softmax) values for each layer in the batch of the given input. The input is a 2-D tensor (Tensor) of size (batch_size x input_feature_dimensions). The output tensor has the same shape and contains the logsoftmax values of the corresponding input.

Input does not need to explicitly be a 2D vector; rather, it will be coerced into one. For an arbitrary n-dimensional tensor input \in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis provided, then input will be coerced into a 2-dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}]. For the default case where axis=1, this means the input tensor will be coerced into a 2D tensor of dimensions [a_0, a_1 * ... * a_{n-1}], where a_0 is often the batch size. In this situation, we must have a_0 = N and a_1 * ... * a_{n-1} = D. Each of these dimensions must be matched correctly, or else the operator will throw errors.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is 1)

Describes the axis of the inputs when coerced to 2D; defaults to one because the 0th axis most likely describes the batch_size

Inputs

input : T

The input tensor that's coerced into a 2D matrix of size (NxD) as described above.

Outputs

output : T

The output values with the same shape as input tensor (the original size without coercion).

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

logsoftmax

node = onnx.helper.make_node(
    'LogSoftmax',
    inputs=['x'],
    outputs=['y'],
)
x = np.array([[-1, 0, 1]]).astype(np.float32)
# expected output [[-2.40760589, -1.40760589, -0.40760589]]
y = x - np.log(np.sum(np.exp(x), axis=1))
expect(node, inputs=[x], outputs=[y],
       name='test_logsoftmax_example_1')

logsoftmax_axis

def logsoftmax_2d(x):  # type: (np.ndarray) -> np.ndarray
    max_x = np.max(x, axis=1).reshape((-1, 1))
    exp_x = np.exp(x - max_x)
    return x - max_x - np.log(np.sum(exp_x, axis=1).reshape((-1, 1)))

x = np.array([[0, 1, 2, 3], [10000, 10001, 10002, 10003]]).astype(np.float32)
# expected output [[-3.4401896, -2.4401896, -1.44018972, -0.44018969],
#                 [-3.4401896, -2.4401896, -1.44018972, -0.44018969]]
y = logsoftmax_2d(x)

node = onnx.helper.make_node(
    'LogSoftmax',
    inputs=['x'],
    outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
       name='test_logsoftmax_large_number')

x = np.abs(np.random.randn(3, 4, 5).astype(np.float32))
node = onnx.helper.make_node(
    'LogSoftmax',
    inputs=['x'],
    outputs=['y'],
    axis=0,
)
y = logsoftmax_2d(x.reshape(1, 60)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_logsoftmax_axis_0')

node = onnx.helper.make_node(
    'LogSoftmax',
    inputs=['x'],
    outputs=['y'],
    axis=1,
)
y = logsoftmax_2d(x.reshape(3, 20)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_logsoftmax_axis_1')

# default axis is 1
node = onnx.helper.make_node(
    'LogSoftmax',
    inputs=['x'],
    outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
       name='test_logsoftmax_default_axis')

node = onnx.helper.make_node(
    'LogSoftmax',
    inputs=['x'],
    outputs=['y'],
    axis=2,
)
y = logsoftmax_2d(x.reshape(12, 5)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_logsoftmax_axis_2')

Loop

Generic Looping construct. This loop has multiple termination conditions:

1. Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M.
2. Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not.

This table summarizes the operating modes of this operator with equivalent C-style code:

  Operator inputs defined as (max_trip_count, condition_var).

  input ("", ""):
      for (int i=0; ; ++i) {
        cond = ... // Note this value is ignored, but is required in the body
      }

  input ("", cond) // Note this is analogous to a while loop
      bool cond = ...;
      for (int i=0; cond; ++i) {
        cond = ...;
      }

  input ("", 1) // Note this is analogous to a do-while loop
      bool cond = true
      for (int i=0; cond; ++i) {
        cond = ...;
      }

  input (trip_count, "") // Note this is analogous to a for loop
      int trip_count = ...
      for (int i=0; i < trip_count; ++i) {
        cond = ...; // ignored
      }

  input (trip_count, cond)
      int trip_count = ...;
      bool cond = ...;
      for (int i=0; i < trip_count && cond; ++i) {
        cond = ...;
      }

Sample usage - cond as well as trip count

  graph predict-net {
    %a = Constant[value = <Scalar Tensor [3]>]()
    %b = Constant[value = <Scalar Tensor [6]>]()
    %keepgoing = Constant[value = <Scalar Tensor [1]>]()
    %max_trip_count = Constant[value = <Scalar Tensor [10]>]()
    %keepgoing_out, %b_out, %user_defined_vals = Loop[body = <graph body-net>](%max_trip_count, %keepgoing, %b)
    return
  }

  graph body-net (
    %i[INT32, scalar]
    %keepgoing[BOOL, scalar]
    %b[INT32, scalar]
  ) {
    %my_local = Add(%a, %b)
    %b_out = Sub(%a, %b)
    %keepgoing_out = Greater(%my_local, %b_out)
    %user_defined_vals = Add(%b, %b)
    return %keepgoing_out, %b_out, %user_defined_vals
  }

Sample equivalent C code

  {
    /* User-defined code (enclosing scope) */
    int a = 3, b = 6;
    bool keepgoing = true; // Analogous to input cond
    /* End user-defined code */

    /* Implicitly-defined code */
    const int max_trip_count = 10; // Analogous to input M
    int user_defined_vals[]; // Imagine this is resizable
    /* End implicitly-defined code */
    for (int i=0; i < max_trip_count && keepgoing; ++i) {
      /* User-defined code (loop body) */
      int my_local = a + b; // Reading values in the enclosing scope is fine
      b = a - b; // writes fine if we specify b as a loop-carried dependency
      keepgoing = my_local > b; // keepgoing is a loop-carried dependency
      user_defined_vals[i] = b + b;
      /* End user-defined code */
    }
    // my_local = 123; // Can't do this. my_local was defined in the the body

    // These below values are live-out from the loop and therefore accessible
    b_out; user_defined_vals; keepgoing_out;
  }

There are several things of note in this code snippet:

1. Values from the enclosing scope (i.e. variable a here) are in scope and can be referenced in the inputs of the loop.
2. Any variables which you wish to make available in the enclosing scope (i.e. the variables b and keepgoing) must be declared as either loop-carried dependencies (both at the op inputs and output and at the body net input and output) or scan_outputs.
3. Values created in the body cannot be accessed in the enclosing scope.

Note that the semantics of this op support "diagonal" or "wavefront" execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer).

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

body : graph (required)

The graph run each iteration. It has 2+N inputs: (iteration_num, condition, loop carried dependencies...). It has 1+N+K outputs: (condition, loop carried dependencies..., scan_outputs...). Each scan_output is created by concatenating the value of the specified output value at the end of each iteration of the loop. It is an error if the dimensions or data type of these scan_outputs change across loop iterations.

Inputs (3 - ∞)

M : I

A maximum trip-count for the loop specified at runtime. Optional. pass empty string to skip.

cond : B

A boolean termination condition. Pass empty string to skip.

v_initial (variadic, heterogeneous) : V

The initial values of any loop-carried dependencies (values that change across loop iterations)

Outputs (1 - ∞)

v_final_and_scan_outputs (variadic, heterogeneous) : V

Final N loop carried dependency values then K scan_outputs

Type Constraints

V : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

All Tensor types

I : int64

Only int64

B : bool

Only bool

LpNormalization

Given a matrix, apply Lp-normalization along the provided axis.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is -1)

The axis on which to apply normalization, -1 mean last axis.

p : int (default is 2)

The order of the normalization, only 1 or 2 are supported.

Inputs

input : T

Input matrix

Outputs

output : T

Matrix after normalization

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

LpPool

LpPool consumes an input tensor X and applies Lp pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. Lp pooling consisting of computing the Lp norm on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing.

Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: LpPool-1

Attributes

auto_pad : string (default is NOTSET)

auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.

kernel_shape : list of ints (required)

The size of the kernel along each axis.

p : int (default is 2)

p value of the Lp norm used to pool over the input data.

pads : list of ints

Padding for the beginning and ending along each axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each axis.

strides : list of ints

Stride along each axis.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size.

Outputs

Y : T

Output data tensor from Lp pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

MatMul

Matrix product that behaves like numpy.matmul: https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.matmul.html

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: MatMul-1

Inputs

A : T

N-dimensional matrix A

B : T

N-dimensional matrix B

Outputs

Y : T

Matrix multiply results from A * B

Type Constraints

T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)

Constrain input and output types to float/int tensors.

Examples

matmul

node = onnx.helper.make_node(
    'MatMul',
    inputs=['a', 'b'],
    outputs=['c'],
)

# 2d
a = np.random.randn(3, 4).astype(np.float32)
b = np.random.randn(4, 3).astype(np.float32)
c = np.matmul(a, b)
expect(node, inputs=[a, b], outputs=[c],
       name='test_matmul_2d')

# 3d
a = np.random.randn(2, 3, 4).astype(np.float32)
b = np.random.randn(2, 4, 3).astype(np.float32)
c = np.matmul(a, b)
expect(node, inputs=[a, b], outputs=[c],
       name='test_matmul_3d')

# 4d
a = np.random.randn(1, 2, 3, 4).astype(np.float32)
b = np.random.randn(1, 2, 4, 3).astype(np.float32)
c = np.matmul(a, b)
expect(node, inputs=[a, b], outputs=[c],
       name='test_matmul_4d')

Max

Element-wise max of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Max-1, Max-6

Inputs (1 - ∞)

data_0 (variadic) : T

List of tensors for max.

Outputs

max : T

Output tensor.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

max

data_0 = np.array([3, 2, 1]).astype(np.float32)
data_1 = np.array([1, 4, 4]).astype(np.float32)
data_2 = np.array([2, 5, 3]).astype(np.float32)
result = np.array([3, 5, 4]).astype(np.float32)
node = onnx.helper.make_node(
    'Max',
    inputs=['data_0', 'data_1', 'data_2'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
       name='test_max_example')

node = onnx.helper.make_node(
    'Max',
    inputs=['data_0'],
    outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
       name='test_max_one_input')

result = np.maximum(data_0, data_1)
node = onnx.helper.make_node(
    'Max',
    inputs=['data_0', 'data_1'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
       name='test_max_two_inputs')

MaxPool

MaxPool consumes an input tensor X and applies max pooling across the tensor according to kernel sizes, stride sizes, and pad lengths. max pooling consisting of computing the max on all values of a subset of the input tensor according to the kernel size and downsampling the data into the output tensor Y for further processing. The output spatial shape will be following:

output_spatial_shape[i] = floor((input_spatial_shape[i] + pad_shape[i] - kernel_spatial_shape[i]) / strides_spatial_shape[i] + 1)

* pad_shape[i] is sum of pads along axis i

auto_pad is a DEPRECATED attribute. If you are using them currently, the output spatial shape will be following:

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - kernel_spatial_shape[i] + 1) / strides_spatial_shape[i])
SAME_UPPER or SAME_LOWER: output_spatial_shape[i] = ceil(input_spatial_shape[i] / strides_spatial_shape[i])

And pad shape will be following if SAME_UPPER or SAME_LOWER:

pad_shape[i] = (output_spatial_shape[i] - 1) * strides_spatial_shape[i] + kernel_spatial_shape[i] - input_spatial_shape[i]

The output of each pooling window is maximum number of elements exclude pad.

Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: MaxPool-1

Attributes

auto_pad : string (default is NOTSET)

auto_pad must be either NOTSET, SAME_UPPER, SAME_LOWER or VALID. Where default value is NOTSET, which means explicit padding is used. SAME_UPPER or SAME_LOWER mean pad the input so that the output size match the input.In case of odd number add the extra padding at the end for SAME_UPPER and at the beginning for SAME_LOWER. VALID mean no padding. DEPRECATION NOTE: auto_pad is only intended to support legacy uses, and for framework authors, one is explicitly encouraged to use explicit padding specified in the pads attribute.

kernel_shape : list of ints (required)

The size of the kernel along each axis.

pads : list of ints

Padding for the beginning and ending along each axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each axis.

storage_order : int (default is 0)

The storage order of the tensor. 0 is row major, and 1 is column major.

strides : list of ints

Stride along each axis.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].

Outputs (1 - 2)

Y : T

Output data tensor from average or max pooling across the input tensor. Dimensions will vary based on various kernel, stride, and pad sizes. Floor value of the dimension is used

Indices (optional) : I

Indices tensor from max pooling across the input tensor. The dimensions of indices are the same as output tensor. The values in indices of are the indices of the selected values during pooling. The indices are computed as flatten 1-D tensor, and the indices do not consider padding. So the values in indices are in [0, N x C x D1 x ... x Dn).

