Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

… dygraph_second_pr_utils

paddle/fluid/eager/CMakeLists.txt

@@ -2,5 +2,4 @@ add_subdirectory(api)
 add_subdirectory(tests)
 cc_library(grad_node_info SRCS grad_node_info.cc DEPS pten pten_api)
 cc_library(autograd_meta SRCS autograd_meta.cc DEPS pten pten_api)
-
 cc_library(utils SRCS utils.cc DEPS pten pten_api global_utils layer proto_desc operator op_registry variable_helper memcpy scale_op autograd_meta)

paddle/fluid/eager/api/all.cc

@@ -0,0 +1,18 @@
+// Copyright (c) 2021 PaddlePaddle 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 "paddle/fluid/eager/api/all.h"
+
+namespace egr {}  // namespace egr

paddle/fluid/eager/api/all.h

@@ -0,0 +1,17 @@
+// Copyright (c) 2021 PaddlePaddle 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.
+//
+#pragma once
+
+#include "paddle/fluid/eager/api/utils/global_utils.h"

paddle/fluid/eager/eager_tensor.h

@@ -14,7 +14,6 @@
 
 #pragma once
 // framework deps
-#include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/framework/pten_utils.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/variable.h"

paddle/fluid/eager/tensor_wrapper.h

@@ -0,0 +1,91 @@
+// Copyright (c) 2021 PaddlePaddle 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.
+
+/**
+ * We now still need TensorWrapper and it is designed to Copy
+ * tensor in autograd mode.
+ *
+ * Since in autograd usage, we need to pass autograd_meta to
+ * backward computation however in tensor interface add to much
+ * autograd_related method is not a good choice.
+ *
+ * In TensorWrapper we will keep autograd info to backward, only
+ * for input var, but for output var it will only copy autograd
+ * with no grad **/
+
+#pragma once
+#include "paddle/fluid/eager/autograd_meta.h"
+#include "paddle/fluid/eager/grad_node_info.h"
+#include "paddle/fluid/eager/utils.h"
+
+namespace egr {
+class TensorWrapper {
+ public:
+  TensorWrapper() = default;
+  explicit TensorWrapper(const egr::EagerTensor& tensor,
+                         bool full_reserved = false) {
+    /**
+     * Normally, we should fully reserved all non-output or non-leaf fwd tensor
+     * here. And for fwd output tensor, we should not reserve its autogradmeta,
+     * to avoid recursive depends on GradNodeBase
+     * **/
+    full_reserved_ = full_reserved;
+    if (full_reserved_) {
+      VLOG(6) << "Fully reserved tensor: " << tensor.name();
+      intermidiate_tensor_ = tensor;
+      return;
+    }
+
+    // shallow copy tensor_impl here
+    intermidiate_tensor_.set_impl(tensor.impl());
+    intermidiate_tensor_.ResetVar(tensor.Var());
+    intermidiate_tensor_.set_name(tensor.name() + "@Saved");
+    PADDLE_ENFORCE_NOT_NULL(
+        EagerUtils::unsafe_autograd_meta(tensor),
+        paddle::platform::errors::Fatal(
+            "Full reserved Tensor should not have null autograd meta, since "
+            "tensor_wrapper is used to build backward info. There is no way "
+            "for us to build it with null autograd_meta."));
+    // copy output_rank
+    out_rank_info_ = EagerUtils::OutRankInfo(tensor);
+  }
+
+  egr::EagerTensor recover(const std::shared_ptr<GradNodeBase>& grad_node) {
+    VLOG(6) << "Recover tensor for wrapper";
+    if ((!intermidiate_tensor_.defined()) &&
+        (!intermidiate_tensor_.Var().IsInitialized())) {
+      VLOG(6) << "Return NULL tensor Here. ";
+      return egr::EagerTensor();
+    }
+
+    // if it's full_reserved just return the full copy of tensor
+    if (full_reserved_) {
+      return intermidiate_tensor_;
+    } else {
+      std::shared_ptr<GradNodeBase> new_grad_node = grad_node;
+      auto p_ab_autograd_meta =
+          std::make_shared<AutogradMeta>(Edge(new_grad_node, out_rank_info_));
+      intermidiate_tensor_.set_autograd_meta(
+          std::static_pointer_cast<paddle::experimental::AbstractAutogradMeta>(
+              p_ab_autograd_meta));
+      return intermidiate_tensor_;
+    }
+  }
+
+ private:
+  bool full_reserved_ = false;
+  std::pair<size_t, size_t> out_rank_info_;
+  egr::EagerTensor intermidiate_tensor_;
+};
+}  // namespace egr

paddle/fluid/eager/tests/CMakeLists.txt

@@ -1,3 +1,3 @@
-set(eager_deps pten pten_api pten_tensor utils)
+set(eager_deps pten pten_api pten_tensor utils global_utils autograd_meta grad_node_info)
 add_subdirectory(data_structure_tests)
 add_subdirectory(task_tests)

