diff options
Diffstat (limited to 'tensorflow/cc')
-rw-r--r-- | tensorflow/cc/BUILD | 31 | ||||
-rw-r--r-- | tensorflow/cc/client/client_session.cc | 18 | ||||
-rw-r--r-- | tensorflow/cc/client/client_session.h | 28 | ||||
-rw-r--r-- | tensorflow/cc/client/client_session_test.cc | 21 | ||||
-rw-r--r-- | tensorflow/cc/framework/gradient_checker.cc | 12 | ||||
-rw-r--r-- | tensorflow/cc/framework/gradient_checker_test.cc | 16 | ||||
-rw-r--r-- | tensorflow/cc/gradients/image_grad.cc | 74 | ||||
-rw-r--r-- | tensorflow/cc/gradients/image_grad_test.cc | 157 | ||||
-rw-r--r-- | tensorflow/cc/gradients/math_grad_test.cc | 6 | ||||
-rw-r--r-- | tensorflow/cc/gradients/nn_grad.cc | 85 | ||||
-rw-r--r-- | tensorflow/cc/gradients/nn_grad_test.cc | 22 | ||||
-rw-r--r-- | tensorflow/cc/saved_model/BUILD | 13 | ||||
-rw-r--r-- | tensorflow/cc/saved_model/loader.cc | 95 |
13 files changed, 520 insertions, 58 deletions
diff --git a/tensorflow/cc/BUILD b/tensorflow/cc/BUILD index a98f0b00b2..588a45ea43 100644 --- a/tensorflow/cc/BUILD +++ b/tensorflow/cc/BUILD @@ -121,6 +121,7 @@ cc_library( deps = [ ":array_grad", ":data_flow_grad", + ":image_grad", ":math_grad", ":nn_grad", ], @@ -332,6 +333,36 @@ tf_cc_test( ) cc_library( + name = "image_grad", + srcs = ["gradients/image_grad.cc"], + deps = [ + ":cc_ops", + ":cc_ops_internal", + ":grad_op_registry", + ":gradients", + ], + alwayslink = 1, +) + +tf_cc_test( + name = "gradients_image_grad_test", + srcs = ["gradients/image_grad_test.cc"], + deps = [ + ":cc_ops", + ":client_session", + ":grad_op_registry", + ":grad_testutil", + ":gradient_checker", + ":image_grad", + ":testutil", + "//tensorflow/core:lib_internal", + "//tensorflow/core:test", + "//tensorflow/core:test_main", + "//tensorflow/core:testlib", + ], +) + +cc_library( name = "math_grad", srcs = ["gradients/math_grad.cc"], deps = [ diff --git a/tensorflow/cc/client/client_session.cc b/tensorflow/cc/client/client_session.cc index ba056a8f3a..0e61089a59 100644 --- a/tensorflow/cc/client/client_session.cc +++ b/tensorflow/cc/client/client_session.cc @@ -127,4 +127,22 @@ Status ClientSession::Run(const RunOptions& run_options, const FeedType& inputs, target_node_names, outputs, run_metadata); } +Status ClientSession::MakeCallable(const CallableOptions& callable_options, + CallableHandle* out_handle) { + TF_RETURN_IF_ERROR(impl()->MaybeExtendGraph()); + return impl()->session_->MakeCallable(callable_options, out_handle); +} + +Status ClientSession::RunCallable(CallableHandle handle, + const std::vector<Tensor>& feed_tensors, + std::vector<Tensor>* fetch_tensors, + RunMetadata* run_metadata) { + return impl()->session_->RunCallable(handle, feed_tensors, fetch_tensors, + run_metadata); +} + +Status ClientSession::ReleaseCallable(CallableHandle handle) { + return impl()->session_->ReleaseCallable(handle); +} + } // end namespace tensorflow diff --git a/tensorflow/cc/client/client_session.h b/tensorflow/cc/client/client_session.h index 5fb4109f7d..7dd653eec4 100644 --- a/tensorflow/cc/client/client_session.h +++ b/tensorflow/cc/client/client_session.h @@ -87,7 +87,33 @@ class ClientSession { const std::vector<Operation>& run_outputs, std::vector<Tensor>* outputs, RunMetadata* run_metadata) const; - // TODO(keveman): Add support for partial run. + /// \brief A handle to a subgraph, created with + /// `ClientSession::MakeCallable()`. + typedef int64 CallableHandle; + + /// \brief Creates a `handle` for invoking the subgraph defined by + /// `callable_options`. + /// NOTE: This API is still experimental and may change. + Status MakeCallable(const CallableOptions& callable_options, + CallableHandle* out_handle); + + /// \brief Invokes the subgraph named by `handle` with the given options and + /// input