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

I : tensor(int64)

Constrain index tensor to int64

Examples

maxpool_1d_default

"""
iutput_shape: [1, 3, 32]
output_shape: [1, 3, 31]
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2],
)
x = np.random.randn(1, 3, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = [2]
strides = [1]
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0], 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_1d_default')

maxpool_2d_default

"""
iutput_shape: [1, 3, 32, 32]
output_shape: [1, 3, 31, 31]
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_default')

maxpool_2d_pads

"""
iutput_shape: [1, 3, 28, 28]
output_shape: [1, 3, 30, 30]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[3, 3],
    pads=[2, 2, 2, 2]
)
x = np.random.randn(1, 3, 28, 28).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (3, 3)
strides = (1, 1)
pad_bottom = pad_top = pad_right = pad_left = 2
pad_shape = [pad_top + pad_bottom, pad_left + pad_right]
out_shape = get_output_shape('VALID', np.add(x_shape[2:], pad_shape), kernel_shape, strides)
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_pads')

maxpool_2d_precomputed_pads

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2]

)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[
    [13, 14, 15, 15, 15],
    [18, 19, 20, 20, 20],
    [23, 24, 25, 25, 25],
    [23, 24, 25, 25, 25],
    [23, 24, 25, 25, 25]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_precomputed_pads')

maxpool_2d_precomputed_same_upper

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 3, 3]
pad_shape: [2, 2] -> [1, 1, 1, 1] by axis
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[3, 3],
    strides=[2, 2],
    auto_pad='SAME_UPPER'
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[7, 9, 10],
                [17, 19, 20],
                [22, 24, 25]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_precomputed_same_upper')

maxpool_2d_precomputed_strides

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
    strides=[2, 2]
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[7, 9],
                [17, 19]]]]).astype(np.float32)

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_precomputed_strides')

maxpool_2d_same_lower

"""
iutput_shape: [1, 3, 32, 32]
output_shape: [1, 3, 32, 32]
pad_shape: [1, 1] -> [1, 0, 1, 0] by axis
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
    auto_pad='SAME_LOWER'
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape('SAME_LOWER', x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape('SAME_LOWER', x_shape[2:], kernel_shape, strides, out_shape)
pad_bottom = pad_shape[0] // 2
pad_top = pad_shape[0] - pad_bottom
pad_right = pad_shape[1] // 2
pad_left = pad_shape[1] - pad_right
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_same_lower')

maxpool_2d_same_upper

"""
iutput_shape: [1, 3, 32, 32]
output_shape: [1, 3, 32, 32]
pad_shape: [1, 1] -> [0, 1, 0, 1] by axis
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2],
    auto_pad='SAME_UPPER'
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (2, 2)
strides = (1, 1)
out_shape = get_output_shape('SAME_UPPER', x_shape[2:], kernel_shape, strides)
pad_shape = get_pad_shape('SAME_UPPER', x_shape[2:], kernel_shape, strides, out_shape)
pad_top = pad_shape[0] // 2
pad_bottom = pad_shape[0] - pad_top
pad_left = pad_shape[1] // 2
pad_right = pad_shape[1] - pad_left
padded = np.pad(x, ((0, 0), (0, 0), (pad_top, pad_bottom), (pad_left, pad_right)), mode='constant',
                constant_values=np.nan)
y = pool(padded, x_shape, kernel_shape, strides, out_shape, pad_shape, 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_same_upper')

maxpool_2d_strides

"""
iutput_shape: [1, 3, 32, 32]
output_shape: [1, 3, 10, 10]
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[5, 5],
    strides=[3, 3]
)
x = np.random.randn(1, 3, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = (5, 5)
strides = (3, 3)
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, (0, 0), 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_2d_strides')

maxpool_3d_default

"""
iutput_shape: [1, 3, 32, 32, 32]
output_shape: [1, 3, 31, 31, 31]
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y'],
    kernel_shape=[2, 2, 2],
)
x = np.random.randn(1, 3, 32, 32, 32).astype(np.float32)
x_shape = np.shape(x)
kernel_shape = [2, 2, 2]
strides = [1, 1, 1]
out_shape = get_output_shape('VALID', x_shape[2:], kernel_shape, strides)
padded = x
y = pool(padded, x_shape, kernel_shape, strides, out_shape, [0, 0, 0], 'MAX')

expect(node, inputs=[x], outputs=[y], name='test_maxpool_3d_default')

maxpool_with_argmax_2d_precomputed_pads

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 5, 5]
pad_shape: [4, 4] -> [2, 2, 2, 2] by axis
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y', 'z'],
    kernel_shape=[5, 5],
    pads=[2, 2, 2, 2]
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[
    [13, 14, 15, 15, 15],
    [18, 19, 20, 20, 20],
    [23, 24, 25, 25, 25],
    [23, 24, 25, 25, 25],
    [23, 24, 25, 25, 25]]]]).astype(np.float32)
z = np.array([[[
    [12, 13, 14, 14, 14],
    [17, 18, 19, 19, 19],
    [22, 23, 24, 24, 24],
    [22, 23, 24, 24, 24],
    [22, 23, 24, 24, 24]]]]).astype(np.int64)

expect(node, inputs=[x], outputs=[y, z], name='test_maxpool_with_argmax_2d_precomputed_pads')

maxpool_with_argmax_2d_precomputed_strides

"""
input_shape: [1, 1, 5, 5]
output_shape: [1, 1, 2, 2]
"""
node = onnx.helper.make_node(
    'MaxPool',
    inputs=['x'],
    outputs=['y', 'z'],
    kernel_shape=[2, 2],
    strides=[2, 2],
    storage_order=1
)
x = np.array([[[
    [1, 2, 3, 4, 5],
    [6, 7, 8, 9, 10],
    [11, 12, 13, 14, 15],
    [16, 17, 18, 19, 20],
    [21, 22, 23, 24, 25],
]]]).astype(np.float32)
y = np.array([[[[7, 9],
                [17, 19]]]]).astype(np.float32)
z = np.array([[[[6, 16],
                [8, 18]]]]).astype(np.int64)

expect(node, inputs=[x], outputs=[y, z], name='test_maxpool_with_argmax_2d_precomputed_strides')

MaxRoiPool

ROI max pool consumes an input tensor X and region of interests (RoIs) to apply max pooling across each RoI, to produce output 4-D tensor of shape (num_rois, channels, pooled_shape[0], pooled_shape[1]).

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

pooled_shape : list of ints (required)

ROI pool output shape (height, width).

spatial_scale : float (default is 1.0)

Multiplicative spatial scale factor to translate ROI coordinates from their input scale to the scale used when pooling.

Inputs

X : T

Input data tensor from the previous operator; dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data.

rois : T

RoIs (Regions of Interest) to pool over. Should be a 2-D tensor of shape (num_rois, 5) given as [[batch_id, x1, y1, x2, y2], ...].

Outputs

Y : T

RoI pooled output 4-D tensor of shape (num_rois, channels, pooled_shape[0], pooled_shape[1]).

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

MaxUnpool

MaxUnpool essentially computes the partial inverse of the MaxPool op. The input information to this op is typically the the output information from a MaxPool op. The first input tensor X is the tensor that needs to be unpooled, which is typically the pooled tensor (first output) from MaxPool. The second input tensor, I, contains the indices to the (locally maximal) elements corrsponding to the elements in the first input tensor X. Input tensor I is typically the second output of the MaxPool op. The third (optional) input is a tensor that specifies the output size of the unpooling operation.

MaxUnpool is intended to do 'partial' inverse of the MaxPool op. 'Partial' because all the non-maximal values from the original input to MaxPool are set to zero in the output of the MaxUnpool op. Pooling the result of an unpooling operation should give back the original input to the unpooling op.

MaxUnpool can produce the same output size for several input sizes, which makes unpooling op ambiguous. The third input argument, output_size, is meant to disambiguate the op and produce output tensor of known/predictable size.

In addition to the inputs, MaxUnpool takes three attributes, namely kernel_shape, strides, and pads, which define the exact unpooling op. The attributes typically have the same values as the corrsponding pooling op that the unpooling op is trying to invert.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Attributes

kernel_shape : list of ints (required)

The size of the kernel along each axis.

pads : list of ints

Padding for the beginning and ending along each axis, it can take any value greater than or equal to 0. The value represent the number of pixels added to the beginning and end part of the corresponding axis. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`. This attribute cannot be used simultaneously with auto_pad attribute. If not present, the padding defaults to 0 along start and end of each axis.

strides : list of ints

Stride along each axis.

Inputs (2 - 3)

X : T1

Input data tensor that has to be unpooled. This tensor is typically the first output of the MaxPool op.Dimensions for image case are (N x C x H x W), where N is the batch size, C is the number of channels, and H and W are the height and the width of the data. For non-image case, the dimensions are in the form of (N x C x D1 x D2 ... Dn), where N is the batch size. Optionally, if dimension denotation is in effect, the operation expects the input data tensor to arrive with the dimension denotation of [DATA_BATCH, DATA_CHANNEL, DATA_FEATURE, DATA_FEATURE ...].

I : T2

Input data tensor containing the indices corresponding to elements in the first input tensor X.This tensor is typically the second output of the MaxPool op.Dimensions must be the same as input tensor X. The indices are linear, i.e. computed considering the tensor as flattened 1-D tensor, assuming row-major storage. Also, the linear indices should not consider padding. So the values in indices are in the range [0, N x C x D1 x ... x Dn).

output_shape (optional) : T2

The shape of the output can be explicitly set which will cause pads values to be auto generated. If 'output_shape' is specified, 'pads' values are ignored.

Outputs

output : T1

Output data tensor that contains the result of the unpooling.

Type Constraints

T1 : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

T2 : tensor(int64)

Constrain index tensor to int64

Examples

with_output_shape

node = onnx.helper.make_node(
    'MaxUnpool',
    inputs=['xT', 'xI', 'output_shape'],
    outputs=['y'],
    kernel_shape=[2, 2],
    strides=[2, 2]
)
xT = np.array([[[[5, 6],
                 [7, 8]]]], dtype=np.float32)
xI = np.array([[[[5, 7],
                 [13, 15]]]], dtype=np.int64)
output_shape = np.array((1, 1, 5, 5), dtype=np.int64)
y = np.array([[[[0, 0, 0, 0, 0],
                [0, 5, 0, 6, 0],
                [0, 0, 0, 0, 0],
                [0, 7, 0, 8, 0],
                [0, 0, 0, 0, 0]]]], dtype=np.float32)
expect(node, inputs=[xT, xI, output_shape], outputs=[y], name='test_maxunpool_export_with_output_shape')

without_output_shape

node = onnx.helper.make_node(
    'MaxUnpool',
    inputs=['xT', 'xI'],
    outputs=['y'],
    kernel_shape=[2, 2],
    strides=[2, 2]
)
xT = np.array([[[[1, 2],
                 [3, 4]]]], dtype=np.float32)
xI = np.array([[[[5, 7],
                 [13, 15]]]], dtype=np.int64)
y = np.array([[[[0, 0, 0, 0],
                [0, 1, 0, 2],
                [0, 0, 0, 0],
                [0, 3, 0, 4]]]], dtype=np.float32)
expect(node, inputs=[xT, xI], outputs=[y], name='test_maxunpool_export_without_output_shape')

Mean

Element-wise mean of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Mean-1, Mean-6

Inputs (1 - ∞)

data_0 (variadic) : T

List of tensors for mean.

Outputs

mean : T

Output tensor.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

mean

data_0 = np.array([3, 0, 2]).astype(np.float32)
data_1 = np.array([1, 3, 4]).astype(np.float32)
data_2 = np.array([2, 6, 6]).astype(np.float32)
result = np.array([2, 3, 4]).astype(np.float32)
node = onnx.helper.make_node(
    'Mean',
    inputs=['data_0', 'data_1', 'data_2'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
       name='test_mean_example')

node = onnx.helper.make_node(
    'Mean',
    inputs=['data_0'],
    outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
       name='test_mean_one_input')

result = np.divide(np.add(data_0, data_1), 2.)
node = onnx.helper.make_node(
    'Mean',
    inputs=['data_0', 'data_1'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
       name='test_mean_two_inputs')

Min

Element-wise min of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Min-1, Min-6

Inputs (1 - ∞)

data_0 (variadic) : T

List of tensors for min.

Outputs

min : T

Output tensor.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

min

data_0 = np.array([3, 2, 1]).astype(np.float32)
data_1 = np.array([1, 4, 4]).astype(np.float32)
data_2 = np.array([2, 5, 0]).astype(np.float32)
result = np.array([1, 2, 0]).astype(np.float32)
node = onnx.helper.make_node(
    'Min',
    inputs=['data_0', 'data_1', 'data_2'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
       name='test_min_example')

node = onnx.helper.make_node(
    'Min',
    inputs=['data_0'],
    outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
       name='test_min_one_input')

result = np.minimum(data_0, data_1)
node = onnx.helper.make_node(
    'Min',
    inputs=['data_0', 'data_1'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
       name='test_min_two_inputs')

Mul

Performs element-wise binary multiplication (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Mul-1, Mul-6

Inputs

A : T

First operand.

B : T

Second operand.

Outputs

C : T

Result, has same element type as two inputs

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

mul

node = onnx.helper.make_node(
    'Mul',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array([4, 5, 6]).astype(np.float32)
z = x * y  # expected output [4., 10., 18.]
expect(node, inputs=[x, y], outputs=[z],
       name='test_mul_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = x * y
expect(node, inputs=[x, y], outputs=[z],
       name='test_mul')

mul_broadcast

node = onnx.helper.make_node(
    'Mul',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = x * y
expect(node, inputs=[x, y], outputs=[z],
       name='test_mul_bcast')

Multinomial

Generate a tensor of samples from a multinomial distribution according to the probabilities of each of the possible outcomes.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Attributes

dtype : int (default is 6)

(Optional) The data type for the elements of the output tensor, if not specified, we will use int32.

sample_size : int (default is 1)

Number of times to sample.

seed : float

(Optional) Seed to the random generator, if not specified we will auto generate one.

Inputs

input : T1

Input tensor with shape [batch_size, class_size], where class_size is the number of all possible outcomes. Each value along the axis zero represents the unnormalized log-probability of each corresponding outcome in a batch.

Outputs

output : T2

Output tensor with shape [batch_size, sample_size], where sample_size is the number of times to sample. Each value along the axis zero represents the outcome of the corresponding sample in a batch.

Type Constraints

T1 : tensor(float16), tensor(float), tensor(double)

Constrain input types to float tensors.

T2 : tensor(int32), tensor(int64)

Constrain output types to integral tensors.

