[Eager Grad] Support eager grad interface #40170

veyron95 · 2022-03-04T09:37:33Z

PR types

New features

PR changes

APIs

Describe

Refator RunBackward(...) interface to adapt both backward(...) and grad(...) mode
Add cpp unit test file grad_test.cc to make sure the basic function of grad(...) interface in cpp-end is work
Expose run_partial_grad(...) interface in eager_functions.cc for python user to use
Support fluid.dygraph.grad(...) interface for Eager mode ( by calling run_partial_grad(...) interface )
Supplement python unit test cases to make sure the execution path from python user interface to cpp-end is work
Add ClearTensorWrappers and IsClearTensorWrappers virtual func in GradNodeBase
Support no_grad_vars, retrain_graph arguments

Note: Have not support double grad yet, create_graph under developing

… support_partial_grad

JiabinYang

some comments and make sure your deps build logic got approve from jim

paddle/fluid/eager/accumulation/accumulation_node.cc

paddle/fluid/eager/auto_code_generator/eager_generator.cc

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

paddle/fluid/eager/backward.cc

paddle/fluid/eager/auto_code_generator/eager_generator.cc

jim19930609 · 2022-03-14T01:39:54Z

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -2176,10 +2193,11 @@ static std::string GenerateGradNodeHeaderContents(
            TENSOR_WRAPPER_MEMBER_TEMPLATE, struct_tensor_wrapper_name);

        const char* SET_TENSOR_WRAPPER_BODY_TEMPLATE =
-            "%s = egr::TensorWrapper(%s, %s /*full_reserved*/);";
+            "%s = egr::TensorWrapper(%s, %s /*full_reserved*/);\n"
+            "     TensorWrappersSet.emplace_back(%s);";


emplace_back() will invoke "copy constructor" in case of lvalue, or "move constructor" in case of rvalue. So we usually use "emplace_back" together with "std::move", otherwise it's not gonna move (like in this case).

Let's say you used "emplace_back(std::move(%s))" here, then you're in big trouble. The problem is variable "x" after "std::move(x)" will become a dangling name, which doesn't actually own any underlying memory. Thus if anyone else uses "x" after your std::move, For instance "SetAttribute(x)", it's gonna cause trouble.

The idea could work with "std::vector<TensorWrapper*>". However, I strongly recommend the other way around (see comments above).

Thank you very much, this comments help me a lot, i do not use vector to hold all the tensor_wrapper now.

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

paddle/fluid/eager/backward.cc

jim19930609 · 2022-03-14T02:07:53Z

paddle/fluid/eager/backward.cc

+        // all the next_nodes of current node will be inserted to
+        // potential_stop_node
+        if (is_potential_stop_nodes) {
+          potential_stop_nodes->emplace(next_node);


Unrecommended use of "emplace", refer to the other comment on "emplace_back and std::move"

Thx, but std::unordered_set<GradNodeBase*>* potential_stop_nodes could not call emplace_back.

paddle/fluid/eager/backward.cc

paddle/fluid/eager/grad_node_info.h

… support_partial_grad

chenwhql

LGTM for PADDLE_ENFORCE

XiaoguangHu01

LGTM

This reverts commit 4db8cf2.

* [Eager] Support eager grad interface, draft version * Support eager grad interface with allow_unused and multi startup_op * Fix code format * Fix allow_unused case, return PyNone if tensor not initialize * Support output's stop_gradient related to create_graph * Support grad exception case in eager mode, fix coverage CI * Update ToPyObject, return PyNone if not initialize * AccumulationNode add FLAGS_retain_grad_for_all_tensor * Fix ci issue * Fix CI issue * fix, use core.eager.Tensor * Add func SetBufferSlotRankZeros for GradTensorHolder * Support retain_graph by using ClearTensorWrappers * Support retain_graph by using ClearTensorWrappers * Update retain_graph and no_grad_vars related test case * Update code gen logic for ClearTensorWrappers * Fix by override statement * fix override func args * Support retain_graph, update unit tests * Updated ClearTensorWrappers logic * fix grad python interface * Use deep copy and update unit tests * Polish code * Polish code * Fix CI issue, Deep copy only use when user set grad_tensors * Fix CI, use Backward instead RunBackward * Fix CI, Declare kernel explicitly in test file * Polish, remove vector of TensorWrapper * Refactor the logic of grad/backward, polish codes * Update code after merge upstream develop * Polish after merge upstream develop * Update to adapt new GradNodeBase superclass * Fix error introduced during conflict resolution * Update purify potential_startup_nodes logic * Fix errors * Polish code * Remove useless args for ToPyObject * Remove useless TensorWrappersSet * Fix code-format, re-install pre-commit * Fix pre-process logic for potential_startup_ops * Update unit tests, use eager mode

veyron95 changed the title ~~[Eager Grad] Support eager grad interface, draft version~~ [WIP, Eager Grad] Support eager grad interface Mar 7, 2022