tensors. + /// + /// The order of tensors in `feed_tensors` must match the order of names in + /// `CallableOptions::feed()` and the order of tensors in `fetch_tensors` will + /// match the order of names in `CallableOptions::fetch()` when this subgraph + /// was created. + /// NOTE: This API is still experimental and may change. + Status RunCallable(CallableHandle handle, + const std::vector<Tensor>& feed_tensors, + std::vector<Tensor>* fetch_tensors, + RunMetadata* run_metadata); + + /// \brief Releases resources associated with the given `handle` in this + /// session. + /// NOTE: This API is still experimental and may change. + Status ReleaseCallable(CallableHandle handle); private: class Impl; diff --git a/tensorflow/cc/client/client_session_test.cc b/tensorflow/cc/client/client_session_test.cc index ea5cf5a1f1..559ffea7e8 100644 --- a/tensorflow/cc/client/client_session_test.cc +++ b/tensorflow/cc/client/client_session_test.cc @@ -95,5 +95,26 @@ TEST(ClientSessionTest, MultiThreaded) { test::ExpectTensorEqual<int>(outputs[0], test::AsTensor<int>({-1, 2}, {2})); } +TEST(ClientSessionTest, Callable) { + Scope root = Scope::NewRootScope(); + auto a = Placeholder(root, DT_INT32); + auto b = Placeholder(root, DT_INT32); + auto c = Add(root, a, b); + ClientSession session(root); + std::vector<Tensor> outputs; + + CallableOptions options; + options.add_feed(a.node()->name()); + options.add_feed(b.node()->name()); + options.add_fetch(c.node()->name()); + ClientSession::CallableHandle callable; + TF_CHECK_OK(session.MakeCallable(options, &callable)); + TF_EXPECT_OK(session.RunCallable( + callable, {test::AsTensor<int>({1}, {}), test::AsTensor<int>({41}, {})}, + &outputs, nullptr)); + test::ExpectTensorEqual<int>(outputs[0], test::AsTensor<int>({42}, {})); + TF_EXPECT_OK(session.ReleaseCallable(callable)); +} + } // namespace } // namespace tensorflow diff --git a/tensorflow/cc/framework/gradient_checker.cc b/tensorflow/cc/framework/gradient_checker.cc index de2645cb44..e9f9c59e3a 100644 --- a/tensorflow/cc/framework/gradient_checker.cc +++ b/tensorflow/cc/framework/gradient_checker.cc @@ -247,7 +247,7 @@ Status ComputeNumericJacobianTranspose(const Scope& scope, const OutputList& xs, auto y_pos_flat = y_pos[y_idx].flat<Y_T>(); auto y_neg_flat = y_neg[y_idx].flat<Y_T>(); const int64 y_size = y_shapes[y_idx].num_elements(); - const Y_T scale = Y_T{2 * delta}; + const Y_T scale = 2 * delta; auto jacobian = (*jacobian_ts)[x_idx * y_num + y_idx].matrix<JAC_T>(); for (int c = 0; c < y_size; ++c) { SetJacobian<Y_T, JAC_T>(&jacobian, r * x_stride + unit_dimension, @@ -351,7 +351,14 @@ Status ComputeGradientErrorInternal(const Scope& scope, const OutputList& xs, auto jac_n = jacobian_ns[i].matrix<JAC_T>(); for (int r = 0; r < jacobian_ts[i].dim_size(0); ++r) { for (int c = 0; c < jacobian_ts[i].dim_size(1); ++c) { - *max_error = std::max(*max_error, std::fabs(jac_t(r, c) - jac_n(r, c))); + auto cur_error = std::fabs(jac_t(r, c) - jac_n(r, c)); + // Treat any NaN as max_error and immediately return. + // (Note that std::max may ignore NaN arguments.) + if (std::isnan(cur_error)) { + *max_error = cur_error; + return Status::OK(); + } + *max_error = std::max(*max_error, cur_error); } } } @@ -409,6 +416,7 @@ Status ComputeGradientError(const Scope& scope, const Output& x, const Output& y, const TensorShape& y_shape, JAC_T* max_error); INSTANTIATE_GRAD_ERR_TYPE(float, float, float); +INSTANTIATE_GRAD_ERR_TYPE(double, float, double); INSTANTIATE_GRAD_ERR_TYPE(double, double, double); INSTANTIATE_GRAD_ERR_TYPE(complex64, float, float); INSTANTIATE_GRAD_ERR_TYPE(float, complex64, float); diff --git a/tensorflow/cc/framework/gradient_checker_test.cc b/tensorflow/cc/framework/gradient_checker_test.cc index d4f0a7f5ab..8dd762c282 