Neg

Neg takes one input data (Tensor) and produces one output data (Tensor) where each element flipped sign, y = -x, is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Neg-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float), tensor(int32), tensor(int8), tensor(int16), tensor(int64), tensor(float16), tensor(double)

Constrain input and output types to signed numeric tensors.

Examples

neg

node = onnx.helper.make_node(
    'Neg',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-4, 2]).astype(np.float32)
y = np.negative(x)  # expected output [4., -2.],
expect(node, inputs=[x], outputs=[y],
       name='test_neg_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.negative(x)
expect(node, inputs=[x], outputs=[y],
       name='test_neg')

Not

Returns the negation of the input tensor element-wise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(bool)

Constrains input/output to boolean tensors.

Examples

not

node = onnx.helper.make_node(
    'Not',
    inputs=['x'],
    outputs=['not'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
expect(node, inputs=[x], outputs=[np.logical_not(x)],
       name='test_not_2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
expect(node, inputs=[x], outputs=[np.logical_not(x)],
       name='test_not_3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
expect(node, inputs=[x], outputs=[np.logical_not(x)],
       name='test_not_4d')

OneHot

Produces a one-hot tensor based on inputs. The locations represented by the index values in the 'indices' input tensor will have 'on_value' and the other locations will have 'off_value' in the output tensor, where 'on_value' and 'off_value' are specified as part of required input argument 'values', which is a two-element tensor of format [off_value, on_value]. The rank of the output tensor will be one greater than the rank of the input tensor. The additional dimension is for one-hot representation. The additional dimension will be inserted at the position specified by 'axis'. If 'axis' is not specified then then additional dimension will be inserted as the innermost dimension, i.e. axis=-1. The size of the additional dimension is specified by required scalar input 'depth'. The type of the output tensor is the same as the type of the 'values' input. Any entries in the 'indices' input tensor with values outside the range [0, depth) will result in one-hot representation with all 'off_value' values in the output tensor.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Attributes

axis : int (default is -1)

(Optional) Axis along which one-hot representation in added. Default: axis=-1. axis=-1 means that the additional dimension will be inserted as the innermost/last dimension in the output tensor.

Inputs

indices : T1

Input tensor containing indices. The values must be non-negative integers. Any entries in the 'indices' input tensor with values outside the range [0, depth) will result in one-hot representation with all 'off_value' values in the output tensor.In case 'indices' is of non-integer type, the values will be casted to int64 before use.

depth : T2

Scalar specifying the number of classes in one-hot tensor. This is also the size of the one-hot dimension (specified by 'axis' attribute) added on in the output tensor and the values in the 'indices' input tensor are expected to be in the range [0, depth). TheIn case 'depth' is of non-integer type, it will be casted to int64 before use.

values : T3

Rank 1 tensor containing exactly two elements, in the format [off_value, on_value], where 'on_value' is the value used for filling locations specified in 'indices' input tensor, and 'off_value' is the value used for filling locations other than those specified in 'indices' input tensor.

Outputs

output : T3

Tensor of rank one greater than input tensor 'indices', i.e. rank(output) = rank(indices) + 1. The data type for the elements of the output tensor is the same as the type of input 'values' is used.

Type Constraints

T1 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrains input to only numeric types.

T2 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrains input to only numeric types.

T3 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain to any tensor type.

Examples

with_axis

axisValue = 1
on_value = 3
off_value = 1
output_type = np.float32
node = onnx.helper.make_node(
    'OneHot',
    inputs=['indices', 'depth', 'values'],
    outputs=['y'],
    axis=axisValue
)
indices = np.array([[1, 9],
                    [2, 4]], dtype=np.float32)
depth = np.array([10], dtype=np.float32)
values = np.array([off_value, on_value], dtype=output_type)
y = one_hot(indices, depth, axis=axisValue, dtype=output_type)
y = y * (on_value - off_value) + off_value
expect(node, inputs=[indices, depth, values], outputs=[y], name='test_onehot_with_axis')

without_axis

on_value = 5
off_value = 2
output_type = np.int32
node = onnx.helper.make_node(
    'OneHot',
    inputs=['indices', 'depth', 'values'],
    outputs=['y']
)
indices = np.array([0, 7, 8], dtype=np.int64)
depth = np.array([12], dtype=np.float32)
values = np.array([off_value, on_value], dtype=output_type)
y = one_hot(indices, depth, dtype=output_type)
y = y * (on_value - off_value) + off_value
expect(node, inputs=[indices, depth, values], outputs=[y], name='test_onehot_without_axis')

Or

Returns the tensor resulted from performing the or logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Or-1

Inputs

A : T

First input operand for the logical operator.

B : T

Second input operand for the logical operator.

Outputs

C : T1

Result tensor.

Type Constraints

T : tensor(bool)

Constrains input to boolean tensor.

T1 : tensor(bool)

Constrains output to boolean tensor.

Examples

or

node = onnx.helper.make_node(
    'Or',
    inputs=['x', 'y'],
    outputs=['or'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
y = (np.random.randn(3, 4) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or4d')

or_broadcast

node = onnx.helper.make_node(
    'Or',
    inputs=['x', 'y'],
    outputs=['or'],
)

# 3d vs 1d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(5) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or_bcast3v1d')

# 3d vs 2d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(4, 5) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or_bcast3v2d')

# 4d vs 2d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or_bcast4v2d')

# 4d vs 3d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(4, 5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or_bcast4v3d')

# 4d vs 4d
x = (np.random.randn(1, 4, 1, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 1, 5, 6) > 0).astype(np.bool)
z = np.logical_or(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_or_bcast4v4d')

PRelu

PRelu takes input data (Tensor) and slope tensor as input, and produces one output data (Tensor) where the function f(x) = slope * x for x < 0, f(x) = x for x >= 0., is applied to the data tensor elementwise. This operator supports unidirectional broadcasting (tensor slope should be unidirectional broadcastable to input tensor X); for more details please check the doc.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: PRelu-1, PRelu-6, PRelu-7

Inputs

X : T

Input tensor

slope : T

Slope tensor. The shape of slope can be smaller then first input X; if so, its shape must be unidirectional broadcastable to X

Outputs

Y : T

Output tensor (same size as X)

Type Constraints

T : tensor(float16), tensor(float), tensor(double), tensor(uint32), tensor(uint64), tensor(int32), tensor(int64)

Constrain input and output types to float/int tensors.

Examples

prelu

node = onnx.helper.make_node(
    'PRelu',
    inputs=['x', 'slope'],
    outputs=['y'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
slope = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * slope

expect(node, inputs=[x, slope], outputs=[y],
       name='test_prelu_example')

prelu_broadcast

node = onnx.helper.make_node(
    'PRelu',
    inputs=['x', 'slope'],
    outputs=['y'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
slope = np.random.randn(5).astype(np.float32)
y = np.clip(x, 0, np.inf) + np.clip(x, -np.inf, 0) * slope

expect(node, inputs=[x, slope], outputs=[y],
       name='test_prelu_broadcast')

Pad

Given data tensor, pads, mode, and value. Example: Insert 0 pads to the beginning of the second dimension. data = [ [1.0, 1.2], [2.3, 3.4], [4.5, 5.7], ] pads = [0, 2, 0, 0] output = [ [ [0.0, 0.0, 1.0, 1.2], [0.0, 0.0, 2.3, 3.4], [0.0, 0.0, 4.5, 5.7], ], ]

Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: Pad-1

Attributes

mode : string (default is constant)

Three modes: constant(default), reflect, edge

pads : list of ints (required)

List of integers indicating the number of padding elements to add or remove (if negative) at the beginning and end of each axis. For 2D it is the number of pixels. `pads` rank should be double of the input's rank. `pads` format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis `i` and xi_end, the number of pixels added at the end of axis `i`.

value : float (default is 0.0)

One float, indicates the value to be filled.

Inputs

data : T

Input tensor.

Outputs

output : T

Tensor after padding.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

constant_pad

node = onnx.helper.make_node(
    'Pad',
    inputs=['x'],
    outputs=['y'],
    mode='constant',
    value=1.2,
    pads=[0, 0, 1, 3, 0, 0, 2, 4],
)
x = np.random.randn(1, 3, 4, 5).astype(np.float32)
y = np.pad(
    x,
    pad_width=((0, 0), (0, 0), (1, 2), (3, 4)),
    mode='constant',
    constant_values=1.2,
)

expect(node, inputs=[x], outputs=[y],
       name='test_constant_pad')

reflection_and_edge_pad

for mode in ['edge', 'reflect']:
    node = onnx.helper.make_node(
        'Pad',
        inputs=['x'],
        outputs=['y'],
        mode=mode,
        pads=[0, 0, 1, 1, 0, 0, 1, 1]
    )
    x = np.random.randn(1, 3, 4, 5).astype(np.float32)
    y = np.pad(
        x,
        pad_width=((0, 0), (0, 0), (1, 1), (1, 1)),
        mode=mode,
    )

    expect(node, inputs=[x], outputs=[y],
           name='test_{}_pad'.format(mode))

Pow

Pow takes input data (Tensor) and exponent Tensor, and produces one output data (Tensor) where the function f(x) = x^exponent, is applied to the data tensor elementwise. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Pow-1

Inputs

X : T

First operand, base of the exponent.

Y : T

Second operand, power of the exponent.

Outputs

Z : T

Output tensor (same size as X)

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

pow

node = onnx.helper.make_node(
    'Pow',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array([4, 5, 6]).astype(np.float32)
z = np.power(x, y)  # expected output [1., 32., 729.]
expect(node, inputs=[x, y], outputs=[z],
       name='test_pow_example')

x = np.arange(60).reshape(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = np.power(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_pow')

pow_broadcast

node = onnx.helper.make_node(
    'Pow',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array(2).astype(np.float32)
z = np.power(x, y)  # expected output [1., 4., 9.]
expect(node, inputs=[x, y], outputs=[z],
       name='test_pow_bcast_scalar')

node = onnx.helper.make_node(
    'Pow',
    inputs=['x', 'y'],
    outputs=['z'],
)
x = np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32)
y = np.array([1, 2, 3]).astype(np.float32)
# expected output [[1, 4, 27], [4, 25, 216]]
z = np.power(x, y).astype(np.float32)
expect(node, inputs=[x, y], outputs=[z],
       name='test_pow_bcast_array')

RNN

Computes an one-layer simple RNN. This operator is usually supported via some custom implementation such as CuDNN.

Notations:

X - input tensor

i - input gate

t - time step (t-1 means previous time step)

Wi - W parameter weight matrix for input gate

Ri - R recurrence weight matrix for input gate

Wbi - W parameter bias vector for input gate

Rbi - R parameter bias vector for input gate

WBi - W parameter weight matrix for backward input gate

RBi - R recurrence weight matrix for backward input gate

WBbi - WR bias vectors for backward input gate

RBbi - RR bias vectors for backward input gate

H - Hidden state

num_directions - 2 if direction == bidirectional else 1

Activation functions:

Relu(x)                - max(0, x)

Tanh(x)                - (1 - e^{-2x})/(1 + e^{-2x})

Sigmoid(x)             - 1/(1 + e^{-x})

(NOTE: Below are optional)

Affine(x)              - alpha*x + beta

LeakyRelu(x)           - x if x >= 0 else alpha * x

ThresholdedRelu(x)     - x if x >= alpha else 0

ScaledTanh(x)          - alpha*Tanh(beta*x)

HardSigmoid(x)         - min(max(alpha*x + beta, 0), 1)

Elu(x)                 - x if x >= 0 else alpha*(e^x - 1)

Softsign(x)            - x/(1 + |x|)

Softplus(x)            - log(1 + e^x)

Equations (Default: f=Tanh):

- Ht = f(Xt*(Wi^T) + Ht-1*(Ri^T) + Wbi + Rbi)

This operator has optional inputs/outputs. See the doc for more details about the representation of optional arguments. An empty string may be used in the place of an actual argument's name to indicate a missing argument. Trailing optional arguments (those not followed by an argument that is present) may also be simply omitted.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: RNN-1

Attributes

activation_alpha : list of floats

Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.For example with LeakyRelu, the default alpha is 0.01.

activation_beta : list of floats

Optional scaling values used by some activation functions. The values are consumed in the order of activation functions, for example (f, g, h) in LSTM. Default values are the same as of corresponding ONNX operators.

activations : list of strings (default is ['Tanh', 'Tanh'])

One (or two if bidirectional) activation function for input gate. The activation function must be one of the activation functions specified above. Optional: Default `Tanh` if not specified.

clip : float

Cell clip threshold. Clipping bounds the elements of a tensor in the range of [-threshold, +threshold] and is applied to the input of activations. No clip if not specified.

direction : string (default is forward)

Specify if the RNN is forward, reverse, or bidirectional. Must be one of forward (default), reverse, or bidirectional.

hidden_size : int

Number of neurons in the hidden layer

Inputs (3 - 6)

X : T

The input sequences packed (and potentially padded) into one 3-D tensor with the shape of `[seq_length, batch_size, input_size]`.

W : T

The weight tensor for input gate. Concatenation of `Wi` and `WBi` (if bidirectional). The tensor has shape `[num_directions, hidden_size, input_size]`.

R : T

The recurrence weight tensor. Concatenation of `Ri` and `RBi` (if bidirectional). The tensor has shape `[num_directions, hidden_size, hidden_size]`.

B (optional) : T

The bias tensor for input gate. Concatenation of `[Wbi, Rbi]` and `[WBbi, RBbi]` (if bidirectional). The tensor has shape `[num_directions, 2*hidden_size]`. Optional: If not specified - assumed to be 0.

sequence_lens (optional) : T1

Optional tensor specifying lengths of the sequences in a batch. If not specified - assumed all sequences in the batch to have length `seq_length`. It has shape `[batch_size]`.

initial_h (optional) : T

Optional initial value of the hidden. If not specified - assumed to be 0. It has shape `[num_directions, batch_size, hidden_size]`.

