handle edge cases

Author: bowang007

py/torch_tensorrt/dynamo/conversion/impl/argmax.py

@@ -1,47 +1,83 @@
-from typing import Optional
+from typing import Optional, Union
 
 import tensorrt as trt
 from torch.fx.node import Target
 from torch_tensorrt.dynamo._SourceIR import SourceIR
+from torch_tensorrt.dynamo.conversion._ConversionContext import ConversionContext
 from torch_tensorrt.dynamo.conversion.converter_utils import (
     cast_trt_tensor,
+    flatten_dims,
     get_axes_for_reduce_op,
 )
 from torch_tensorrt.fx.converters.converter_utils import (
     get_positive_dim,
     set_layer_name,
 )
-from torch_tensorrt.fx.types import TRTNetwork, TRTTensor
+from torch_tensorrt.fx.types import TRTTensor
 
 from . import squeeze
 
 
 def argmax(
-    network: TRTNetwork,
+    ctx: ConversionContext,
     target: Target,
     source_ir: Optional[SourceIR],
     name: str,
     input: TRTTensor,
-    dim: int = 0,
+    dim: Union[int, None],
     keep_dim: bool = False,
 ) -> TRTTensor:
     if not isinstance(input, TRTTensor):
         raise RuntimeError(
             f"argmax received input {input} that is not part " "of the TensorRT region!"
         )
+
     if input.dtype == trt.int32:
-        input = cast_trt_tensor(network, input, trt.float32, name)
-    if dim < 0:
+        input = cast_trt_tensor(ctx, input, trt.float32, name)
+
+    # Three different cases here:
+    # 1. dim == None, flatten input tensor first, keep_dim will be ignore and the output rank == input rank
+    # 2. input rank == 1: TopK layer does not support 1 dimensional topk operation. Broadcast input to rank == 2
+    # 3. normal cases, no additional handlings
+    out = input
+
+    if dim is None:
+        shuffle_layer = ctx.net.add_shuffle(input)
+        shuffle_layer.reshape_dims = (*flatten_dims(input, 0, -1), 1)
+        set_layer_name(shuffle_layer, target, name + "_flatten")
+        out = shuffle_layer.get_output(0)
+    elif len(input.shape) == 1:
+        shuffle_layer = ctx.net.add_shuffle(input)
+        shuffle_layer.reshape_dims = (*input.shape, 1)
+        set_layer_name(shuffle_layer, target, name + "_broadcast")
+        out = shuffle_layer.get_output(0)
+    elif dim < 0:
         dim = len(tuple(input.shape)) + dim
-    reduce_mask = get_axes_for_reduce_op(get_positive_dim(dim, len(input.shape)))
-    topk_layer = network.add_topk(input, trt.TopKOperation.MAX, 1, reduce_mask)
+
+    reduce_mask = get_axes_for_reduce_op(0)
+    if dim is not None:
+        reduce_mask = get_axes_for_reduce_op(get_positive_dim(dim, len(out.shape)))
+
+    topk_layer = ctx.net.add_topk(out, trt.TopKOperation.MAX, 1, reduce_mask)
     set_layer_name(topk_layer, target, name)
 
     out = topk_layer.get_output(1)
 
-    if not keep_dim:
+    if dim is None:
+        out_shuffle_layer = ctx.net.add_shuffle(out)
+        out_shuffle_layer.reshape_dims = (1,) * len(input.shape) if keep_dim else ()
+        set_layer_name(out_shuffle_layer, target, name + "_broadcast")
+        out = out_shuffle_layer.get_output(0)
+    elif len(input.shape) == 1:
         out = squeeze.squeeze(
-            network, target, SourceIR.ATEN, name + "_squeeze", out, dim
+            ctx,
+            target,
+            SourceIR.ATEN,
+            name + "_squeeze",
+            out,
+            1 if keep_dim else [0, 1],
         )
+    elif not keep_dim:
+        out = squeeze.squeeze(ctx, target, SourceIR.ATEN, name + "_squeeze", out, dim)
 
     return out

tests/py/dynamo/conversion/test_argmax_aten.py

@@ -0,0 +1,40 @@
+import torch
+import torch.nn as nn
+from parameterized import parameterized
+from torch.testing._internal.common_utils import run_tests
+
+from .harness import DispatchTestCase
+
+
+class TestArgmaxConverter(DispatchTestCase):
+    @parameterized.expand(
+        [
+            # input dimension == 1
+            ("dim_1_keep_dim_true", (3,), 0, True),
+            ("dim_1_keep_dim_true", (3,), 0, False),
+            # dim == None
+            ("dim_none", (3,), None, True),
+            ("dim_none", (3, 3), None, True),
+            ("dim_none", (3, 3, 3), None, False),
+            # # common cases
+            ("dim_1_keep_dim_true", (3, 3), 1, True),
+            ("dim_1_keep_dim_false", (3, 3), 1, False),
+            ("dim_0_keep_dim_true", (4, 4, 4), 0, True),
+            ("dim_0_keep_dim_false", (4, 4, 4), 0, False),
+        ]
+    )
+    def test_argmax(self, _, input_shape, dim, keep_dim):
+        class ArgMax(nn.Module):
+            def __init__(self):
+                super().__init__()
+
+            def forward(self, input):
+                return torch.ops.aten.argmax.default(input, dim, keep_dim)
+
+        input = [torch.randn(*input_shape)]
+
+        self.run_test(ArgMax(), input)
+
+
+if __name__ == "__main__":
+    run_tests()