100644 --- a/tensorflow/cc/framework/gradient_checker_test.cc +++ b/tensorflow/cc/framework/gradient_checker_test.cc @@ -28,12 +28,14 @@ namespace { using ops::Complex; using ops::Const; +using ops::Div; using ops::MatMul; using ops::Placeholder; using ops::Real; using ops::Split; using ops::Square; using ops::Stack; +using ops::Sub; using ops::Unstack; TEST(GradientCheckerTest, BasicFloat) { @@ -104,6 +106,20 @@ TEST(GradientCheckerTest, Complex64ToFloat) { EXPECT_LT(max_error, 1e-4); } +// When calculating gradients that are undefined, test we get NaN +// as the computed error rather than 0. +TEST(GradientCheckerTest, BasicNan) { + Scope scope = Scope::NewRootScope(); + TensorShape shape({2, 4, 3}); + auto x = Placeholder(scope, DT_FLOAT, Placeholder::Shape(shape)); + // y = x/(x-x) should always return NaN + auto y = Div(scope, x, Sub(scope, x, x)); + float max_error; + TF_ASSERT_OK((ComputeGradientError<float, float, float>( + scope, {x}, {shape}, {y}, {shape}, &max_error))); + EXPECT_TRUE(std::isnan(max_error)); +} + TEST(GradientCheckerTest, MatMulGrad) { Scope scope = Scope::NewRootScope(); diff --git a/tensorflow/cc/gradients/image_grad.cc b/tensorflow/cc/gradients/image_grad.cc new file mode 100644 index 0000000000..882709e1e2 --- /dev/null +++ b/tensorflow/cc/gradients/image_grad.cc @@ -0,0 +1,74 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include <vector> +#include "tensorflow/cc/framework/grad_op_registry.h" +#include "tensorflow/cc/framework/gradients.h" +#include "tensorflow/cc/ops/image_ops_internal.h" +#include "tensorflow/cc/ops/standard_ops.h" + +namespace tensorflow { +namespace ops { +namespace { + +Status ResizeNearestNeighborGradHelper(const Scope& scope, const Operation& op, + const std::vector<Output>& grad_inputs, + std::vector<Output>* grad_outputs) { + bool align_corners; + TF_RETURN_IF_ERROR( + GetNodeAttr(op.node()->attrs(), "align_corners", &align_corners)); + // The internal gradient implementation needs the shape of the input image. + // x_shape = shape(x)[1:3] + // = slice(shape(x), {1}, {3 - 1}) + auto x_shape = Slice(scope, Shape(scope, op.input(0)), {1}, {2}); + grad_outputs->push_back(internal::ResizeNearestNeighborGrad( + scope, grad_inputs[0], x_shape, + internal::ResizeNearestNeighborGrad::AlignCorners(align_corners))); + grad_outputs->push_back(NoGradient()); + return scope.status(); +} +REGISTER_GRADIENT_OP("ResizeNearestNeighbor", ResizeNearestNeighborGradHelper); + +Status ResizeBilinearGradHelper(const Scope& scope, const Operation& op, + const std::vector<Output>& grad_inputs, + std::vector<Output>* grad_outputs) { + bool align_corners; + TF_RETURN_IF_ERROR( + GetNodeAttr(op.node()->attrs(), "align_corners", &align_corners)); + grad_outputs->push_back(internal::ResizeBilinearGrad( + scope, grad_inputs[0], op.input(0), + internal::ResizeBilinearGrad::AlignCorners(align_corners))); + grad_outputs->push_back(NoGradient()); + return scope.status(); +} +REGISTER_GRADIENT_OP("ResizeBilinear", ResizeBilinearGradHelper); + +Status ResizeBicubicGradHelper(const Scope& scope, const Operation& op, + const std::vector<Output>& grad_inputs, + std::vector<Output>* grad_outputs) { + bool align_corners; + TF_RETURN_IF_ERROR( + GetNodeAttr(op.node()->attrs(), "align_corners", &align_corners)); + grad_outputs->push_back(internal::ResizeBicubicGrad( + scope, grad_inputs[0], op.input(0), + internal::ResizeBicubicGrad::AlignCorners(align_corners))); + grad_outputs->push_back(NoGradient()); + return scope.status(); +} +REGISTER_GRADIENT_OP("ResizeBicubic", ResizeBicubicGradHelper); + +} // anonymous namespace +} // namespace ops +} // namespace tensorflow diff --git a/tensorflow/cc/gradients/image_grad_test.cc b/tensorflow/cc/gradients/image_grad_test.cc new