History

Siyavash Najafzade 3744d26060 Implement manip gradient in cc.		2020-03-19 15:49:42 +03:30
..
array_grad_test.cc	Add gradient for tensorflow::ops::Fill	2018-07-18 08:35:54 -07:00
array_grad.cc	Temporary rollback of QuantizeAndDequantizeV2Grad.	2020-02-21 14:06:30 -08:00
data_flow_grad_test.cc
data_flow_grad.cc
grad_testutil.cc	Remove using directives. Test appertaining to ops have been moved into namespace tensorflow::ops; all other tests now use explicit using-declarations.	2018-01-02 10:49:13 -08:00
grad_testutil.h
image_grad_test.cc
image_grad.cc	Adds python gradients for ScaleAndTranslate ops, which are now used internal by tf.resize_images_v2 op.	2019-03-14 21:08:05 -07:00
manip_grad_test.cc	Implement manip gradient in cc.	2020-03-19 15:49:42 +03:30
manip_grad.cc	Implement manip gradient in cc.	2020-03-19 15:49:42 +03:30
math_grad_test.cc
math_grad.cc	Define _USE_MATH_DEFINES for windows builds.	2019-12-20 14:43:49 -08:00
nn_grad_test.cc	disabling subtests that test 3D pooling ops, and removing the no_rocm tag from //tensorflow/cc:gradients_nn_grad_test	2019-12-24 20:49:51 -08:00
nn_grad.cc	minor spelling tweaks	2020-01-16 14:36:52 +09:00
README.md

README.md

C++ gradients

Gradients are currently being ported from python to C++ (in this directory).

Contributions are welcome and much appreciated; please follow the instructions below.

Create the op gradient function in foo_grad.cc corresponding to the foo_grad.py file where the op originated (i.e. array_grad.py op gradients should be written in array_grad.cc).

Write the op gradient with the following naming scheme:

Status OpNameGrad(const Scope& scope, const Operation& op,
                  const std::vector<Output>& grad_inputs,
                  std::vector<Output>* grad_outputs) {
  ...
  return scope.status();
}
REGISTER_GRADIENT_OP("OpName", OpNameGrad);

Ops gradients are implemented by using the C++ API.
Tests should be included in foo_grad_test.cc. Please see array_grad_test.cc for an many examples. Tests are as simple as, creating a placeholder input for the op's inputs and calling RunTest (RunTest uses a gradient checker to verify that the theoretical gradient matches the numeric gradient). For example:
```
TEST_F(ArrayGradTest, IdentityGrad) {
  TensorShape shape({5, 2});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
  auto y = Identity(scope_, x);
  RunTest(x, shape, y, shape);
}
```

NOTE: There are some ops that require features from the C++ API that are not yet implemented.

Ops that require PartialTensorShape information cannot yet be implemented.
Ops that require SparseTensor or IndexSlices (currently only in python) cannot yet be implemented.
Maybe more.

For questions: Please create an issue assigned to suharshs.