diff options
author | Yan Facai (颜发才) <facai.yan@gmail.com> | 2018-08-14 15:05:02 +0800 |
---|---|---|
committer | Yan Facai (颜发才) <facai.yan@gmail.com> | 2018-08-14 15:05:02 +0800 |
commit | 8f4bb0fed956a35b767c0984d9587636cba661bb (patch) | |
tree | bcf8460276aeb730449afec0d930818ac73f4163 /tensorflow/cc/gradients | |
parent | 0552c2976cf6e4f0f09556d0a3ae5a76509e9a46 (diff) | |
parent | cb53bfaf48588962f5799391d3a7a411dda72b49 (diff) |
Merge remote-tracking branch 'upstream/master' into ENH/unsafe_div
Diffstat (limited to 'tensorflow/cc/gradients')
-rw-r--r-- | tensorflow/cc/gradients/array_grad.cc | 18 | ||||
-rw-r--r-- | tensorflow/cc/gradients/array_grad_test.cc | 8 | ||||
-rw-r--r-- | tensorflow/cc/gradients/math_grad.cc | 20 | ||||
-rw-r--r-- | tensorflow/cc/gradients/math_grad_test.cc | 11 |
4 files changed, 56 insertions, 1 deletions
diff --git a/tensorflow/cc/gradients/array_grad.cc b/tensorflow/cc/gradients/array_grad.cc index b353accddc..e9173227aa 100644 --- a/tensorflow/cc/gradients/array_grad.cc +++ b/tensorflow/cc/gradients/array_grad.cc @@ -120,6 +120,24 @@ Status SplitGrad(const Scope& scope, const Operation& op, } REGISTER_GRADIENT_OP("Split", SplitGrad); +Status FillGrad(const Scope& scope, const Operation& op, + const std::vector<Output>& grad_inputs, + std::vector<Output>* grad_outputs) { + // y = fill(fill_shape, x) + // No gradient returned for the fill_shape argument. + grad_outputs->push_back(NoGradient()); + // The gradient for x (which must be a scalar) is just the sum of + // all the gradients from the shape it fills. + // We use ReduceSum to implement this, which needs an argument providing + // the indices of all the dimensions of the incoming gradient. + // grad(x) = reduce_sum(grad(y), [0..rank(grad(y))]) + auto all_dims = Range(scope, Const(scope, 0), Rank(scope, grad_inputs[0]), + Const(scope, 1)); + grad_outputs->push_back(ReduceSum(scope, grad_inputs[0], all_dims)); + return scope.status(); +} +REGISTER_GRADIENT_OP("Fill", FillGrad); + Status DiagGrad(const Scope& scope, const Operation& op, const std::vector<Output>& grad_inputs, std::vector<Output>* grad_outputs) { diff --git a/tensorflow/cc/gradients/array_grad_test.cc b/tensorflow/cc/gradients/array_grad_test.cc index d09275b648..f41de3dc20 100644 --- a/tensorflow/cc/gradients/array_grad_test.cc +++ b/tensorflow/cc/gradients/array_grad_test.cc @@ -108,6 +108,14 @@ TEST_F(ArrayGradTest, SplitGrad) { RunTest({x}, {x_shape}, y.output, {y_shape, y_shape}); } +TEST_F(ArrayGradTest, FillGrad) { + TensorShape x_shape({}); + auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)); + TensorShape y_shape({2, 5, 3}); + auto y = Fill(scope_, {2, 5, 3}, x); + RunTest(x, x_shape, y, y_shape); +} + TEST_F(ArrayGradTest, DiagGrad) { TensorShape x_shape({5, 2}); auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)); diff --git a/tensorflow/cc/gradients/math_grad.cc b/tensorflow/cc/gradients/math_grad.cc index cd215f740d..1329b568ab 100644 --- a/tensorflow/cc/gradients/math_grad.cc +++ b/tensorflow/cc/gradients/math_grad.cc @@ -1022,6 +1022,26 @@ Status ProdGrad(const Scope& scope, const Operation& op, } REGISTER_GRADIENT_OP("Prod", ProdGrad); +Status SegmentSumGrad(const Scope& scope, const Operation& op, + const std::vector<Output>& grad_inputs, + std::vector<Output>* grad_outputs) { + // The SegmentSum operation sums segments of the Tensor that have the same + // index in the segment_ids parameter. + // i.e z = [2, 3, 4, 5], segment_ids [0, 0, 0, 1] + // will produce [2 + 3 + 4, 5] = [9, 5] + // The gradient that will flow back to the gather operation will look like + // [x1, x2], it will have the same shape as the output of the SegmentSum + // operation. The differentiation step of the SegmentSum operation just + // broadcast the gradient in order to retrieve the z's shape. + // dy/dz = [x1, x1, x1, x2] + grad_outputs->push_back(Gather(scope, grad_inputs[0], op.input(1))); + + // stop propagation along segment_ids + grad_outputs->push_back(NoGradient()); + return scope.status(); +} +REGISTER_GRADIENT_OP("SegmentSum", SegmentSumGrad); + // MatMulGrad helper function used to compute two MatMul operations // based on input matrix transposition combinations. Status MatMulGradHelper(const Scope& scope, const bool is_batch, diff --git a/tensorflow/cc/gradients/math_grad_test.cc b/tensorflow/cc/gradients/math_grad_test.cc index 147428cc39..c16938322c 100644 --- a/tensorflow/cc/gradients/math_grad_test.cc +++ b/tensorflow/cc/gradients/math_grad_test.cc @@ -45,10 +45,10 @@ using ops::Placeholder; using ops::Pow; using ops::Prod; using ops::RealDiv; +using ops::SegmentSum; using ops::SquaredDifference; using ops::Sub; using ops::Sum; -using ops::Where3; // TODO(andydavis) Test gradient function against numeric gradients output. // TODO(andydavis) As more gradients are added move common test functions @@ -932,5 +932,14 @@ TEST_F(NaryGradTest, Prod) { RunTest({x}, {x_shape}, {y}, {y_shape}); } +TEST_F(NaryGradTest, SegmentSum) { + TensorShape x_shape({3, 4}); + auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)); + auto y = SegmentSum(scope_, x, {0, 0, 1}); + // the sum is always on the first dimension + TensorShape y_shape({2, 4}); + RunTest({x}, {x_shape}, {y}, {y_shape}); +} + } // namespace } // namespace tensorflow |