file mode 100644 index 0000000000..2e55c7561b --- /dev/null +++ b/tensorflow/cc/gradients/image_grad_test.cc @@ -0,0 +1,157 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/cc/client/client_session.h" +#include "tensorflow/cc/framework/grad_op_registry.h" +#include "tensorflow/cc/framework/gradient_checker.h" +#include "tensorflow/cc/framework/testutil.h" +#include "tensorflow/cc/gradients/grad_testutil.h" +#include "tensorflow/cc/ops/image_ops.h" +#include "tensorflow/cc/ops/standard_ops.h" +#include "tensorflow/core/framework/tensor_testutil.h" +#include "tensorflow/core/lib/core/status_test_util.h" + +namespace tensorflow { +namespace { + +using ops::Const; +using ops::ResizeBicubic; +using ops::ResizeBilinear; +using ops::ResizeNearestNeighbor; + +class ImageGradTest : public ::testing::Test { + protected: + ImageGradTest() : scope_(Scope::NewRootScope()) {} + + enum OpType { RESIZE_NEAREST, RESIZE_BILINEAR, RESIZE_BICUBIC }; + + template <typename T> + Tensor MakeData(const TensorShape& data_shape) { + DataType data_type = DataTypeToEnum<T>::v(); + Tensor data(data_type, data_shape); + auto data_flat = data.flat<T>(); + for (int i = 0; i < data_flat.size(); ++i) { + data_flat(i) = T(i); + } + return data; + } + + template <typename T> + void MakeOp(const OpType op_type, const Tensor& x_data, const Input& y_shape, + const bool align_corners, Output* x, Output* y) { + *x = Const<T>(scope_, x_data); + switch (op_type) { + case RESIZE_NEAREST: + *y = ResizeNearestNeighbor( + scope_, *x, y_shape, + ResizeNearestNeighbor::AlignCorners(align_corners)); + return; + case RESIZE_BILINEAR: + *y = ResizeBilinear(scope_, *x, y_shape, + ResizeBilinear::AlignCorners(align_corners)); + return; + case RESIZE_BICUBIC: + *y = ResizeBicubic(scope_, *x, y_shape, + ResizeBicubic::AlignCorners(align_corners)); + return; + } + assert(false); + } + + template <typename T> + void TestResizedShapeForType(const OpType op_type, const bool align_corners) { + TensorShape x_shape({1, 2, 2, 1}); + Tensor x_data = MakeData<T>(x_shape); + Output x, y; + MakeOp<T>(op_type, x_data, {4, 6}, align_corners, &x, &y); + + ClientSession session(scope_); + std::vector<Tensor> outputs; + TF_ASSERT_OK(session.Run({y}, &outputs)); + EXPECT_EQ(outputs.size(), 1); + EXPECT_EQ(outputs[0].shape(), TensorShape({1, 4, 6, 1})); + } + + void TestResizedShape(OpType op_type) { + for (const bool align_corners : {true, false}) { + TestResizedShapeForType<Eigen::half>(op_type, align_corners); + TestResizedShapeForType<float>(op_type, align_corners); + TestResizedShapeForType<double>(op_type, align_corners); + } + } + + template <typename X_T, typename Y_T, typename JAC_T> + void TestResizeToSmallerAndAlign(const OpType op_type, + const bool align_corners) { + TensorShape x_shape({1, 4, 6, 1}); + Tensor x_data = MakeData<X_T>(x_shape); + Output x, y; + MakeOp<X_T>(op_type, x_data, {2, 3}, align_corners, &x, &y); + JAC_T max_error; + TF_ASSERT_OK((ComputeGradientError<X_T, Y_T, JAC_T>( + scope_, x, x_data, y, {1, 2, 3, 1}, &max_error))); + EXPECT_LT(max_error, 1e-3); + } + + template <typename X_T, typename Y_T, typename JAC_T> + void TestResizeToLargerAndAlign(const OpType op_type, + const bool align_corners) { + TensorShape x_shape({1, 2, 3, 1}); + Tensor x_data = MakeData<X_T>(x_shape); + Output x, y; + MakeOp<X_T>(op_type, x_data, {4, 6}, align_corners, &x, &y); + JAC_T max_error; + TF_ASSERT_OK((ComputeGradientError<X_T, Y_T, JAC_T>( + scope_, x, x_data, y, {1, 4, 6, 1}, &max_error))); + EXPECT_LT(max_error, 1e-3); + } + + template <typename X_T, typename Y_T, typename JAC_T> + void TestResize(OpType op_type) { + for (const bool align_corners : {true, false}) { + TestResizeToSmallerAndAlign<X_T, Y_T, JAC_T>(op_type, align_corners); + TestResizeToLargerAndAlign<X_T, Y_T, JAC_T>(op_type, align_corners); + } + } + + Scope