veyron95 added 2 commits March 7, 2022 12:42

[Eager] Support eager grad interface, draft version

9fc70fe

Support eager grad interface with allow_unused and multi startup_op

ba8d79e

veyron95 force-pushed the support_partial_grad branch from 0d1beae to ba8d79e Compare March 7, 2022 12:45

veyron95 added 23 commits March 8, 2022 02:23

Fix code format

137db9d

Fix allow_unused case, return PyNone if tensor not initialize

1a18aa2

Support output's stop_gradient related to create_graph

d09ec3b

Support grad exception case in eager mode, fix coverage CI

f84f2be

Update ToPyObject, return PyNone if not initialize

733672e

AccumulationNode add FLAGS_retain_grad_for_all_tensor

68b1991

Fix ci issue

7665d63

Fix CI issue

86393f5

fix, use core.eager.Tensor

c653ec0

Add func SetBufferSlotRankZeros for GradTensorHolder

9156cea

Support retain_graph by using ClearTensorWrappers

6fd613d

Support retain_graph by using ClearTensorWrappers

58731e9

Update retain_graph and no_grad_vars related test case

a88f9b1

Update code gen logic for ClearTensorWrappers

778719b

Fix by override statement

65cf9e3

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

af7b919

… support_partial_grad

fix override func args

4d3b57d

Support retain_graph, update unit tests

415ff65

Updated ClearTensorWrappers logic

bb283ce

fix grad python interface

e548c22

Use deep copy and update unit tests

519c9a6

Polish code

1fbc61b

Polish code

c0a2b8b

veyron95 changed the title ~~[WIP, Eager Grad] Support eager grad interface~~ [Eager Grad] Support eager grad interface Mar 11, 2022

Fix CI, Declare kernel explicitly in test file

af7f058

JiabinYang reviewed Mar 14, 2022

View reviewed changes

jim19930609 reviewed Mar 14, 2022

View reviewed changes

paddle/fluid/eager/grad_node_info.h Outdated Show resolved Hide resolved

veyron95 added 12 commits March 14, 2022 13:14

Polish, remove vector of TensorWrapper

f397b8f

Refactor the logic of grad/backward, polish codes

e3f9826

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

27830a9

… support_partial_grad

Update code after merge upstream develop

7ede919

Polish after merge upstream develop

f9adf49

Update to adapt new GradNodeBase superclass

2fe3b9f

Fix error introduced during conflict resolution

90e97d6

Update purify potential_startup_nodes logic

d18697a

Fix errors

1b5eac2

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

f4e42e2

… support_partial_grad

Polish code

58a03b5

Remove useless args for ToPyObject

b04e9a9

jim19930609 previously approved these changes Mar 15, 2022

View reviewed changes

Remove useless TensorWrappersSet

c7bd6fc

veyron95 dismissed jim19930609’s stale review via c7bd6fc March 15, 2022 07:19

veyron95 added 4 commits March 16, 2022 06:34

Fix code-format, re-install pre-commit

ac85d81

Fix pre-process logic for potential_startup_ops

8312d2d

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

441bc81

… support_partial_grad

Update unit tests, use eager mode

326eee5

jim19930609 approved these changes Mar 17, 2022

View reviewed changes

chenwhql approved these changes Mar 17, 2022

View reviewed changes

XiaoguangHu01 approved these changes Mar 17, 2022

View reviewed changes

veyron95 merged commit 4db8cf2 into PaddlePaddle:develop Mar 17, 2022

veyron95 added a commit that referenced this pull request Mar 17, 2022

Revert "[Eager Grad] Support eager grad interface (#40170)"

4b269ba

This reverts commit 4db8cf2.

veyron95 mentioned this pull request Mar 17, 2022

Revert "[Eager Grad] Support eager grad interface" #40653

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Eager Grad] Support eager grad interface #40170

[Eager Grad] Support eager grad interface #40170

veyron95 commented Mar 4, 2022 •

edited

Loading

JiabinYang left a comment

jim19930609 Mar 14, 2022

veyron95 Mar 14, 2022

jim19930609 Mar 14, 2022

veyron95 Mar 14, 2022

chenwhql left a comment

XiaoguangHu01 left a comment

[Eager Grad] Support eager grad interface #40170

[Eager Grad] Support eager grad interface #40170

Conversation

veyron95 commented Mar 4, 2022 • edited Loading

PR types

PR changes

Describe

JiabinYang left a comment

Choose a reason for hiding this comment

jim19930609 Mar 14, 2022

Choose a reason for hiding this comment

veyron95 Mar 14, 2022

Choose a reason for hiding this comment

jim19930609 Mar 14, 2022

Choose a reason for hiding this comment

veyron95 Mar 14, 2022

Choose a reason for hiding this comment

chenwhql left a comment

Choose a reason for hiding this comment

XiaoguangHu01 left a comment

Choose a reason for hiding this comment

veyron95 commented Mar 4, 2022 •

edited

Loading