Outputs (0 - 2)

Y (optional) : T

A tensor that concats all the intermediate output values of the hidden. It has shape `[seq_length, num_directions, batch_size, hidden_size]`.

Y_h (optional) : T

The last output value of the hidden. It has shape `[num_directions, batch_size, hidden_size]`.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

T1 : tensor(int32)

Constrain seq_lens to integer tensor.

Examples

defaults

input = np.array([[[1., 2.], [3., 4.], [5., 6.]]]).astype(np.float32)

input_size = 2
hidden_size = 4
weight_scale = 0.1

node = onnx.helper.make_node(
    'RNN',
    inputs=['X', 'W', 'R'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = weight_scale * np.ones((1, hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, hidden_size, hidden_size)).astype(np.float32)

rnn = RNN_Helper(X=input, W=W, R=R)
_, Y_h = rnn.step()
expect(node, inputs=[input, W, R], outputs=[Y_h.astype(np.float32)], name='test_simple_rnn_defaults')

initial_bias

input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]]).astype(np.float32)

input_size = 3
hidden_size = 5
custom_bias = 0.1
weight_scale = 0.1

node = onnx.helper.make_node(
    'RNN',
    inputs=['X', 'W', 'R', 'B'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = weight_scale * np.ones((1, hidden_size, input_size)).astype(np.float32)
R = weight_scale * np.ones((1, hidden_size, hidden_size)).astype(np.float32)

# Adding custom bias
W_B = custom_bias * np.ones((1, hidden_size)).astype(np.float32)
R_B = np.zeros((1, hidden_size)).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

rnn = RNN_Helper(X=input, W=W, R=R, B=B)
_, Y_h = rnn.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)],
       name='test_simple_rnn_with_initial_bias')

seq_length

input = np.array([[[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]],
                  [[10., 11., 12.], [13., 14., 15.], [16., 17., 18.]]]).astype(np.float32)

input_size = 3
hidden_size = 5

node = onnx.helper.make_node(
    'RNN',
    inputs=['X', 'W', 'R', 'B'],
    outputs=['', 'Y'],
    hidden_size=hidden_size
)

W = np.random.randn(1, hidden_size, input_size).astype(np.float32)
R = np.random.randn(1, hidden_size, hidden_size).astype(np.float32)

# Adding custom bias
W_B = np.random.randn(1, hidden_size).astype(np.float32)
R_B = np.random.randn(1, hidden_size).astype(np.float32)
B = np.concatenate((W_B, R_B), axis=1)

rnn = RNN_Helper(X=input, W=W, R=R, B=B)
_, Y_h = rnn.step()
expect(node, inputs=[input, W, R, B], outputs=[Y_h.astype(np.float32)], name='test_rnn_seq_length')

RandomNormal

Generate a tensor with random values drawn from a normal distribution. The shape of the tensor is specified by the shape argument and the parameter of the normal distribution specified by mean and scale.

The data type is specified by the 'dtype' argument. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

dtype : int (default is 1)

The data type for the elements of the output tensor. Default is TensorProto::FLOAT.

mean : float (default is 0.0)

The mean of the normal distribution.

scale : float (default is 1.0)

The standard deviation of the normal distribution.

seed : float

(Optional) Seed to the random generator, if not specified we will auto generate one.

shape : list of ints (required)

The shape of the output tensor.

Inputs

Outputs

output : T

Output tensor of random values drawn from normal distribution

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain output types to float tensors.

RandomNormalLike

Generate a tensor with random values drawn from a normal distribution. The shape of the output tensor is copied from the shape of the input tensor, and the parameters of the normal distribution are specified by mean and scale.

The data type is specified by the 'dtype' argument, or copied from the input tensor if not provided. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message, and be valid as an output type.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

dtype : int

(Optional) The data type for the elements of the output tensor, if not specified, we will usethe data type of the input tensor.

mean : float (default is 0.0)

The mean of the normal distribution.

scale : float (default is 1.0)

The standard deviation of the normal distribution.

seed : float

(Optional) Seed to the random generator, if not specified we will auto generate one.

Inputs

input : T1

Input tensor to copy shape and optionally type information from.

Outputs

output : T2

Output tensor of random values drawn from normal distribution

Type Constraints

T1 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain to any tensor type. If the dtype attribute is not provided this must be a valid output type.

T2 : tensor(float16), tensor(float), tensor(double)

Constrain output types to float tensors.

RandomUniform

Generate a tensor with random values drawn from a uniform distribution. The shape of the tensor is specified by the shape argument and the range by low and high.

The data type is specified by the 'dtype' argument. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

dtype : int (default is 1)

The data type for the elements of the output tensor. If not specified, default is TensorProto::FLOAT.

high : float (default is 1.0)

Upper boundary of the output values.

low : float (default is 0.0)

Lower boundary of the output values.

seed : float

(Optional) Seed to the random generator, if not specified we will auto generate one.

shape : list of ints (required)

The shape of the output tensor.

Inputs

Outputs

output : T

Output tensor of random values drawn from uniform distribution

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain output types to float tensors.

RandomUniformLike

Generate a tensor with random values drawn from a uniform distribution. The shape of the output tensor is copied from the shape of the input tensor, and the parameters of the uniform distribution are specified by low and high.

The data type is specified by the 'dtype' argument, or copied from the input tensor if not provided. The 'dtype' argument must be one of the data types specified in the 'DataType' enum field in the TensorProto message and be valid as an output type.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

dtype : int

(Optional) The data type for the elements of the output tensor, if not specified, we will usethe data type of the input tensor.

high : float (default is 1.0)

Upper boundary of the output values.

low : float (default is 0.0)

Lower boundary of the output values.

seed : float

(Optional) Seed to the random generator, if not specified we will auto generate one.

Inputs

input : T1

Input tensor to copy shape and optionally type information from.

Outputs

output : T2

Output tensor of random values drawn from uniform distribution

Type Constraints

T1 : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain to any tensor type. If the dtype attribute is not provided this must be a valid output type.

T2 : tensor(float16), tensor(float), tensor(double)

Constrain output types to float tensors.

Reciprocal

Reciprocal takes one input data (Tensor) and produces one output data (Tensor) where the reciprocal is, y = 1/x, is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Reciprocal-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

reciprocal

node = onnx.helper.make_node(
    'Reciprocal',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-4, 2]).astype(np.float32)
y = np.reciprocal(x)  # expected output [-0.25, 0.5],
expect(node, inputs=[x], outputs=[y],
       name='test_reciprocal_example')

x = np.random.rand(3, 4, 5).astype(np.float32) + 0.5
y = np.reciprocal(x)
expect(node, inputs=[x], outputs=[y],
       name='test_reciprocal')

ReduceL1

Computes the L1 norm of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceL1',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sum(a=np.abs(data), axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[78.]]]

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l1_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(a=np.abs(data), axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l1_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [2]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceL1',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[3., 7.], [11., 15.], [19., 23.]]

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l1_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l1_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [2]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceL1',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[3.], [7.]], [[11.], [15.]], [[19.], [23.]]]

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l1_keep_dims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(a=np.abs(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l1_keep_dims_random')

ReduceL2

Computes the L2 norm of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceL2',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sqrt(np.sum(
    a=np.square(data), axis=axes, keepdims=keepdims == 1))
#print(reduced)
#[[[25.49509757]]]

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l2_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sqrt(np.sum(
    a=np.square(data), axis=axes, keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l2_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [2]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceL2',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sqrt(np.sum(
    a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))
#print(reduced)
#[[2.23606798, 5.],
# [7.81024968, 10.63014581],
# [13.45362405, 16.2788206]]

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l2_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sqrt(np.sum(
    a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l2_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [2]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceL2',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims
)

data = np.reshape(np.arange(1, np.prod(shape) + 1, dtype=np.float32), shape)
#print(data)
#[[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]], [[9., 10.], [11., 12.]]]

reduced = np.sqrt(np.sum(
    a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))
#print(reduced)
#[[[2.23606798], [5.]]
# [[7.81024968], [10.63014581]]
# [[13.45362405], [16.2788206 ]]]

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_l2_keep_dims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sqrt(np.sum(
    a=np.square(data), axis=tuple(axes), keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_l2_keep_dims_random')

ReduceLogSum

Computes the log sum of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

keepdims

node = onnx.helper.make_node(
    'ReduceLogSum',
    inputs=['data'],
    outputs=["reduced"]
)
data = np.random.ranf([3, 4, 5]).astype(np.float32)
reduced = np.log(np.sum(data, keepdims=True))
expect(node, inputs=[data], outputs=[reduced],
       name='test_reduce_log_sum_default')

nokeepdims

node = onnx.helper.make_node(
    'ReduceLogSum',
    inputs=['data'],
    outputs=["reduced"],
    axes=[2, 1],
    keepdims=0
)
data = np.random.ranf([3, 4, 5]).astype(np.float32)
reduced = np.log(np.sum(data, axis=(2, 1), keepdims=False))
expect(node, inputs=[data], outputs=[reduced],
       name='test_reduce_log_sum_desc_axes')

node = onnx.helper.make_node(
    'ReduceLogSum',
    inputs=['data'],
    outputs=["reduced"],
    axes=[0, 1],
    keepdims=0
)
data = np.random.ranf([3, 4, 5]).astype(np.float32)
reduced = np.log(np.sum(data, axis=(0, 1), keepdims=False))
expect(node, inputs=[data], outputs=[reduced],
       name='test_reduce_log_sum_asc_axes')

ReduceLogSumExp

Computes the log sum exponent of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceLogSumExp',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims
)

data = np.array(
    [[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]],
    dtype=np.float32)
reduced = np.log(np.sum(np.exp(data),
                        axis=axes,
                        keepdims=keepdims == 1))
# print(reduced)
# [[[60.00671387]]]

expect(node, inputs=[data], outputs=[reduced],
      name='test_reduce_log_sum_exp_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.log(np.sum(np.exp(data),
                        axis=axes,
                        keepdims=keepdims == 1))
expect(node, inputs=[data], outputs=[reduced],
      name='test_reduce_log_sum_exp_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0
node = onnx.helper.make_node(
    'ReduceLogSumExp',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims
)

data = np.array(
    [[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]],
    dtype=np.float32)
reduced = np.log(np.sum(
    np.exp(data), axis=tuple(axes), keepdims=keepdims == 1))
# print(reduced)
#[[20., 2.31326175]
# [40.00004578, 2.31326175]
# [60.00671387, 2.31326175]]

expect(node, inputs=[data], outputs=[reduced],
      name='test_reduce_log_sum_exp_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.log(np.sum(
    np.exp(data), axis=tuple(axes), keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
    name='test_reduce_log_sum_exp_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1
node = onnx.helper.make_node(
    'ReduceLogSumExp',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims
)

data = np.array(
    [[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]],
    dtype=np.float32)
reduced = np.log(np.sum(np.exp(data),
                        axis=tuple(axes),
                        keepdims=keepdims == 1))
# print(reduced)
# [[[20., 2.31326175]]
# [[40.00004578, 2.31326175]]
# [[60.00671387, 2.31326175]]]

expect(node, inputs=[data], outputs=[reduced],
      name='test_reduce_log_sum_exp_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.log(np.sum(np.exp(data),
                        axis=tuple(axes),
                        keepdims=keepdims == 1))

expect(node, inputs=[data], outputs=[reduced],
      name='test_reduce_log_sum_exp_keepdims_random')

ReduceMax

Computes the max of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1
node = onnx.helper.make_node(
    'ReduceMax',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.maximum.reduce(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
[[[60.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_default_axes_keepdim_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.maximum.reduce(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceMax',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[20., 2.]
# [40., 2.]
# [60., 2.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceMax',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[20., 2.]]
# [[40., 2.]]
# [[60., 2.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.maximum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_max_keepdims_random')

ReduceMean

Computes the mean of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceMean',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.mean(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[18.25]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.mean(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceMean',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[12.5, 1.5]
# [35., 1.5]
# [57.5, 1.5]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceMean',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[12.5, 1.5]]
# [[35., 1.5]]
# [[57.5, 1.5]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.mean(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_mean_keepdims_random')

ReduceMin

Computes the min of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceMin',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.minimum.reduce(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[1.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.minimum.reduce(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceMin',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[5., 1.]
# [30., 1.]
# [55., 1.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceMin', inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[5, 1], [20, 2]], [[30, 1], [40, 2]], [[55, 1], [60, 2]]], dtype=np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[5., 1.]]
# [[30., 1.]]
# [[55., 1.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.minimum.reduce(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_min_keepdims_random')

ReduceProd

Computes the product of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceProd',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.prod(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[4.790016e+08]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.prod(data, axis=axes, keepdims=keepdims == 1)
expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceProd',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[3., 8.]
# [35., 48.]
# [99., 120.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceProd',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[3., 8.]]
# [[35., 48.]]
# [[99., 120.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.prod(data, axis=tuple(axes), keepdims=keepdims == 1)
expect(node, inputs=[data], outputs=[reduced], name='test_reduce_prod_keepdims_random')

ReduceSum

Computes the sum of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceSum',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(data, axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[78.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(data, axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceSum',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[4., 6.]
# [12., 14.]
# [20., 22.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceSum',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[4., 6.]]
# [[12., 14.]]
# [[20., 22.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(data, axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_keepdims_random')

ReduceSumSquare

Computes the sum square of the input tensor's element along the provided axes. The resulted tensor has the same rank as the input if keepdims equal 1. If keepdims equal 0, then the resulted tensor have the reduced dimension pruned.