scope_; +}; + +TEST_F(ImageGradTest, TestNearestNeighbor) { + TestResizedShape(RESIZE_NEAREST); + TestResize<float, float, float>(RESIZE_NEAREST); + TestResize<double, double, double>(RESIZE_NEAREST); +} + +TEST_F(ImageGradTest, TestBilinear) { + TestResizedShape(RESIZE_BILINEAR); + TestResize<float, float, float>(RESIZE_BILINEAR); + // Note that Y_T is always float for this op. We choose + // double for the jacobian to capture the higher precision + // between X_T and Y_T. + TestResize<double, float, double>(RESIZE_BILINEAR); +} + +TEST_F(ImageGradTest, TestBicubic) { + TestResizedShape(RESIZE_BICUBIC); + TestResize<float, float, float>(RESIZE_BICUBIC); + // Note that Y_T is always float for this op. We choose + // double for the jacobian to capture the higher precision + // between X_T and Y_T. + TestResize<double, float, double>(RESIZE_BICUBIC); +} + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/cc/gradients/math_grad_test.cc b/tensorflow/cc/gradients/math_grad_test.cc index fd7b6fe662..1c9bdff5e1 100644 --- a/tensorflow/cc/gradients/math_grad_test.cc +++ b/tensorflow/cc/gradients/math_grad_test.cc @@ -475,11 +475,7 @@ TEST_F(CWiseUnaryGradTest, Tan_Complex) { auto x_fn = [this](const int i) { return CRV({{1, 0}, {0, 1}, {2, -1}, {1, 2}, {3, 4}}); }; - // TODO(kbsriram) - // Enable when tan kernel supports complex inputs - if (false) { - TestCWiseGrad<complex64, complex64>(TAN, x_fn); - } + TestCWiseGrad<complex64, complex64>(TAN, x_fn); } TEST_F(CWiseUnaryGradTest, Atan) { diff --git a/tensorflow/cc/gradients/nn_grad.cc b/tensorflow/cc/gradients/nn_grad.cc index c73482d5f4..588e96cb19 100644 --- a/tensorflow/cc/gradients/nn_grad.cc +++ b/tensorflow/cc/gradients/nn_grad.cc @@ -47,6 +47,72 @@ Status SoftmaxGrad(const Scope& scope, const Operation& op, } REGISTER_GRADIENT_OP("Softmax", SoftmaxGrad); +bool IsZero(const Scope& scope, const Output& grad) { + string op_type_name = grad.op().node()->type_string(); + if (op_type_name == "ZerosLike" || op_type_name == "Zeros") { + return true; + } + // The Operation we were provided is not named something obvious so + // we need to actually look at its contents. + // The original python code did this by calling a utility function called + // tensor_util.constant_value. + // There is no C++ equivalent to tensor_util.constant_value so we do nothing + // for the moment. + return false; +} + +// Multiply after broadcasting vec to match dimensions of mat. +// Args: +// vec: A 1-D tensor of dimension [D0] +// mat: A 2-D tensor of dimesnion [D0, D1] +// +// Returns: +// A tensor of dimension [D0, D1], the result fo vec * mat. +Output BroadcastMul(const Scope& scope, const Output& vec, const Output& mat) { + auto reshaped = ExpandDims(scope, vec, -1); + return Multiply(scope, reshaped, mat); +} + +Status SoftmaxCrossEntropyWithLogitsGrad(const Scope& scope, + const Operation& op, + const std::vector<Output>& grad_inputs, + std::vector<Output>* grad_outputs) { + // Softmax gradient with cross entropy logits function. + // We multiply the backprop for cost with the gradients - op.output[1]. + // There is no gradient for labels. + + // The outputs of the network are at input index 0. + auto logits = op.input(0); + // The "truth" labels are at index 1. + auto softmax_grad = op.output(1); + + // The loss is the output at index 0, and backprop is the output at index 1. + auto grad_loss = grad_inputs[0]; + auto grad_grad = grad_inputs[1]; + + auto grad = BroadcastMul(scope, grad_loss, softmax_grad); + if (!IsZero(scope, grad_grad)) { + std::vector<int> axis; + auto logits_softmax = Softmax(scope, logits); + + auto grad_grad_expand = ExpandDims(scope, grad_grad, 1); + auto