| Status | Accepted |
|---|---|
| RFC # | 335 |
| Author(s) | Saurabh Saxena (Google) |
| Sponsor | Rohan Jain (Google) |
| Updated | 2020-12-01 |
We propose performing gradient computation entirely in C++. We mainly focus on reverse-mode autodiff here and leave forward-mode AD as future work. A C API for this implementation is also left as future work but we imagine it to be a straightforward wrapper over the provided C++ APIs. The APIs discussed here are experimental and hence do not have any backwards compatibility guarantees yet.
Performance: Running op gradient functions in Python means we incur python overheads for a transformation pass that could entirely be done in C/C++. This most severely affects eager mode training of compute-light pipelines as we show in the performance impact section.
Support C++ training: Lack of autodiff support in C/C++ could be detrimental to high performance training pipelines at Google scale. This has often come up in discussions internally.
Autodiff in non-python TF: Relying on python gradient functions means we cannot support eager-mode autodiff from other language front-ends such as Java.
In addition, we try to address some shortcomings of the current GradientTape design:
- The tape is currently in-charge of filling zeros for missing incoming gradients so that the gradient functions do not need any None-grad checks. This is sub-optimal since the tape now needs to keep around additional metadata(or actual tensors) from the forward pass around in order to be able to create these zeros. To prevent excessive memory usage we have a hand-curated list of ops with indices of incoming gradients that are allowed to be None. This list is highly likely incomplete and is hard to keep up to date. We address this by moving this logic from the tape to individual gradient functions. Since most ops have a single output this is usually a no-op.
- We have a similar problem with inputs/outputs of the forward op. Since the tape does not know which of the op’s inputs/outputs the gradient function might require, it will need to keep all tensors around which can be prohibitive. To limit the memory usage here we auto-generate a list of ops with indices of inputs/outputs that can be dropped by parsing ASTs of gradient functions. This again indicates that this should be handled on a per-op level. We address this by allowing individual gradient function instances to incref needed tensors from the forward pass.
The gradients infrastructure will be built on top of the abstract interfaces for op execution which provide a backend agnostic way of tracing and executing ops. We provide APIs for authoring GradientFunctions and registering them into a GradientRegistry for name based lookup. We provide a gradient Tape API that is close to python’s tf.GradientTape and shares most of the implementation with the existing tape.
An op’s gradient is defined by subclassing GradientFunction
class GradientFunction {
public:
virtual Status Compute(
AbstractContext* ctx,
absl::Span<AbstractTensorHandle* const> grad_outputs,
absl::Span<AbstractTensorHandle*> grad_inputs) = 0;
virtual ~GradientFunction() {}
};
GradientFunction::Compute receives gradients wrt op’s outputs in grad_outputs and needs to populate gradients wrt op’s inputs in grad_inputs. This is the same signature we use for authoring python gradients with the addition of an AbstractContext, which provides an API creating operations (eagerly or traced). In python this context is stored in a global variable and is implicitly captured. For the C++ API we chose to pass this context explicitly.
The reason GradientFunction is a class and not a callable is so that each op’s gradient function can keep the necessary state needed from forward pass for the gradient computation (see ExpGradientFunction below for an example).
Examples:
class AddGradientFunction : public GradientFunction {
public:
Status Compute(AbstractContext* ctx,
absl::Span<AbstractTensorHandle* const> grad_outputs,
absl::Span<AbstractTensorHandle*> grad_inputs) override {
// Tape never calls a gradient function if there are no incoming grads.
DCHECK(grad_outputs[0]);
grad_inputs[0] = grad_outputs[0];
grad_inputs[1] = grad_outputs[0];
grad_inputs[0]->Ref();
grad_inputs[1]->Ref();
return Status::OK();
}
~AddGradientFunction() override {}
};
class ExpGradientFunction : public GradientFunction {
public:
explicit ExpGradientFunction(AbstractTensorHandle* exp) : exp_(exp) {
exp->Ref();
}
Status ExpGradientFunction::Compute(
AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> grad_outputs,
absl::Span<AbstractTensorHandle*> grad_inputs) {
vector<AbstractTensorHandle*> conj_outputs(1);
TF_RETURN_IF_ERROR(
Conj(ctx, {exp_.get()}, absl::MakeSpan(conj_outputs), "Conj_Exp_Grad"));
AbstractTensorHandlePtr conj_output_releaser(conj_outputs[0]);
TF_RETURN_IF_ERROR(
Mul(ctx, {conj_outputs[0], grad_outputs[0]}, grad_inputs, "Mul_Exp_Grad"));
return Status::OK();
}
~ExpGradientFunction() override {}
private:
AbstractTensorHandlePtr exp_;
};
C++ gen_ops
Authoring gradient functions requires calling elementary ops in C++. Using low level op building APIs everywhere would be very verbose so we plan to auto-generate C++ APIs for op-building using registered OpDefs, similar to how we do for python. This is part of a larger TensorFlow core team’s effort to provide efficient python bindings which are light wrappers for C++ API calls. For now we have been handwriting C++ op building functions using the unified execution APIs. Once ready we will replace these with auto-generated C++ op functions.