The above behavior is similar to numpy, with the exception that numpy default keepdims to False instead of True.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

A list of integers, along which to reduce. The default is to reduce over all the dimensions of the input tensor.

keepdims : int (default is 1)

Keep the reduced dimension or not, default 1 mean keep reduced dimension.

Inputs

data : T

An input tensor.

Outputs

reduced : T

Reduced output tensor.

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

default_axes_keepdims

shape = [3, 2, 2]
axes = None
keepdims = 1

node = onnx.helper.make_node(
    'ReduceSumSquare',
    inputs=['data'],
    outputs=['reduced'],
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(np.square(data), axis=axes, keepdims=keepdims == 1)
#print(reduced)
#[[[650.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_default_axes_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(np.square(data), axis=axes, keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_default_axes_keepdims_random')

do_not_keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 0

node = onnx.helper.make_node(
    'ReduceSumSquare',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[10., 20.]
# [74., 100.]
# [202., 244.]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_do_not_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_do_not_keepdims_random')

keepdims

shape = [3, 2, 2]
axes = [1]
keepdims = 1

node = onnx.helper.make_node(
    'ReduceSumSquare',
    inputs=['data'],
    outputs=['reduced'],
    axes=axes,
    keepdims=keepdims)

data = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]], [[9, 10], [11, 12]]], dtype=np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)
#print(reduced)
#[[[10., 20.]]
# [[74., 100.]]
# [[202., 244.]]]

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_keepdims_example')

np.random.seed(0)
data = np.random.uniform(-10, 10, shape).astype(np.float32)
reduced = np.sum(np.square(data), axis=tuple(axes), keepdims=keepdims == 1)

expect(node, inputs=[data], outputs=[reduced], name='test_reduce_sum_square_keepdims_random')

Relu

Relu takes one input data (Tensor) and produces one output data (Tensor) where the rectified linear function, y = max(0, x), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Relu-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

relu

node = onnx.helper.make_node(
    'Relu',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf)

expect(node, inputs=[x], outputs=[y],
       name='test_relu')

Reshape

Reshape the input tensor similar to numpy.reshape. First input is the data tensor, second input is a shape tensor which specifies the output shape. It outputs the reshaped tensor. At most one dimension of the new shape can be -1. In this case, the value is inferred from the size of the tensor and the remaining dimensions. A dimension could also be 0, in which case the actual dimension value is unchanged (i.e. taken from the input tensor).

Version

This version of the operator has been available since version 5 of the default ONNX operator set.

Other versions of this operator: Reshape-1

Inputs

data : T

An input tensor.

shape : tensor(int64)

Specified shape for output.

Outputs

reshaped : T

Reshaped data.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

reshape

original_shape = [2, 3, 4]
test_cases = {
    'reordered_dims': np.array([4, 2, 3], dtype=np.int64),
    'reduced_dims': np.array([3, 8], dtype=np.int64),
    'extended_dims': np.array([3, 2, 2, 2], dtype=np.int64),
    'one_dim': np.array([24], dtype=np.int64),
    'negative_dim': np.array([6, -1, 2], dtype=np.int64),
}
data = np.random.random_sample(original_shape).astype(np.float32)

for test_name, shape in test_cases.items():
    node = onnx.helper.make_node(
        'Reshape',
        inputs=['data', 'shape'],
        outputs=['reshaped'],
    )

    reshaped = np.reshape(data, shape)
    expect(node, inputs=[data, shape], outputs=[reshaped],
           name='test_reshape_' + test_name)

Scan

Scan can be used to iterate over one or more scan_input tensors, constructing zero or more scan_output tensors. It combines ideas from general recurrences, functional programming constructs such as scan, fold, map, and zip and is intended to enable generalizations of RNN-like constructs for sequence-to-sequence processing. Other tensors (referred to as state_variables here) can be used to carry a state when iterating from one element to another (similar to hidden-state in RNNs, also referred to as loop-carried dependences in the context of loops). Many common usages involve a single scan_input tensor (where functionality similar to scan, fold and map can be obtained). When more than one scan_input is used, a behavior similar to zip is obtained.

The attribute body must be a graph, specifying the computation to be performed in every iteration. It takes as input the current values of the state_variables and the current iterated element of the scan_inputs. It must return the (updated) values of the state_variables and zero or more scan_output_element tensors. The values of the scan_output_element tensors are concatenated over all the iterations to produce the scan_output values of the scan construct (similar to the concatenated intermediate hidden-state values of RNN-like constructs). All the output tensors (state_variables as well as scan_output_element tensors) are required to have the same shape in each iteration of the loop (a restriction imposed to enable efficient memory allocation).

The scan operation returns the final values of the state_variables as well as the scan_outputs.

The optional attribute scan_input_directions specifies the direction (forward or backward) for each scan input. If this attribute is omitted, all sequences are scanned in the forward direction. A bidirectional scan may be performed by specifying the same tensor input twice in the scan_inputs, once with a forward direction, and once with a backward direction.

The scan_output of the operation is produced by concatenating the scan_output_element values produced by the body in each iteration. The optional attribute scan_output_directions specifies the direction in which scan_output is constructed (by appending or prepending the scan_output_element to scan_output in each iteration) for each scan_output. If this attribute is omitted, the scan_output_element is appended to the scan_output in each iteration.

The optional attribute axes specifies the axis to be scanned for each scan_input. If omitted, every scan_input will be scanned in axis 0. For example, if axis 0 is the batch axis and axis 1 is the time axis (to be scanned), specify an axis value of 1. Note that scanning a non-zero axis may be less efficient than scanning axis zero.

Note that because of the ONNX restriction that only the last parameter of an operator can be variadic, the initial-states and scan-inputs are listed together as one input parameter. Similarly, the final-states and scan-outputs are listed together as one output parameter. The attribute num_scan_inputs indicates the number M of scan-inputs.

The behavior of

  Scan <
      num_scan_inputs = m,
      body = loop-body,
      axes = [axis_1, ..., axis_m]
  > (init_1, ..., init_n, scan_1, ..., scan_m)

is equivalent to the following pseudo-code:

  // scan_i.shape[axis_i] denotes the (max) sequence-length of scan_i
  // scan_i.shape[axis_i] is required to be equal to scan_j.shape[axis_j] for all i,j.
  sequence_length = scan_1.shape[axis_1];

  // initialize state-variables
  st_1 = init_1; ... st_n = init_n;
  // initialize scan-output variables: [] denotes an empty tensor
  scan_out_1 = []; ...; scan_out_k = [];
  // identify number of iterations:

  // execute loop
  for (int t = 0; t < sequence_length; ++t) {
      // generate the scan-input elements: the notation T<axis=k>[t] indicates the sub-tensor
      // of rank one less than T obtained by indexing T at position t along axis k.
      si_1 = scan_1<axis=axis_1>[t];
      ... ;
      si_m = scan_m<axis=axis_m>[t];
      // execute loop-body
      st_1, ..., st_n, so_1, ..., so_k = loop-body(st_1, ..., st_n, si_1, ..., si_m)
      // accumulate the scan-output elements
      scan_out_1 = Concat<axis=0>(scan_out_1, so_1); ... ; scan_out_k = Concat<axis=0>(scan_out_k, so_k);
  }

  return st_1, ..., st_n, scan_out_1, ..., scan_out_k;

Sample usage: Encoding RNN using a Scan

The following example shows how a simple RNN over an input tensor %X, with weight tensor %Wi, recurrence weight tensor %Ri, bias tensors %Wbi and %Rbi, and initial hidden-state %H_0 can be encoded as a ScanLoop. Note that the loop-body is a nested graph, and it directly computes %Wi, %Ri, %Wbi, and %Rbi (typically constants or initializers in the body graph). If these values are computed in the outer graph, they need to be passed in as extra state_variables.

  graph rnn-encoding {
    %H_0 = ... 
    %X = ...
    %Y_h, %Y = Scan[body = <graph rnn-cell-1>, num_scan_inputs=1](%H_0, %X)
    return %Y, %Y_h
  }

  graph rnn-cell-1 (
    %H_tminus1[FLOAT, tensor]
    %X_t[FLOAT, tensor]
  ) {
    %Wi = ...
    %Ri = ...
    %Wbi = ...
    %Rbi = ...
    %t1 = X_t * (Wi^T)
    %t2 = H_tminus1*(Ri^T)
    %t3 = Add(%t1, %t2)
    %t4 = Add(%t3, %Wbi)
    %t5 = Add(%t4, %Rbi)
    %Ht = Tanh(%t5)
    %Accumulate = Identity(%Ht)
    return %Ht, %Accumulate
  }

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Scan-8

Attributes

axes : list of ints

An optional list of M flags. The i-th element of the list specifies the axis to be scanned (the sequence axis) for the i-th scan_input. If omitted, 0 will be used as the scan axis for every scan_input.

body : graph (required)

The graph run each iteration. It has N+M inputs: (loop state variables..., scan_input_elts...). It has N+K outputs: (loop state variables..., scan_output_elts...). Each scan_output is created by concatenating the value of the specified scan_output_elt value at the end of each iteration of the loop. It is an error if the dimensions of these values change across loop iterations.

num_scan_inputs : int (required)

An attribute specifying the number of scan_inputs M.

scan_input_directions : list of ints

An optional list of M flags. The i-th element of the list specifies the direction to be scanned for the i-th scan_input tensor: 0 indicates forward direction and 1 indicates reverse direction. If omitted, all scan_input tensors will be scanned in the forward direction.

scan_output_directions : list of ints

An optional list of K flags, one for each scan_output. The i-th element of the list specifies whether the i-th scan_output should be constructed by appending or prepending a new value in each iteration: 0 indicates appending and 1 indicates prepending. If omitted, all scan_output tensors will be produced by appending a value in each iteration.

Inputs (1 - ∞)

initial_state_and_scan_inputs (variadic, heterogeneous) : V

Initial values of the loop's N state variables followed by M scan_inputs

Outputs (1 - ∞)

final_state_and_scan_outputs (variadic, heterogeneous) : V

Final values of the loop's N state variables followed by K scan_outputs

Type Constraints

I : tensor(int64)

Int64 tensor

V : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

All Tensor types

Examples

scan

# Given an input sequence [x1, ..., xN], sum up its elements using a scan
# returning the final state (x1+x2+...+xN) as well the scan_output
# [x1, x1+x2, ..., x1+x2+...+xN]
#
# create graph to represent scan body
sum_in = onnx.helper.make_tensor_value_info('sum_in', onnx.TensorProto.FLOAT, [2])
next = onnx.helper.make_tensor_value_info('next', onnx.TensorProto.FLOAT, [2])
sum_out = onnx.helper.make_tensor_value_info('sum_out', onnx.TensorProto.FLOAT, [2])
scan_out = onnx.helper.make_tensor_value_info('scan_out', onnx.TensorProto.FLOAT, [2])
add_node = onnx.helper.make_node(
    'Add',
    inputs=['sum_in', 'next'],
    outputs=['sum_out']
)
id_node = onnx.helper.make_node(
    'Identity',
    inputs=['sum_out'],
    outputs=['scan_out']
)
scan_body = onnx.helper.make_graph(
    [add_node, id_node],
    'scan_body',
    [sum_in, next],
    [sum_out, scan_out]
)
# create scan op node
no_sequence_lens = ''   # optional input, not supplied
node = onnx.helper.make_node(
    'Scan',
    inputs=[no_sequence_lens, 'initial', 'x'],
    outputs=['y', 'z'],
    num_scan_inputs=1,
    body=scan_body
)
# create inputs for batch-size 1, sequence-length 3, inner dimension 2
initial = np.array([0, 0]).astype(np.float32).reshape((1, 2))
x = np.array([1, 2, 3, 4, 5, 6]).astype(np.float32).reshape((1, 3, 2))
# final state computed = [1 + 3 + 5, 2 + 4 + 6]
y = np.array([9, 12]).astype(np.float32).reshape((1, 2))
# scan-output computed
z = np.array([1, 2, 4, 6, 9, 12]).astype(np.float32).reshape((1, 3, 2))

expect(node, inputs=[initial, x], outputs=[y, z],
       name='test_scan_sum')

Scatter

Given data, updates and indices input tensors of rank r >= 1, write the values provided by updates into the first input, data, along axis dimension of data (by default outer-most one as axis=0) at corresponding indices. For each entry in updates, the target index in data is specified by corresponding entry in indices for dimension = axis, and index in source for dimension != axis. For instance, in a 2-D tensor case, data[indices[i][j]][j] = updates[i][j] if axis = 0, or data[i][indices[i][j]] = updates[i][j] if axis = 1, where i and j are loop counters from 0 up to the respective size in updates - 1.

Example 1: data = [ [0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0], ] indices = [ [1, 0, 2], [0, 2, 1], ] updates = [ [1.0, 1.1, 1.2], [2.0, 2.1, 2.2], ] output = [ [2.0, 1.1, 0.0] [1.0, 0.0, 2.2] [0.0, 2.1, 1.2] ]

Example 2: data = [[1.0, 2.0, 3.0, 4.0, 5.0]] indices = [[1, 3]] updates = [[1.1, 2.1]] axis = 1 output = [[1.0, 1.1, 3.0, 2.1, 5.0]]

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Attributes

axis : int (default is 0)

Which axis to scatter on. Negative value means counting dimensions from the back. Accepted range in [-r, r-1]

Inputs

data : T

Tensor of rank r >= 1.

indices : Tind

Tensor of int32/int64 indices, of r >= 1 (same rank as input).

updates : T

Tensor of rank r >=1 (same rank and shape as indices)

Outputs

output : T

Tensor of rank r >= 1 (same rank as input).

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Input and output types can be of any tensor type.