logits_softmax_expand = ExpandDims(scope, logits_softmax, 2); + auto matmul_result = + BatchMatMul(scope, grad_grad_expand, logits_softmax_expand); + axis.push_back(1); + auto squeeze_result = Squeeze(scope, matmul_result, Squeeze::Axis(axis)); + auto subtraction_result = Subtract(scope, grad_grad, squeeze_result); + auto multiply_result = Multiply(scope, subtraction_result, logits_softmax); + grad = Add(scope, grad, multiply_result); + } + auto minus_log_softmax = Multiply(scope, LogSoftmax(scope, logits), -1.0f); + grad_outputs->push_back(grad); + grad_outputs->push_back(BroadcastMul(scope, grad_loss, minus_log_softmax)); + return scope.status(); +} +REGISTER_GRADIENT_OP("SoftmaxCrossEntropyWithLogits", + SoftmaxCrossEntropyWithLogitsGrad); + Status LogSoftmaxGrad(const Scope& scope, const Operation& op, const std::vector<Output>& grad_inputs, std::vector<Output>* grad_outputs) { @@ -195,9 +261,9 @@ Status MaxPool3DGradHelper(const Scope& scope, const Operation& op, TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "padding", &padding)); TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "data_format", &data_format)); MaxPool3DGrad::Attrs grad_attrs; - auto dx = MaxPool3DGrad(scope, op.input(0), op.output(0), grad_inputs[0], - ksize, strides, padding, - grad_attrs.DataFormat(data_format)); + auto dx = + MaxPool3DGrad(scope, op.input(0), op.output(0), grad_inputs[0], ksize, + strides, padding, grad_attrs.DataFormat(data_format)); grad_outputs->push_back(dx); return scope.status(); } @@ -216,10 +282,9 @@ Status AvgPoolGradHelper(const Scope& scope, const Operation& op, TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "padding", &padding)); TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "data_format", &data_format)); internal::AvgPoolGrad::Attrs grad_attrs; - auto dx = - internal::AvgPoolGrad(scope, Shape(scope, op.input(0)), grad_inputs[0], - ksize, strides, padding, - grad_attrs.DataFormat(data_format)); + auto dx = internal::AvgPoolGrad(scope, Shape(scope, op.input(0)), + grad_inputs[0], ksize, strides, padding, + grad_attrs.DataFormat(data_format)); grad_outputs->push_back(dx); return scope.status(); } @@ -238,9 +303,9 @@ Status AvgPool3DGradHelper(const Scope& scope, const Operation& op, TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "padding", &padding)); TF_RETURN_IF_ERROR(GetNodeAttr(attrs, "data_format", &data_format)); AvgPool3DGrad::Attrs grad_attrs; - auto dx = AvgPool3DGrad(scope, Shape(scope, op.input(0)), grad_inputs[0], - ksize, strides, padding, - grad_attrs.DataFormat(data_format)); + auto dx = + AvgPool3DGrad(scope, Shape(scope, op.input(0)), grad_inputs[0], ksize, + strides, padding, grad_attrs.DataFormat(data_format)); grad_outputs->push_back(dx); return scope.status(); } diff --git a/tensorflow/cc/gradients/nn_grad_test.cc b/tensorflow/cc/gradients/nn_grad_test.cc index b4d457a9d1..aa72cf7ba2 100644 --- a/tensorflow/cc/gradients/nn_grad_test.cc +++ b/tensorflow/cc/gradients/nn_grad_test.cc @@ -25,6 +25,8 @@ limitations under the License. namespace tensorflow { namespace { +using ops::AvgPool; +using ops::AvgPool3D; using ops::BiasAdd; using ops::Conv2D; using ops::Elu; @@ -33,11 +35,9 @@ using ops::FractionalMaxPool; using ops::L2Loss; using ops::LogSoftmax; using ops::LRN; -using ops::AvgPool; -using ops::AvgPool3D; using ops::MaxPool; -using ops::MaxPoolV2; using ops::MaxPool3D; +using ops::MaxPoolV2; using ops::Placeholder; using ops::Relu; using ops::Relu6; @@ -111,6 +111,20 @@ TEST_F(NNGradTest, SoftmaxGrad) { RunTest(x, shape, y, shape); } +TEST_F(NNGradTest, SoftmaxCrossEntropyWithLogitsGrad) { + TensorShape logits_shape({5, 3}); + TensorShape loss_shape({5}); + + auto logits = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(logits_shape)); + auto