We provide a registry to store the mapping from op type to factory functions that return the GradientFunction for an op’s instance. The factory function takes as input the ForwardOperation, which contains metadata from the forward operation, and returns a GradientFunction. This allows gradient function authors to control which inputs/outputs of the forward op to keep around by increasing the ref-count on AbstractTensorHandle.
Additionally, we provide a utility function RegisterNotDifferentiable to mark an op as non-differentiable. This can be used to implement tf.no_gradient. We also provide a NotDifferentiableGradientFunction which returns nullptr output gradients. This can be used to implement tf.stop_gradient.
struct ForwardOperation {
public:
string op_name;
std::vector<AbstractTensorHandle*> inputs;
std::vector<AbstractTensorHandle*> outputs;
std::vector<int64> skip_input_indices;
AttrBuilder attrs;
};
using GradientFunctionFactory = std::function<GradientFunction*(const ForwardOperation& op)>;
// Map from op name to a GradientFunctionFactory.
class GradientRegistry {
public:
Status Register(const string& op,
GradientFunctionFactory gradient_function_factory);
Status Lookup(const ForwardOperation& op,
std::unique_ptr<GradientFunction>* gradient_function) const;
private:
absl::flat_hash_map<string, GradientFunctionFactory> registry_;
};
class NotDifferentiableGradientFunction : public GradientFunction {
Status Compute(
AbstractContext* ctx,
absl::Span<AbstractTensorHandle* const> grad_outputs,
absl::Span<AbstractTensorHandle*> grad_inputs) override;
};
Status RegisterNotDifferentiable(GradientRegistry* registry, const string& op);
Examples:
GradientFunction* AddRegisterer(const ForwardOperation& op) {
return new AddGradientFunction;
}
GradientFunction* ExpRegisterer(const ForwardOperation& op) {
return new ExpGradientFunction(op.outputs[0]);
}
The API for C++ Tape is very similar to python’s tf.GradientTape. The implementation for this interface is almost entirely shared with the C++ tape in c/eager/tape.h.
class Tape {
public:
explicit Tape::Tape(bool is_persistent);
// Adds this tensor to the list of watched tensors.
//
// This is a no-op if the tensor is already being watched either from an
// earlier call to GradientTape::Watch or being an output of an op with
// watched inputs.
void Watch(const AbstractTensorHandle* tensor);
// Records an operation with given inputs and outputs
// on the tape and marks all its outputs as watched if at
// least one input of the op is watched and has a trainable dtype.
// op_name is optional and is used for debugging only.
void RecordOperation(absl::Span<AbstractTensorHandle* const> inputs,
absl::Span<AbstractTensorHandle* const> outputs,
GradientFunction* gradient_function,
const string& op_name = "");
// Returns whether any tensor in a list of tensors is being watched and has
// a trainable dtype.
bool ShouldRecord(absl::Span<const AbstractTensorHandle* const> tensors) const;
// Unwatches this tensor on the tape. Mainly used for cleanup when deleting
// eager tensors.
void DeleteTrace(const AbstractTensorHandle*);
// Computes the gradient and stores the result in result.
// Raises an error if result is not the same length as sources.
Status ComputeGradient(
AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> targets,
absl::Span<AbstractTensorHandle* const> sources,
absl::Span<AbstractTensorHandle* const> output_gradients,
absl::Span<AbstractTensorHandle*> result);
// Returns whether the tape is persistent, i.e., whether the tape will hold
// onto its internal state after a call to ComputeGradient.