Tind : tensor(int32), tensor(int64)

Constrain indices to integer types

Examples

scatter_with_axis

node = onnx.helper.make_node(
    'Scatter',
    inputs=['data', 'indices', 'updates'],
    outputs=['y'],
    axis=1,
)
data = np.array([[1.0, 2.0, 3.0, 4.0, 5.0]], dtype=np.float32)
indices = np.array([[1, 3]], dtype=np.int64)
updates = np.array([[1.1, 2.1]], dtype=np.float32)

y = np.array([[1.0, 1.1, 3.0, 2.1, 5.0]], dtype=np.float32)

expect(node, inputs=[data, indices, updates], outputs=[y],
       name='test_scatter_with_axis')

scatter_without_axis

node = onnx.helper.make_node(
    'Scatter',
    inputs=['data', 'indices', 'updates'],
    outputs=['y'],
)
data = np.zeros((3, 3), dtype=np.float32)
indices = np.array([[1, 0, 2], [0, 2, 1]], dtype=np.int64)
updates = np.array([[1.0, 1.1, 1.2], [2.0, 2.1, 2.2]], dtype=np.float32)

y = np.array([
    [2.0, 1.1, 0.0],
    [1.0, 0.0, 2.2],
    [0.0, 2.1, 1.2]
], dtype=np.float32)

expect(node, inputs=[data, indices, updates], outputs=[y],
       name='test_scatter_without_axis')

Selu

Selu takes one input data (Tensor) and produces one output data (Tensor) where the scaled exponential linear unit function, y = gamma * (alpha * e^x - alpha) for x <= 0, y = gamma * x for x > 0, is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Selu-1

Attributes

alpha : float (default is 1.67326)

Coefficient of SELU default to 1.67326319217681884765625 (i.e., float32 approximation of 1.6732632423543772848170429916717).

gamma : float (default is 1.0507)

Coefficient of SELU default to 1.05070102214813232421875 (i.e., float32 approximation of 1.0507009873554804934193349852946).

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

selu

node = onnx.helper.make_node(
    'Selu',
    inputs=['x'],
    outputs=['y'],
    alpha=2.0,
    gamma=3.0
)

x = np.array([-1, 0, 1]).astype(np.float32)
# expected output [-3.79272318, 0., 3.]
y = np.clip(x, 0, np.inf) * 3.0 + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0 * 3.0
expect(node, inputs=[x], outputs=[y],
       name='test_selu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) * 3.0 + (np.exp(np.clip(x, -np.inf, 0)) - 1) * 2.0 * 3.0
expect(node, inputs=[x], outputs=[y],
       name='test_selu')

selu_default

default_alpha = 1.67326319217681884765625
default_gamma = 1.05070102214813232421875
node = onnx.helper.make_node(
    'Selu',
    inputs=['x'],
    outputs=['y'],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, 0, np.inf) * default_gamma + \
    (np.exp(np.clip(x, -np.inf, 0)) - 1) * default_alpha * default_gamma
expect(node, inputs=[x], outputs=[y],
       name='test_selu_default')

Shape

Takes a tensor as input and outputs an 1D int64 tensor containing the shape of the input tensor.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

data : T

An input tensor.

Outputs

shape : T1

Shape of the input tensor

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Input tensor can be of arbitrary type.

T1 : tensor(int64)

Constrain output to int64 tensor.

Examples

shape

node = onnx.helper.make_node(
    'Shape',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([
    [1, 2, 3],
    [4, 5, 6],
]).astype(np.float32)
y = np.array([
    2, 3,
]).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
       name='test_shape_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.array(x.shape).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
       name='test_shape')

Sigmoid

Sigmoid takes one input data (Tensor) and produces one output data (Tensor) where the sigmoid function, y = 1 / (1 + exp(-x)), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Sigmoid-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

sigmoid

node = onnx.helper.make_node(
    'Sigmoid',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = 1.0 / (1.0 + np.exp(np.negative(x)))  # expected output [0.26894143, 0.5, 0.7310586]
expect(node, inputs=[x], outputs=[y],
       name='test_sigmoid_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = 1.0 / (1.0 + np.exp(np.negative(x)))
expect(node, inputs=[x], outputs=[y],
       name='test_sigmoid')

Sign

Calculate the sign of the given input tensor element-wise. If input > 0, output 1. if input < 0, output -1. if input == 0, output 0.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The sign of the input tensor computed element-wise. It has the same shape and type of the input.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to all numeric tensors.

Examples

sign

node = onnx.helper.make_node(
    'Sign',
    inputs=['x'],
    outputs=['y'],
)

x = np.array(range(-5, 6)).astype(np.float32)
y = np.sign(x)
expect(node, inputs=[x], outputs=[y],
       name='test_sign')

Sin

Calculates the sine of the given input tensor, element-wise.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The sine of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

sin

node = onnx.helper.make_node(
    'Sin',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.sin(x)
expect(node, inputs=[x], outputs=[y],
       name='test_sin_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.sin(x)
expect(node, inputs=[x], outputs=[y],
       name='test_sin')

Sinh

Calculates the hyperbolic sine of the given input tensor element-wise.

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The hyperbolic sine values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

sinh

node = onnx.helper.make_node(
    'Sinh',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.sinh(x)  # expected output [-1.17520118,  0.,  1.17520118]
expect(node, inputs=[x], outputs=[y],
       name='test_sinh_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.sinh(x)
expect(node, inputs=[x], outputs=[y],
       name='test_sinh')

Size

Takes a tensor as input and outputs a int64 scalar that equals to the total number of elements of the input tensor.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

data : T

An input tensor.

Outputs

size : T1

Total number of elements of the input tensor

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Input tensor can be of arbitrary type.

T1 : tensor(int64)

Constrain output to int64 tensor, which should be a scalar though.

Examples

size

node = onnx.helper.make_node(
    'Size',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([
    [1, 2, 3],
    [4, 5, 6],
]).astype(np.float32)
y = np.array(6).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
       name='test_size_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.array(x.size).astype(np.int64)

expect(node, inputs=[x], outputs=[y],
       name='test_size')

Slice

Produces a slice of the input tensor along multiple axes. Similar to numpy: https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html Slices uses axes, starts and ends attributes to specify the start and end dimension for each axis in the list of axes, it uses this information to slice the input data tensor. If a negative value is passed for any of the start or end indices, it represent number of elements before the end of that dimension. If the value passed to start or end is larger than the n (the number of elements in this dimension), it represents n. For slicing to the end of a dimension with unknown size, it is recommended to pass in INT_MAX. If axes are omitted, they are set to [0, ..., ndim-1]. Example 1: data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] axes = [0, 1] starts = [1, 0] ends = [2, 3] result = [ [5, 6, 7], ] Example 2: data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] starts = [0, 1] ends = [-1, 1000] result = [ [2, 3, 4], ]

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

Axes that `starts` and `ends` apply to. It's optional. If not present, will be treated as [0, 1, ..., len(`starts`) - 1].

ends : list of ints (required)

Ending indices (exclusive) of corresponding axis in axes`

starts : list of ints (required)

Starting indices of corresponding axis in `axes`

Inputs

data : T

Tensor of data to extract slices from.

Outputs

output : T

Sliced data tensor.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

slice

node = onnx.helper.make_node(
    'Slice',
    inputs=['x'],
    outputs=['y'],
    axes=[0, 1],
    starts=[0, 0],
    ends=[3, 10],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[0:3, 0:10]

expect(node, inputs=[x], outputs=[y],
       name='test_slice')

slice_default_axes

node = onnx.helper.make_node(
    'Slice',
    inputs=['x'],
    outputs=['y'],
    starts=[0, 0, 3],
    ends=[20, 10, 4],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, :, 3:4]

expect(node, inputs=[x], outputs=[y],
       name='test_slice_default_axes')

slice_end_out_of_bounds

node = onnx.helper.make_node(
    'Slice',
    inputs=['x'],
    outputs=['y'],
    axes=[1],
    starts=[1],
    ends=[1000],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, 1:1000]

expect(node, inputs=[x], outputs=[y],
       name='test_slice_end_out_of_bounds')

slice_neg

node = onnx.helper.make_node(
    'Slice',
    inputs=['x'],
    outputs=['y'],
    axes=[1],
    starts=[0],
    ends=[-1],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, 0:-1]

expect(node, inputs=[x], outputs=[y],
       name='test_slice_neg')

slice_start_out_of_bounds

node = onnx.helper.make_node(
    'Slice',
    inputs=['x'],
    outputs=['y'],
    axes=[1],
    starts=[1000],
    ends=[1000],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[:, 1000:1000]

expect(node, inputs=[x], outputs=[y],
       name='test_slice_start_out_of_bounds')

Softmax

The operator computes the softmax (normalized exponential) values for each layer in the batch of the given input. The input is a 2-D tensor (Tensor) of size (batch_size x input_feature_dimensions). The output tensor has the same shape and contains the softmax values of the corresponding input.

Input does not need to explicitly be a 2D vector; rather, it will be coerced into one. For an arbitrary n-dimensional tensor input \in [a_0, a_1, ..., a_{k-1}, a_k, ..., a_{n-1}] and k is the axis provided, then input will be coerced into a 2-dimensional tensor with dimensions [a_0 * ... * a_{k-1}, a_k * ... * a_{n-1}]. For the default case where axis=1, this means the input tensor will be coerced into a 2D tensor of dimensions [a_0, a_1 * ... * a_{n-1}], where a_0 is often the batch size. In this situation, we must have a_0 = N and a_1 * ... * a_{n-1} = D. Each of these dimensions must be matched correctly, or else the operator will throw errors.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is 1)

Describes the axis of the inputs when coerced to 2D; defaults to one because the 0th axis most likely describes the batch_size

Inputs

input : T

The input tensor that's coerced into a 2D matrix of size (NxD) as described above.

Outputs

output : T

The output values with the same shape as input tensor (the original size without coercion).

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

softmax

node = onnx.helper.make_node(
    'Softmax',
    inputs=['x'],
    outputs=['y'],
)
x = np.array([[-1, 0, 1]]).astype(np.float32)
# expected output [[0.09003058, 0.24472848, 0.66524094]]
y = np.exp(x) / np.sum(np.exp(x), axis=1)
expect(node, inputs=[x], outputs=[y],
       name='test_softmax_example')

softmax_axis

def softmax_2d(x):  # type: (np.ndarray) -> np.ndarray
    max_x = np.max(x, axis=1).reshape((-1, 1))
    exp_x = np.exp(x - max_x)
    return exp_x / np.sum(exp_x, axis=1).reshape((-1, 1))

x = np.array([[0, 1, 2, 3], [10000, 10001, 10002, 10003]]).astype(np.float32)
# expected output [[0.0320586, 0.08714432, 0.23688284, 0.64391428],
#                 [0.0320586, 0.08714432, 0.23688284, 0.64391428]]
y = softmax_2d(x)

node = onnx.helper.make_node(
    'Softmax',
    inputs=['x'],
    outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
       name='test_softmax_large_number')

x = np.abs(np.random.randn(3, 4, 5).astype(np.float32))
node = onnx.helper.make_node(
    'Softmax',
    inputs=['x'],
    outputs=['y'],
    axis=0,
)
y = softmax_2d(x.reshape(1, 60)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_softmax_axis_0')

node = onnx.helper.make_node(
    'Softmax',
    inputs=['x'],
    outputs=['y'],
    axis=1,
)
y = softmax_2d(x.reshape(3, 20)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_softmax_axis_1')

# default axis is 1
node = onnx.helper.make_node(
    'Softmax',
    inputs=['x'],
    outputs=['y'],
)
expect(node, inputs=[x], outputs=[y],
       name='test_softmax_default_axis')

node = onnx.helper.make_node(
    'Softmax',
    inputs=['x'],
    outputs=['y'],
    axis=2,
)
y = softmax_2d(x.reshape(12, 5)).reshape(3, 4, 5)
expect(node, inputs=[x], outputs=[y],
       name='test_softmax_axis_2')

Softplus

Softplus takes one input data (Tensor) and produces one output data (Tensor) where the softplus function, y = ln(exp(x) + 1), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

X : T

1D input tensor

Outputs

Y : T

1D input tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

softplus

node = onnx.helper.make_node(
    'Softplus',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.log(np.exp(x) + 1)  # expected output [0.31326166, 0.69314718, 1.31326163]
expect(node, inputs=[x], outputs=[y],
       name='test_softplus_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.log(np.exp(x) + 1)
expect(node, inputs=[x], outputs=[y],
       name='test_softplus')

Softsign

Calculates the softsign (x/(1+|x|)) of the given input tensor element-wise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The softsign (x/(1+|x|)) values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

softsign

node = onnx.helper.make_node(
    'Softsign',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.array([-0.5, 0, 0.5]).astype(np.float32)
expect(node, inputs=[x], outputs=[y],
       name='test_softsign_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = x / (1 + np.abs(x))
expect(node, inputs=[x], outputs=[y],
       name='test_softsign')

SpaceToDepth

SpaceToDepth rearranges blocks of spatial data into depth. More specifically, this op outputs a copy of the input tensor where values from the height and width dimensions are moved to the depth dimension.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

blocksize : int (required)

Blocks of [blocksize, blocksize] are moved.

Inputs

input : T

Input tensor of [N,C,H,W], where N is the batch axis, C is the channel or depth, H is the height and W is the width.

Outputs

output : T

Output tensor of [N, C * blocksize * blocksize, H/blocksize, W/blocksize].

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Split

Split a tensor into a list of tensors, along the specified 'axis'. Lengths of the parts can be specified using argument 'split'. Otherwise, the tensor is split to equal sized parts.

Version

This version of the operator has been available since version 2 of the default ONNX operator set.