labels = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(logits_shape)); + auto y = + tensorflow::ops::SoftmaxCrossEntropyWithLogits(scope_, logits, labels); + // Note the reversal of the backprop and loss orders. Issue #18734 has been + // opened for this. + RunTest({logits, labels}, {logits_shape, logits_shape}, {y.backprop, y.loss}, + {logits_shape, loss_shape}); +} + TEST_F(NNGradTest, LogSoftmaxGrad) { TensorShape shape({5, 3}); auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape)); @@ -253,7 +267,7 @@ TEST_F(NNGradTest, AvgPool3DGradHelper) { RunTest(x, x_shape, y, y_shape); } -TEST_F(NNGradTest, LRN){ +TEST_F(NNGradTest, LRN) { TensorShape x_shape({1, 1, 2, 1}); auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)); auto y = LRN(scope_, x); diff --git a/tensorflow/cc/saved_model/BUILD b/tensorflow/cc/saved_model/BUILD index 730b1b669b..3d3895c8fa 100644 --- a/tensorflow/cc/saved_model/BUILD +++ b/tensorflow/cc/saved_model/BUILD @@ -39,9 +39,20 @@ cc_library( hdrs = ["reader.h"], deps = [ ":constants", + ] + if_not_mobile([ + # TODO(b/111634734): :lib and :protos_all contain dependencies that + # cannot be built on mobile platforms. Instead, include the appropriate + # tf_lib depending on the build platform. "//tensorflow/core:lib", "//tensorflow/core:protos_all_cc", - ], + ]) + if_mobile([ + # Mobile-friendly SavedModel proto. See go/portable-proto for more info. + "//tensorflow/core:saved_model_portable_proto", + ]) + if_android([ + "//tensorflow/core:android_tensorflow_lib", + ]) + if_ios([ + "//tensorflow/core:ios_tensorflow_lib", + ]), ) tf_cc_test( diff --git a/tensorflow/cc/saved_model/loader.cc b/tensorflow/cc/saved_model/loader.cc index 07807ed2f3..98be66a6ad 100644 --- a/tensorflow/cc/saved_model/loader.cc +++ b/tensorflow/cc/saved_model/loader.cc @@ -74,6 +74,54 @@ void AddAssetsTensorsToInputs(const StringPiece export_dir, } } +// Like Session::Run(), but uses the Make/Run/ReleaseCallable() API to avoid +// leaving behind non-GC'ed state. +// +// Detailed motivation behind this approach, from ashankar@: +// +// Each call to Session::Run() that identifies a new subgraph (based on feeds +// and fetches) creates some datastructures that live as long as the session +// (the partitioned graph, associated executors etc.). +// +// A pathological case of this would be if say the initialization op +// (main_op/legacy_init_op) involves the use of a large constant. Then we +// allocate memory for that large constant that will just stick around till the +// session dies. With this Callable mechanism, that memory will be released +// right after ReleaseCallable returns. +// +// However, the resource manager state remains. +Status RunOnce(const RunOptions& run_options, + const std::vector<std::pair<string, Tensor>>& inputs, + const std::vector<string>& output_tensor_names, + const std::vector<string>& target_node_names, + std::vector<Tensor>* outputs, RunMetadata* run_metadata, + Session* session) { + CallableOptions callable_options; + std::vector<Tensor> feed_tensors; + *callable_options.mutable_run_options() = run_options; + for (const auto& input : inputs) { + const string& name = input.first; + const Tensor& tensor = input.second; + callable_options.add_feed(name); + feed_tensors.push_back(tensor); + } + for (const string& output_tensor_name : output_tensor_names) { + callable_options.add_fetch(output_tensor_name); + } + for (const string& target_node_name : target_node_names) { + callable_options.add_target(target_node_name); + } + + Session::CallableHandle callable_handle; + TF_RETURN_IF_ERROR(session->MakeCallable(callable_options, &callable_handle)); + const Status run_status = session->RunCallable(callable_handle, feed_tensors, + outputs, run_metadata); + // Be