bool IsPersistent() const;
};
Example:
// Computes
// y = tf.exp(inputs[0])
// outputs = grad(y, inputs)
Status ExpGradModel(AbstractContext* ctx,
absl::Span<AbstractTensorHandle* const> inputs,
absl::Span<AbstractTensorHandle*> outputs,
const GradientRegistry& registry) {
Tape tape(/*persistent=*/false);
tape.Watch(inputs[0]);
std::vector<AbstractTensorHandle*> exp_outputs(1);
TF_RETURN_IF_ERROR(ops::Exp(ctx, inputs, absl::MakeSpan(exp_outputs), "Exp"));
std::unique_ptr<GradientFunction> gradient_function;
ForwardOperation forward_op;
forward_op.op_name = “Exp”;
forward_op.inputs.push_back(inputs[0]);
forward_op.outputs.push_back(exp_outputs[0]);
TF_RETURN_IF_ERROR(registry.Lookup(forward_op, &gradient_function));
tape.RecordOperation(inputs, exp_outputs, gradient_function.release());
TF_RETURN_IF_ERROR(tape.ComputeGradient(ctx,
/*targets*/exp_outputs,
/*sources=*/inputs,
/*output_gradients=*/{},
/*result=*/outputs));
for (auto exp_output : exp_outputs) {
exp_output->Unref();
}
return Status::OK();
}
TapeContext
In the final C++ API, we don’t expect users to have to directly call RecordOperation for each op. We would provide an AbstractContext implementation for the tape which would trace ops on the tape and delegate execution to a backing context. A prototype for this is available here.
class TapeContext : public AbstractContext {
public:
explicit TapeContext(AbstractContext* parent_ctx, Tape*, const GradientRegistry&);
// Skipping overridden methods.
private:
AbstractContext* parent_ctx_;
Tape* tape_;
const GradientRegistry& registry_;
};
class TapeOperation : public AbstractOperation {
public:
explicit TapeOperation(AbstractOperation* parent_op, Tape*, const GradientRegistry&);
// Skipping overridden methods.
private:
AbstractOperation* parent_op_;
ForwardOperation forward_op_;
Tape* tape_;
const GradientRegistry& registry_;
};
TapeOperation would populate the ForwardOperation object and record the operation on the tape in the call to AbstractOperation::Execute:
Status TapeOperation::AddInput(AbstractTensorHandle* input) {
TF_RETURN_IF_ERROR(parent_op_->AddInput(input));
forward_op_.inputs.push_back(input);
return Status::OK();
}
Status TapeOperation::Execute(absl::Span<AbstractTensorHandle*> retvals,
int* num_retvals) {
TF_RETURN_IF_ERROR(parent_op_->Execute(retvals, num_retvals));
for (int i = 0; i < *num_retvals; i++) {
forward_op_.outputs.push_back(retvals[i]);
}
// Populate forward_op_.skip_input_indices here.
std::unique_ptr<GradientFunction> backward_fn;
TF_RETURN_IF_ERROR(registry_.Lookup(forward_op_, &backward_fn));
tape_->RecordOperation(forward_op_.inputs, forward_op_.outputs,
backward_fn.release(), parent_op_->Name());
return Status::OK();
}
This way the same C++ gen_ops code can be used to execute ops with/without a tape by simply wrapping the current execution context in a TapeContext.
Some details on memory management
AbstractTensorHandle provides Ref and Unref methods which can be used to manage its lifecycle. Gradient functions and the tape follow these guidelines for memory safety:
- The tape releases tensors returned from gradient functions after they have been consumed.
- Gradient functions which pass-through incoming gradients e.g. AddGradientFunction in the example above, should Ref each copy of the incoming grad since the tape is free to release the incoming grads passed to AddGradientFunction after it has been called.
- It is the user’s responsibility to release tensors that are passed in as default gradients in the call to ComputeGradients.
- More trivially, a gradient function should release any intermediate tensor that is created by it but not returned as an output gradient. This includes outputs of intermediate operations and operations added to the forward pass for reductions e.g. shape computation.
If manual management of ref-counts becomes too cumbersome we could consider adding a stack allocated wrapper for AbstractTensorHandle* that manages its life-cycle. However, this change should probably happen at the level of the unified execution APIs.
A custom GradientFunction for a set of inputs/outputs can be registered using Tape::RecordOperation similar to a gradient function looked up from the gradient registry.
Example:
class CustomGradientFunction: public GradientFunction {
public:
Status ExpGradientFunction::Compute(
AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> grad_outputs,
absl::Span<AbstractTensorHandle*> grad_inputs) {
// Populate grad_inputs.