Other versions of this operator: Split-1

Attributes

axis : int (default is 0)

Which axis to split on.

split : list of ints

length of each output

Inputs

input : T

The tensor to split

Outputs (1 - ∞)

outputs (variadic) : T

One or more outputs forming list of tensors after splitting

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

1d

input = np.array([1., 2., 3., 4., 5., 6.]).astype(np.float32)

node = onnx.helper.make_node(
    'Split',
    inputs=['input'],
    outputs=['output_1', 'output_2', 'output_3'],
    axis=0
)

expected_outputs = [np.array([1., 2.]).astype(np.float32), np.array([3., 4.]).astype(np.float32), np.array([5., 6.]).astype(np.float32)]
expect(node, inputs=[input], outputs=[y for y in expected_outputs], name='test_split_equal_parts_1d')

node = onnx.helper.make_node(
    'Split',
    inputs=['input'],
    outputs=['output_1', 'output_2'],
    axis=0,
    split=[2, 4]
)

expected_outputs = [np.array([1., 2.]).astype(np.float32), np.array([3., 4., 5., 6.]).astype(np.float32)]
expect(node, inputs=[input], outputs=[y for y in expected_outputs], name='test_split_variable_parts_1d')

2d

input = np.array([[1., 2., 3., 4., 5., 6.],
                  [7., 8., 9., 10., 11., 12.]]).astype(np.float32)

node = onnx.helper.make_node(
    'Split',
    inputs=['input'],
    outputs=['output_1', 'output_2'],
    axis=1
)

expected_outputs = [np.array([[1., 2., 3.], [7., 8., 9.]]).astype(np.float32),
                    np.array([[4., 5., 6.], [10., 11., 12.]]).astype(np.float32)]

expect(node, inputs=[input], outputs=[y for y in expected_outputs], name='test_split_equal_parts_2d')

node = onnx.helper.make_node(
    'Split',
    inputs=['input'],
    outputs=['output_1', 'output_2'],
    axis=1,
    split=[2, 4]
)

expected_outputs = [np.array([[1., 2.], [7., 8.]]).astype(np.float32),
                    np.array([[3., 4., 5., 6.], [9., 10., 11., 12.]]).astype(np.float32)]

expect(node, inputs=[input], outputs=[y for y in expected_outputs], name='test_split_variable_parts_2d')

default_values

input = np.array([1., 2., 3., 4., 5., 6.]).astype(np.float32)

# If axis is not specified, split is applied on default axis 0
node = onnx.helper.make_node(
    'Split',
    inputs=['input'],
    outputs=['output_1', 'output_2', 'output_3']
)

expected_outputs = [np.array([1., 2.]).astype(np.float32), np.array([3., 4.]).astype(np.float32), np.array([5., 6.]).astype(np.float32)]
expect(node, inputs=[input], outputs=[y for y in expected_outputs], name='test_split_equal_parts_default_axis')

node = onnx.helper.make_node(
    'Split',
    inputs=['input'],
    outputs=['output_1', 'output_2'],
    split=[2, 4]
)

expected_outputs = [np.array([1., 2.]).astype(np.float32), np.array([3., 4., 5., 6.]).astype(np.float32)]
expect(node, inputs=[input], outputs=[y for y in expected_outputs], name='test_split_variable_parts_default_axis')

Sqrt

Square root takes one input data (Tensor) and produces one output data (Tensor) where the square root is, y = x^0.5, is applied to the tensor elementwise. If x is negative, then it will return NaN.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Sqrt-1

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

sqrt

node = onnx.helper.make_node(
    'Sqrt',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([1, 4, 9]).astype(np.float32)
y = np.sqrt(x)  # expected output [1., 2., 3.]
expect(node, inputs=[x], outputs=[y],
       name='test_sqrt_example')

x = np.abs(np.random.randn(3, 4, 5).astype(np.float32))
y = np.sqrt(x)
expect(node, inputs=[x], outputs=[y],
       name='test_sqrt')

Squeeze

Remove single-dimensional entries from the shape of a tensor. Takes a parameter axes with a list of axes to squeeze. If axes is not provided, all the single dimensions will be removed from the shape. If an axis is selected with shape entry not equal to one, an error is raised.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints

List of positive integers, indicate the dimensions to squeeze.

Inputs

data : T

Tensors with at least max(dims) dimensions.

Outputs

squeezed : T

Reshaped tensor with same data as input.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

squeeze

node = onnx.helper.make_node(
    'Squeeze',
    inputs=['x'],
    outputs=['y'],
    axes=[0],
)
x = np.random.randn(1, 3, 4, 5).astype(np.float32)
y = np.squeeze(x, axis=0)

expect(node, inputs=[x], outputs=[y],
       name='test_squeeze')

Sub

Performs element-wise binary subtraction (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Sub-1, Sub-6

Inputs

A : T

First operand.

B : T

Second operand.

Outputs

C : T

Result, has same element type as two inputs

Type Constraints

T : tensor(uint32), tensor(uint64), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain input and output types to high-precision numeric tensors.

Examples

sub

node = onnx.helper.make_node(
    'Sub',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.array([1, 2, 3]).astype(np.float32)
y = np.array([3, 2, 1]).astype(np.float32)
z = x - y  # expected output [-2., 0., 2.]
expect(node, inputs=[x, y], outputs=[z],
       name='test_sub_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(3, 4, 5).astype(np.float32)
z = x - y
expect(node, inputs=[x, y], outputs=[z],
       name='test_sub')

sub_broadcast

node = onnx.helper.make_node(
    'Sub',
    inputs=['x', 'y'],
    outputs=['z'],
)

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.random.randn(5).astype(np.float32)
z = x - y
expect(node, inputs=[x, y], outputs=[z],
       name='test_sub_bcast')

Sum

Element-wise sum of each of the input tensors (with Numpy-style broadcasting support). All inputs and outputs must have the same data type. This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 8 of the default ONNX operator set.

Other versions of this operator: Sum-1, Sum-6

Inputs (1 - ∞)

data_0 (variadic) : T

List of tensors for sum.

Outputs

sum : T

Output tensor.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

sum

data_0 = np.array([3, 0, 2]).astype(np.float32)
data_1 = np.array([1, 3, 4]).astype(np.float32)
data_2 = np.array([2, 6, 6]).astype(np.float32)
result = np.array([6, 9, 12]).astype(np.float32)
node = onnx.helper.make_node(
    'Sum',
    inputs=['data_0', 'data_1', 'data_2'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1, data_2], outputs=[result],
       name='test_sum_example')

node = onnx.helper.make_node(
    'Sum',
    inputs=['data_0'],
    outputs=['result'],
)
expect(node, inputs=[data_0], outputs=[data_0],
       name='test_sum_one_input')

result = np.add(data_0, data_1)
node = onnx.helper.make_node(
    'Sum',
    inputs=['data_0', 'data_1'],
    outputs=['result'],
)
expect(node, inputs=[data_0, data_1], outputs=[result],
       name='test_sum_two_inputs')

Tan

Calculates the tangent of the given input tensor, element-wise.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Inputs

input : T

Input tensor

Outputs

output : T

The tangent of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

tan

node = onnx.helper.make_node(
    'Tan',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.tan(x)
expect(node, inputs=[x], outputs=[y],
       name='test_tan_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.tan(x)
expect(node, inputs=[x], outputs=[y],
       name='test_tan')

Tanh

Calculates the hyperbolic tangent of the given input tensor element-wise.

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Tanh-1

Inputs

input : T

Input tensor

Outputs

output : T

The hyperbolic tangent values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

tanh

node = onnx.helper.make_node(
    'Tanh',
    inputs=['x'],
    outputs=['y'],
)

x = np.array([-1, 0, 1]).astype(np.float32)
y = np.tanh(x)  # expected output [-0.76159418, 0., 0.76159418]
expect(node, inputs=[x], outputs=[y],
       name='test_tanh_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.tanh(x)
expect(node, inputs=[x], outputs=[y],
       name='test_tanh')

Tile

Constructs a tensor by tiling a given tensor. This is the same as function tile in Numpy, but no broadcast. For example A = [[1, 2], [3, 4]], B = [1, 2], tile(A, B) = [[1, 2, 1, 2], [3, 4, 3, 4]]

Version

This version of the operator has been available since version 6 of the default ONNX operator set.

Other versions of this operator: Tile-1

Inputs

input : T

Input tensor of any shape.

repeats : T1

1D int64 tensor of the same length as input's dimension number, includes numbers of repeated copies along input's dimensions.

Outputs

output : T

Output tensor of the same dimension and type as tensor input. output_dim[i] = input_dim[i] * repeats[i]

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

T1 : tensor(int64)

Constrain repeat's type to int64 tensors.

Examples

tile

node = onnx.helper.make_node(
    'Tile',
    inputs=['x', 'y'],
    outputs=['z']
)

x = np.random.rand(2, 3, 4, 5).astype(np.float32)

repeats = np.random.randint(low=1, high=10, size=(np.ndim(x),)).astype(np.int64)

z = np.tile(x, repeats)

expect(node,
       inputs=[x, repeats],
       outputs=[z],
       name='test_tile')

tile_precomputed

node = onnx.helper.make_node(
    'Tile',
    inputs=['x', 'y'],
    outputs=['z']
)

x = np.array([
    [0, 1],
    [2, 3]
], dtype=np.float32)

repeats = np.array([2, 2], dtype=np.int64)

z = np.array([
    [0, 1, 0, 1],
    [2, 3, 2, 3],
    [0, 1, 0, 1],
    [2, 3, 2, 3]
], dtype=np.float32)

expect(node,
       inputs=[x, repeats],
       outputs=[z],
       name='test_tile_precomputed')

TopK

Retrieve the top-K elements along a specified axis. Given an input tensor of shape [a_1, a_2, ..., a_n, r] and integer argument k, return two outputs: -Value tensor of shape [a_1, a_2, ..., a_{axis-1}, k, a_{axis+1}, ... a_n] which contains the values of the top k elements along the specified axis -Index tensor of shape [a_1, a_2, ..., a_{axis-1}, k, a_{axis+1}, ... a_n] which contains the indices of the top k elements (original indices from the input tensor).

Given two equivalent values, this operator uses the indices along the axis as a tiebreaker. That is, the element with the lower index will appear first.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axis : int (default is -1)

Dimension on which to do the sort.

k : int (required)

Number of top elements to retrieve

Inputs

X : T

Tensor of shape [a_1, a_2, ..., a_n, r]

Outputs

Values : T

Tensor of shape [a_1, a_2, ..., a_{axis-1}, k, a_{axis+1}, ... a_n] containing top K values from the input tensor

Indices : I

Tensor of shape [a_1, a_2, ..., a_{axis-1}, k, a_{axis+1}, ... a_n] containing the corresponding input tensor indices for the top K values.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

I : tensor(int64)

Constrain index tensor to int64

Examples

top_k

node = onnx.helper.make_node(
    'TopK',
    inputs=['x'],
    outputs=['values', 'indices'],
    k=3
)
X = np.array([
    [0, 1, 2, 3],
    [4, 5, 6, 7],
    [8, 9, 10, 11],
], dtype=np.float32)
values_ref = np.array([
    [3, 2, 1],
    [7, 6, 5],
    [11, 10, 9],
], dtype=np.float32)
indices_ref = np.array([
    [3, 2, 1],
    [3, 2, 1],
    [3, 2, 1],
], dtype=np.int64)

expect(node, inputs=[X], outputs=[values_ref, indices_ref],
       name='test_top_k')

Transpose

Transpose the input tensor similar to numpy.transpose. For example, when perm=(1, 0, 2), given an input tensor of shape (1, 2, 3), the output shape will be (2, 1, 3).

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

perm : list of ints

A list of integers. By default, reverse the dimensions, otherwise permute the axes according to the values given.

Inputs

data : T

An input tensor.

Outputs

transposed : T

Transposed output.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

all_permutations

shape = (2, 3, 4)
data = np.random.random_sample(shape).astype(np.float32)
permutations = list(itertools.permutations(np.arange(len(shape))))

for i in range(len(permutations)):
    node = onnx.helper.make_node(
        'Transpose',
        inputs=['data'],
        outputs=['transposed'],
        perm=permutations[i]
    )
    transposed = np.transpose(data, permutations[i])
    expect(node, inputs=[data], outputs=[transposed],
           name='test_transpose_all_permutations_' + str(i))

default

shape = (2, 3, 4)
data = np.random.random_sample(shape).astype(np.float32)

node = onnx.helper.make_node(
    'Transpose',
    inputs=['data'],
    outputs=['transposed']
)

transposed = np.transpose(data)
expect(node, inputs=[data], outputs=[transposed],
       name='test_transpose_default')

Unsqueeze

Insert single-dimensional entries to the shape of a tensor. Takes one required argument axes, a list of dimensions that will be inserted. Dimension indices in axes are as seen in the output tensor. For example: Given a tensor such that tensor with shape [3, 4, 5], then Unsqueeze(tensor, axes=[0, 4]) has shape [1, 3, 4, 5, 1]

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

axes : list of ints (required)

List of positive integers, indicate the dimensions to be inserted

Inputs

data : T

Original tensor

Outputs

expanded : T

Reshaped tensor with same data as input.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Examples

unsqueeze

node = onnx.helper.make_node(
    'Unsqueeze',
    inputs=['x'],
    outputs=['y'],
    axes=[0],
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.expand_dims(x, axis=0)

expect(node, inputs=[x], outputs=[y],
       name='test_unsqueeze')

Upsample

Upsample the input tensor. Each dimension value of the output tensor is: output_dimension = floor(input_dimension * scale).

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Other versions of this operator: Upsample-7

Attributes

mode : string (default is nearest)

Two interpolation modes: nearest (default), and linear (including bilinear, trilinear, etc)

Inputs

X : T

N-D tensor

scales : tensor(float)

The scale array along each dimension. It takes value greater than or equal to 1. The number of elements of 'scales' should be the same as the rank of input 'X'.