paddle/fluid/eager/tests/data_structure_tests/CMakeLists.txt

@@ -1,3 +1,4 @@
 cc_test(test_egr_ds_eager_tensor SRCS eager_tensor_test.cc DEPS ${eager_deps} )
 cc_test(test_egr_ds_auotgrad_meta SRCS autograd_meta_test.cc DEPS ${eager_deps} grad_node_info)
 cc_test(test_egr_ds_grad_node_info SRCS grad_node_info_test.cc DEPS ${eager_deps} grad_node_info)
+cc_test(test_egr_ds_tensor_wrapper SRCS tensor_wrapper_test.cc DEPS ${eager_deps} grad_node_info utils)

paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h

@@ -19,6 +19,9 @@
 #include "paddle/fluid/eager/eager_tensor.h"
 #include "paddle/fluid/eager/grad_node_info.h"
 #include "paddle/pten/api/lib/utils/allocator.h"
+namespace egr {
+class TensorWrapper;
+}
 
 namespace eager_test {
 class GradTestNode : public egr::GradNodeBase {

paddle/fluid/eager/tests/data_structure_tests/tensor_wrapper_test.cc

@@ -0,0 +1,80 @@
+// Copyright (c) 2021 PaddlePaddle 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 "glog/logging.h"
+#include "gtest/gtest.h"
+
+#include "paddle/fluid/eager/tensor_wrapper.h"
+#include "paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h"
+#include "paddle/fluid/eager/utils.h"
+
+TEST(TensorWrapper, Basic) {
+  VLOG(6) << "Test Full reserved";
+  egr::EagerTensor et1;
+  pten::DenseTensorMeta meta = pten::DenseTensorMeta(
+      pten::DataType::FLOAT32, paddle::framework::make_ddim({1, 2}));
+  std::shared_ptr<pten::DenseTensor> dt = std::make_shared<pten::DenseTensor>(
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          paddle::platform::CPUPlace()),
+      meta);
+  auto* dt_ptr = dt->mutable_data<float>();
+  dt_ptr[0] = 5.0f;
+  dt_ptr[1] = 10.0f;
+  et1.set_impl(dt);
+  // Create grad node;
+  auto grad_test_node0 = std::make_shared<eager_test::GradTestNode>(
+      /* val */ 5.0, /* in_num */ 2, /* out_num */ 2);
+  egr::Edge edge0(grad_test_node0, 1, 2);
+  auto auto_grad0 = std::make_shared<egr::AutogradMeta>(edge0);
+  et1.set_autograd_meta(auto_grad0);
+  et1.set_name("et1");
+  auto tw0 = egr::TensorWrapper(et1, true);
+  auto recover_et1 = tw0.recover(std::make_shared<eager_test::GradTestNode>());
+  CHECK_EQ(recover_et1.name(), std::string("et1"));
+  CHECK_EQ(egr::EagerUtils::OutRankInfo(recover_et1).first,
+           egr::EagerUtils::OutRankInfo(et1).first);
+  CHECK_EQ(egr::EagerUtils::OutRankInfo(recover_et1).second,
+           egr::EagerUtils::OutRankInfo(et1).second);
+  VLOG(6) << "Test reconstruct";
+  egr::EagerTensor et2;
+  pten::DenseTensorMeta meta2 = pten::DenseTensorMeta(
+      pten::DataType::FLOAT32, paddle::framework::make_ddim({1, 2}));
+  std::shared_ptr<pten::DenseTensor> dt2 = std::make_shared<pten::DenseTensor>(
+      std::make_shared<paddle::experimental::DefaultAllocator>(
+          paddle::platform::CPUPlace()),
+      meta2);
+  auto* dt_ptr2 = dt->mutable_data<float>();
+  dt_ptr2[0] = 6.0f;
+  dt_ptr2[1] = 11.0f;
+  et2.set_impl(dt2);
+  et2.set_name("et2");
+  auto grad_test_node1 =
+      std::make_shared<eager_test::GradTestNode>(/* val */ 5.0, 2, 2);
+  egr::Edge edge1(grad_test_node1, 1, 2);
+  auto auto_grad1 = std::make_shared<egr::AutogradMeta>(edge1);
+  et2.set_autograd_meta(auto_grad1);
+  auto tw1 = egr::TensorWrapper(et2, false);
+  auto recover_et2 = tw1.recover(grad_test_node1);
+  CHECK_EQ(recover_et2.name(), std::string("et2@Saved"));
+  CHECK_EQ(egr::EagerUtils::OutRankInfo(recover_et2).first,
+           egr::EagerUtils::OutRankInfo(et2).first);
+  CHECK_EQ(egr::EagerUtils::OutRankInfo(recover_et2).second,
+           egr::EagerUtils::OutRankInfo(et2).second);
+  // Test Raw recover
+  egr::EagerTensor et3;
+  auto tw2 = egr::TensorWrapper(et3, true);
+  CHECK(
+      tw2.recover(std::make_shared<eager_test::GradTestNode>()).initialized() ==
+      false);
+}