return Status::OK();
}
~CustomGradientFunction() override {}
};
Status ExpWithCustomGrad(AbstractContext* ctx,
absl::Span<AbstractTensorHandle* const> inputs,
absl::Span<AbstractTensorHandle*> outputs) {
Tape tape(/*persistent=*/false);
tape.Watch(inputs[0]);
std::vector<AbstractTensorHandle*> exp_outputs(1);
TF_RETURN_IF_ERROR(ops::Exp(ctx, inputs, absl::MakeSpan(exp_outputs), "Exp"));
std::unique_ptr<GradientFunction> gradient_function(new CustomGradientFunction);
tape.RecordOperation(inputs, exp_outputs, gradient_function.release());
TF_RETURN_IF_ERROR(tape.ComputeGradient(ctx,
/*targets*/exp_outputs,
/*sources=*/inputs,
/*output_gradients=*/{},
/*result=*/outputs));
for (auto exp_output : exp_outputs) {
exp_output->Unref();
}
return Status::OK();
}
tf.recompute_grad is an application of tf.custom_gradient where we do not record the forward pass on the tape so that we are not holding on to forward pass tensors in memory. (In tf.custom_gradient we allow recording the forward pass on the tape in order for higher-order derivatives to work for cases where the custom gradient function uses intermediate tensors from the forward pass.) This is implemented by executing the forward pass outside the tape (managed by a higher layer) and registering a gradient function that re-runs the forward pass and computes gradients. The same behavior can be achieved using this tape.
Higher order derivatives are computed by either using a persistent tape or by tracing computations on nested tapes. The nesting is managed by a higher layer. This can be implemented using this tape.
A small set of python gradient functions have been optimized to not return gradients for inputs which are not tracked under the tape. This is beneficial in eager mode where unneeded gradients cannot be pruned during execution. In the C++ tape, we support this by providing a skip_input_indices field on the ForwardOperation which stores the list of input indices which are not watched.
In python, if a variable is accessed inside a tf.GradientTapes scope it is automatically tracked, i.e. Tape::Watch is called for the DT_RESOURCE tensor backing the variable on behalf of the user. For now we will leave this out as a higher layer feature and require that variable handles are explicitly tracked by a higher layer. We can revisit this later if needed.
Eventually we plan to implement tf.function and functional control flow gradients in C++ but for now we will leave them in python and directly call the python gradient function.
Gradient function of a gather returns IndexedSlices for efficiency. We need to support IndexedSlices as part of the input and output gradients of a gradient function. Currently there is no good C++ representation for these. One possible representation would be to wrap the component tensors in an IndexedSlicesTensorHandle that subclasses AbstractTensorHandle. This way IndexedSlices would be transparent to the tape. The C++ gen ops can choose to handle IndexedSlices appropriately or simply densify them by calling a C++ equivalent of convert_to_tensor.
class IndexedSlicesTensorHandle : public AbstractTensorHandle {
protected:
explicit IndexedSlicesTensorHandle(AbstractTensorHandle* values,
AbstractTensorHandle* indices,
const PartialTensorShape& dense_shape)
: AbstractTensorHandle(kIndexedSlices),
values_(values),
indices_(indices),
dense_shape_(dense_shape) {
values->Ref();
indices->Ref();
}
// Returns tensor dtype.
DataType DataType() override {
return values_->DataType();
}
// Returns tensor shape.
Status Shape(PartialTensorShape* shape) override {
*shape = dense_shape_;
}
public:
// For LLVM style RTTI.
static bool classof(const AbstractTensorHandle* ptr) {
return ptr->getKind() == kIndexedSlices;
}
~IndexedSlicesTensorHandle() {
values_->Unref();
indices_->Unref();
}
private:
AbstractTensorHandle* values_;
AbstractTensorHandle* indices_;
PartialTensorShape dense_shape_;
}
Because gradient operations will no longer be run in Python, we expect the python framework overhead for usual NN models to be halved. This should substantially improve both the graph building time and time to compute gradients in eager mode.
Initial tests on some MLPs of various depths for MNIST show ~2x speedup in eager mode.
The framework is a fairly lightweight implementation of the existing Tape interface in c/eager/tape.h which was already templated to support different C++ types for gradient functions and tensors. We have been making necessary improvements to the base framework to support this project, e.g., moving default zeros creation logic from the tape to respective gradient functions.
We plan to implement gradient functions under tensorflow/c/gradients. As a proof-of-concept we implemented an MLP for MNIST using an experimental python binding (see python/framework/experimental/tape.py). For that we implemented gradient functions for MatMul, Add, ReLu and Softmax. We are currently working on implementing gradient functions needed for ResNet50.
We further plan to publish a guide for inviting contributions and setup a spreadsheet or some such for tracking completeness.
We plan to rollout C++ gradient functions incrementally. We will port the existing pybind C++ tape to use the new tape implementation. The GradientFunction for ops with registered C++ gradients will directly be called. For others, we will simply register a GradientFunction that calls the python gradient function.
@alextp motivated the design and provided an initial prototype for this project. @amturati implemented various gradient functions to get a MLP training on MNIST. @vnvo2409 has been working on making framework improvements and further implementing C++ gradient functions.