Outputs

Y : T

N-D tensor after resizing

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input 'X' and output 'Y' to all tensor types.

Examples

nearest

node = onnx.helper.make_node(
    'Upsample',
    inputs=['X', 'scales'],
    outputs=['Y'],
    mode='nearest',
)

data = np.array([[[
    [1, 2],
    [3, 4],
]]], dtype=np.float32)

scales = np.array([1.0, 1.0, 2.0, 3.0], dtype=np.float32)

output = np.array([[[
    [1, 1, 1, 2, 2, 2],
    [1, 1, 1, 2, 2, 2],
    [3, 3, 3, 4, 4, 4],
    [3, 3, 3, 4, 4, 4],
]]], dtype=np.float32)

expect(node, inputs=[data, scales], outputs=[output],
       name='test_upsample_nearest')

Xor

Returns the tensor resulted from performing the xor logical operation elementwise on the input tensors A and B (with Numpy-style broadcasting support).

This operator supports multidirectional (i.e., Numpy-style) broadcasting; for more details please check the doc.

Version

This version of the operator has been available since version 7 of the default ONNX operator set.

Other versions of this operator: Xor-1

Inputs

A : T

First input operand for the logical operator.

B : T

Second input operand for the logical operator.

Outputs

C : T1

Result tensor.

Type Constraints

T : tensor(bool)

Constrains input to boolean tensor.

T1 : tensor(bool)

Constrains output to boolean tensor.

Examples

xor

node = onnx.helper.make_node(
    'Xor',
    inputs=['x', 'y'],
    outputs=['xor'],
)

# 2d
x = (np.random.randn(3, 4) > 0).astype(np.bool)
y = (np.random.randn(3, 4) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor2d')

# 3d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor3d')

# 4d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor4d')

xor_broadcast

node = onnx.helper.make_node(
    'Xor',
    inputs=['x', 'y'],
    outputs=['xor'],
)

# 3d vs 1d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(5) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor_bcast3v1d')

# 3d vs 2d
x = (np.random.randn(3, 4, 5) > 0).astype(np.bool)
y = (np.random.randn(4, 5) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor_bcast3v2d')

# 4d vs 2d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(5, 6) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor_bcast4v2d')

# 4d vs 3d
x = (np.random.randn(3, 4, 5, 6) > 0).astype(np.bool)
y = (np.random.randn(4, 5, 6) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor_bcast4v3d')

# 4d vs 4d
x = (np.random.randn(1, 4, 1, 6) > 0).astype(np.bool)
y = (np.random.randn(3, 1, 5, 6) > 0).astype(np.bool)
z = np.logical_xor(x, y)
expect(node, inputs=[x, y], outputs=[z],
       name='test_xor_bcast4v4d')

_experimental ATen

Experimental allowing ATen operations to be accessed directly from Caffe2 to allow for quick prototyping when ONNX is missing standard versions of and op

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Inputs (1 - ∞)

input (variadic) : T

Arbitrary input

Outputs (1 - ∞)

output (variadic) : T

Arbitrary output

Type Constraints

T : tensor(bool), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double)

Constrain output types to bool, int32, int64, float16, float, double tensors.

_experimental Affine

Affine takes one input data (Tensor) and produces one output data (Tensor) where the affine function, y = alpha * x + beta, is applied to the tensor elementwise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

alpha : float (default is 1.0)

Value of alpha

beta : float (default is 0.0)

Value of beta

Inputs

X : T

1D input tensor

Outputs

Y : T

1D output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental ConstantFill

The operator fills the elements of the output tensor with a constant value specified by the 'value' attribute.

The data type is specified by the 'dtype' attribute. The 'dtype' attribute must be one of the data types specified in the 'DataType' enum field in the TensorProto message. If the 'dtype' attribute is not provided, the data type of 'value' is used.

The output tensor shape is specified by the 'shape' attribute. If the number of input is 1, the shape will be identical to that of the input at run time with optional additional dimensions appended at the end as specified by 'extra_shape' attribute. In that case the 'shape' attribute should not be set.

If input_as_shape is set to true, then the input should be a 1D tensor containing the desired output shape (the dimensions specified in extra_shape will also be appended)

NOTE: Currently, it supports data type of float, int32, int64, and bool.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

dtype : int (default is 1)

The data type for the elements of the output tensor.Strictly must be one of the types from DataType enum in TensorProto.

extra_shape : list of ints

The additional dimensions appended at the end of the shape indicatedby the input blob.Cannot set the extra_shape argument when there is no input blob.

input_as_shape : int

1D tensor containing the desired output shape. First input must be in CPU context.

shape : list of ints

The shape of the output tensor. Cannot set the shape argument and pass in an input at the same time.

value : float (default is 0.0)

The value for the elements of the output tensor.

Inputs (0 - 1)

input (optional) : T1

Input tensor (optional) to provide shape information.

Outputs

output : T2

Output tensor of constant values specified by 'value'argument and its type is specified by the 'dtype' argument

Type Constraints

T1 : tensor(float), tensor(int32), tensor(int64), tensor(bool)

Constrain input types to float, int32, int64, bool tensors.

T2 : tensor(float), tensor(int32), tensor(int64), tensor(bool)

Constrain output types to float, int32, int64, bool tensors.

_experimental Crop

Crop and image to the specified spatial dimensions. If scale is given, then optionally start the crop offset by the left/top border amounts. If scale is not provided, crop the borders as provided.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

border : list of ints

A 1-D values of (leftBorder, topBorder, rightBorder, bottomBorder).

scale : list of ints

A 1-D values of (height, width).

Inputs

input : T

Input tensor of shape [N,C,H,W]

Outputs

output : T

Result, has same type as input, with H and W dimensions reduced.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental DynamicSlice

Produces a slice of the input tensor along multiple axes. Similar to numpy: https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html Slices uses axes, starts and ends inputs to specify the start and end dimension for each axis in the list of axes, it uses this information to slice the input data tensor. If a negative value is passed for any of the start or end indices, it represent number of elements before the end of that dimension. If the value passed to start or end is larger than the n (the number of elements in this dimension), it represents n. For slicing to the end of a dimension with unknown size, it is recommended to pass in INT_MAX. If axes are omitted, they are set to [0, ..., ndim-1]. Example 1: data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] axes = [0, 1] starts = [1, 0] ends = [2, 3] result = [ [5, 6, 7], ] Example 2: data = [ [1, 2, 3, 4], [5, 6, 7, 8], ] starts = [0, 1] ends = [-1, 1000] result = [ [2, 3, 4], ]

Version

This version of the operator has been available since version 9 of the default ONNX operator set.

Inputs (3 - 4)

data : T

Tensor of data to extract slices from.

starts : Tind

1-D tensor of starting indices of corresponding axis in `axes`

ends : Tind

1-D tensor of ending indices (exclusive) of corresponding axis in axes

axes (optional) : Tind

1-D tensor of axes that `starts` and `ends` apply to.

Outputs

output : T

Sliced data tensor.

Type Constraints

T : tensor(uint8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(int8), tensor(int16), tensor(int32), tensor(int64), tensor(float16), tensor(float), tensor(double), tensor(string), tensor(bool), tensor(complex64), tensor(complex128)

Constrain input and output types to all tensor types.

Tind : tensor(int32), tensor(int64)

Constrain indices to integer types

Examples

dynamic_slice

node = onnx.helper.make_node(
    'DynamicSlice',
    inputs=['x', 'starts', 'ends', 'axes'],
    outputs=['y'],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
y = x[0:3, 0:10]
starts = np.array([0, 0], dtype=np.int64)
ends = np.array([3, 10], dtype=np.int64)
axes = np.array([0, 1], dtype=np.int64)

expect(node, inputs=[x, starts, ends, axes], outputs=[y],
       name='test_dynamic_slice')

dynamic_slice_default_axes

node = onnx.helper.make_node(
    'DynamicSlice',
    inputs=['x', 'starts', 'ends'],
    outputs=['y'],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
starts = np.array([0, 0, 3], dtype=np.int64)
ends = np.array([20, 10, 4], dtype=np.int64)
y = x[:, :, 3:4]

expect(node, inputs=[x, starts, ends], outputs=[y],
       name='test_dynamic_slice_default_axes')

dynamic_slice_end_out_of_bounds

node = onnx.helper.make_node(
    'DynamicSlice',
    inputs=['x', 'starts', 'ends', 'axes'],
    outputs=['y'],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
starts = np.array([1], dtype=np.int64)
ends = np.array([1000], dtype=np.int64)
axes = np.array([1], dtype=np.int64)
y = x[:, 1:1000]

expect(node, inputs=[x, starts, ends, axes], outputs=[y],
       name='test_dynamic_slice_end_out_of_bounds')

dynamic_slice_neg

node = onnx.helper.make_node(
    'DynamicSlice',
    inputs=['x', 'starts', 'ends', 'axes'],
    outputs=['y'],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
starts = np.array([0], dtype=np.int64)
ends = np.array([-1], dtype=np.int64)
axes = np.array([1], dtype=np.int64)
y = x[:, 0:-1]

expect(node, inputs=[x, starts, ends, axes], outputs=[y],
       name='test_dynamic_slice_neg')

dynamic_slice_start_out_of_bounds

node = onnx.helper.make_node(
    'DynamicSlice',
    inputs=['x', 'starts', 'ends', 'axes'],
    outputs=['y'],
)

x = np.random.randn(20, 10, 5).astype(np.float32)
starts = np.array([1000], dtype=np.int64)
ends = np.array([1000], dtype=np.int64)
axes = np.array([1], dtype=np.int64)
y = x[:, 1000:1000]

expect(node, inputs=[x, starts, ends, axes], outputs=[y],
       name='test_dynamic_slice_start_out_of_bounds')

_experimental GRUUnit

GRUUnit computes the activations of a standard GRU, in a sequence-length aware fashion. Concretely, given the (fused) inputs X (TxNxD), the previous hidden state (NxD), and the sequence lengths (N), computes the GRU activations, avoiding computation if the input is invalid (as in, the value at X[t][n] >= seqLengths[n].

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

drop_states : int

Bool to determine if hidden state is zeroes or passed along for timesteps past the given sequence_length.

Inputs

hidden_prev : T

The previous GRU hidden state.

gates : T

Unactivated gate outputs from forget, update, and output gates, pre-activation.

seq_lengths : T

Array of sequence lengths. len(seq_lengths) should equal batch size N.

t : T

The timestep for this operation.

Outputs

hidden : T

The new GRU hidden state calculated by this op.

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental GivenTensorFill

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

extra_shape : list of ints

input_as_shape : int

shape : list of ints

values : list of floats

Inputs (0 - 1)

shape (optional) : T

The shape of filled tensor

Outputs

X : T

The filled tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental ImageScaler

Scale and bias the input image. Bias values are stored in the same ordering as the image pixel format.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

bias : list of floats

Bias applied to each channel, same size as C.

scale : float (default is 1.0)

The scale to apply.

Inputs

input : T

Input tensor of shape [N,C,H,W]

Outputs

output : T

Result, has same shape and type as input

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental ParametricSoftplus

ParametricSoftplus takes one input data (Tensor) and produces one output data (Tensor) where the softplus function, y = alpha * ln(exp(beta * x) + 1), is applied to the tensor elementwise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

alpha : float

Value of alpha

beta : float

Value of beta

Inputs

X : T

1D input tensor

Outputs

Y : T

1D input tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental Scale

Scale takes one input data (Tensor) and produces one output data (Tensor) whose value is the input data tensor scaled element-wise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

scale : float (default is 1.0)

The scale to apply.

Inputs

input : T

Input data to be scaled

Outputs

output : T

Output data after scaling

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental ScaledTanh

Calculates the scaled hyperbolic tangent of the given input tensor element-wise, alpha * tanh(beta * x).

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

alpha : float

Scaling value

beta : float

Scaling value

Inputs

input : T

Input tensor

Outputs

output : T

The scaled hyperbolic tangent values of the input tensor computed element-wise

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

_experimental ThresholdedRelu

ThresholdedRelu takes one input data (Tensor) and produces one output data (Tensor) where the rectified linear function, y = x for x > alpha, y = 0 otherwise, is applied to the tensor elementwise.

Version

This version of the operator has been available since version 1 of the default ONNX operator set.

Attributes

alpha : float (default is 1.0)

Threshold value

Inputs

X : T

Input tensor

Outputs

Y : T

Output tensor

Type Constraints

T : tensor(float16), tensor(float), tensor(double)

Constrain input and output types to float tensors.

Examples

default

default_alpha = 1.0
node = onnx.helper.make_node(
    'ThresholdedRelu',
    inputs=['x'],
    outputs=['y']
)
x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, default_alpha, np.inf)
y[y == default_alpha] = 0

expect(node, inputs=[x], outputs=[y],
       name='test_thresholdedrelu_default')

thresholdedrelu

alpha = 2.0
node = onnx.helper.make_node(
    'ThresholdedRelu',
    inputs=['x'],
    outputs=['y'],
    alpha=alpha
)

x = np.array([-1.5, 0., 1.2, 2.0, 2.2]).astype(np.float32)
y = np.clip(x, alpha, np.inf)  # expected output [0., 0., 0., 0., 2.2]
y[y == alpha] = 0

expect(node, inputs=[x], outputs=[y],
       name='test_thresholdedrelu_example')

x = np.random.randn(3, 4, 5).astype(np.float32)
y = np.clip(x, alpha, np.inf)
y[y == alpha] = 0

expect(node, inputs=[x], outputs=[y],
       name='test_thresholdedrelu')