sure to call ReleaseCallable() regardless of the outcome of + // RunCallable(). + session->ReleaseCallable(callable_handle).IgnoreError(); + return run_status; +} + bool HasMainOp(const MetaGraphDef& meta_graph_def) { const auto& collection_def_map = meta_graph_def.collection_def(); if (collection_def_map.find(kSavedModelMainOpKey) != @@ -86,10 +134,11 @@ bool HasMainOp(const MetaGraphDef& meta_graph_def) { Status RunMainOp(const RunOptions& run_options, const string& export_dir, const MetaGraphDef& meta_graph_def, const std::vector<AssetFileDef>& asset_file_defs, - Session* session) { - LOG(INFO) << "Running MainOp on SavedModel bundle."; + Session* session, const string& main_op_key) { + LOG(INFO) << "Running MainOp with key " << main_op_key + << " on SavedModel bundle."; const auto& collection_def_map = meta_graph_def.collection_def(); - const auto main_op_it = collection_def_map.find(kSavedModelMainOpKey); + const auto main_op_it = collection_def_map.find(main_op_key); if (main_op_it != collection_def_map.end()) { if (main_op_it->second.node_list().value_size() != 1) { return errors::FailedPrecondition( @@ -99,8 +148,8 @@ Status RunMainOp(const RunOptions& run_options, const string& export_dir, AddAssetsTensorsToInputs(export_dir, asset_file_defs, &inputs); RunMetadata run_metadata; const StringPiece main_op_name = main_op_it->second.node_list().value(0); - return session->Run(run_options, inputs, {}, {main_op_name.ToString()}, - nullptr /* outputs */, &run_metadata); + return RunOnce(run_options, inputs, {}, {main_op_name.ToString()}, + nullptr /* outputs */, &run_metadata, session); } return Status::OK(); } @@ -137,32 +186,8 @@ Status RunRestore(const RunOptions& run_options, const string& export_dir, AddAssetsTensorsToInputs(export_dir, asset_file_defs, &inputs); RunMetadata run_metadata; - return session->Run(run_options, inputs, {}, {restore_op_name.ToString()}, - nullptr /* outputs */, &run_metadata); -} - -Status RunLegacyInitOp(const RunOptions& run_options, const string& export_dir, - const MetaGraphDef& meta_graph_def, - const std::vector<AssetFileDef>& asset_file_defs, - Session* session) { - LOG(INFO) << "Running LegacyInitOp on SavedModel bundle."; - const auto& collection_def_map = meta_graph_def.collection_def(); - const auto init_op_it = collection_def_map.find(kSavedModelLegacyInitOpKey); - if (init_op_it != collection_def_map.end()) { - if (init_op_it->second.node_list().value_size() != 1) { - return errors::FailedPrecondition(strings::StrCat( - "Expected exactly one serving init op in : ", export_dir)); - } - std::vector<std::pair<string, Tensor>> inputs; - AddAssetsTensorsToInputs(export_dir, asset_file_defs, &inputs); - RunMetadata run_metadata; - const StringPiece legacy_init_op_name = - init_op_it->second.node_list().value(0); - return session->Run(run_options, inputs, {}, - {legacy_init_op_name.ToString()}, nullptr /* outputs */, - &run_metadata); - } - return Status::OK(); + return RunOnce(run_options, inputs, {}, {restore_op_name.ToString()}, + nullptr /* outputs */, &run_metadata, session); } Status GetAssetFileDefs(const MetaGraphDef& meta_graph_def, @@ -204,11 +229,11 @@ Status LoadSavedModelInternal(const SessionOptions& session_options, if (HasMainOp(bundle->meta_graph_def)) { TF_RETURN_IF_ERROR(RunMainOp(run_options, export_dir, bundle->meta_graph_def, asset_file_defs, - bundle->session.get())); + bundle->session.get(), kSavedModelMainOpKey)); } else { - TF_RETURN_IF_ERROR(RunLegacyInitOp(run_options, export_dir, - bundle->meta_graph_def, asset_file_defs, - bundle->session.get())); + TF_RETURN_IF_ERROR(RunMainOp( + run_options, export_dir, bundle->meta_graph_def, asset_file_defs, + bundle->session.get(), kSavedModelLegacyInitOpKey)); } return Status::OK(); } |