【PIR API adaptor No.1、20、103、104、120】 Migrate L1Loss/BCELoss/HSigmoidLoss/SmoothL1Loss/KLDivLoss into pir #58708

DrRyanHuang · 2023-11-06T04:46:17Z

PR types

Others

PR changes

APIs

Description

新IR Python API适配升级 #58067

paddle.nn.L1Loss
paddle.nn.BCELoss
paddle.nn.HSigmoidLoss
paddle.nn.SmoothL1Loss
paddle.nn.KLDivLoss

test/legacy_test/test_l1_loss.py 文件下
self.assertTrue('aaa' in result3.name) 不支持pir model，于是将 TestClassL1Loss.run_static 和 TestFunctionalL1Loss.run_static 关闭
test/legacy_test/test_hsigmoid_op.py 没有开启 test_hs_grad_with_sparse，原因是调用的 paddle.static.nn.embedding 和base.DataFeeder 应该没有适配pir

…add t_w_pir_api

test/legacy_test/test_l1_loss.py

test/legacy_test/test_hsigmoid_op.py

test/legacy_test/test_bce_loss.py

as

DrRyanHuang · 2023-11-24T15:53:32Z

目前需要等待 #58876 先合入

MarioLulab · 2023-12-19T03:14:45Z

nn.initializer.Bilinear/Assign 已迁移至 pir 下：#60114
此 pr 可以继续推进了~

0x45f · 2023-12-22T08:55:22Z

HsigmoidLossOp::Vjp遇到报错，报错栈为

2023-12-20 03:24:01 I1219 19:24:00.857872 24887 pass.cc:38] --- detected [0] subgraphs!
2023-12-20 03:24:01 test_hsigmoid_op failed
2023-12-20 03:24:01  .............E..
2023-12-20 03:24:01 ======================================================================
2023-12-20 03:24:01 ERROR: test_check_grad (test_hsigmoid_op.TestHSigmoidOpWithCostumTreeWithoutBias)
2023-12-20 03:24:01 ----------------------------------------------------------------------
2023-12-20 03:24:01 Traceback (most recent call last):
2023-12-20 03:24:01   File "/workspace/Paddle/build/test/legacy_test/test_hsigmoid_op.py", line 490, in test_check_grad
2023-12-20 03:24:01     self.check_grad(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/op_test.py", line 2969, in check_grad
2023-12-20 03:24:01     self.check_grad_with_place(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/op_test.py", line 3288, in check_grad_with_place
2023-12-20 03:24:01     pir_grad = self._get_ir_gradient(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/op_test.py", line 3814, in _get_ir_gradient
2023-12-20 03:24:01     grad_inputs = ir_grad(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 1077, in grad
2023-12-20 03:24:01     input_grad = calc_gradient(outputs, inputs, grad_outputs, no_grad_set)
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 962, in calc_gradient
2023-12-20 03:24:01     input_to_inputgrad_map = calc_gradient_helper(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 901, in calc_gradient_helper
2023-12-20 03:24:01     append_backward_ops(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 759, in append_backward_ops
2023-12-20 03:24:01     input_grads = paddle.framework.core.call_vjp(
2023-12-20 03:24:01 ValueError: 
2023-12-20 03:24:01 
2023-12-20 03:24:01 0   paddle::dialect::VjpInterface::Vjp(pir::Operation*, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 1   paddle::dialect::VjpInterface::Model<paddle::dialect::HsigmoidLossOp>::Vjp(pir::Operation*, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 2   paddle::dialect::HsigmoidLossOp::Vjp(pir::Operation*, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 3   paddle::primitive::hsigmoid_loss_vjp(paddle::Tensor const&, paddle::Tensor const&, paddle::Tensor const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::Tensor const&, paddle::Tensor const&, int, bool, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 4   std::tuple<paddle::Tensor, paddle::Tensor, paddle::Tensor> paddle::primitive::backend::hsigmoid_loss_grad<paddle::primitive::LazyTensor>(paddle::Tensor const&, paddle::Tensor const&, paddle::Tensor const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::Tensor const&, paddle::Tensor const&, int, bool)
2023-12-20 03:24:01 5   paddle::dialect::hsigmoid_loss_grad(pir::Value const&, pir::Value const&, pir::Value const&, paddle::optional<pir::Value> const&, paddle::optional<pir::Value> const&, paddle::optional<pir::Value> const&, pir::Value const&, pir::Value const&, int, bool)
2023-12-20 03:24:01 6   paddle::dialect::HsigmoidLossGradOp::Build(pir::Builder&, pir::OperationArgument&, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, int, bool)
2023-12-20 03:24:01 7   paddle::dialect::IrMetaTensor::share_meta(phi::MetaTensor const&)
2023-12-20 03:24:01 8   phi::enforce::EnforceNotMet::EnforceNotMet(common::ErrorSummary const&, char const*, int)
2023-12-20 03:24:01 9   phi::enforce::GetCurrentTraceBackString[abi:cxx11](bool)
2023-12-20 03:24:01 
2023-12-20 03:24:01 ----------------------
2023-12-20 03:24:01 Error Message Summary:
2023-12-20 03:24:01 ----------------------
2023-12-20 03:24:01 InvalidArgumentError: The current MetaTensor is not initialized.
2023-12-20 03:24:01   [Hint: Expected meta_tensor.initialized() == true, but received meta_tensor.initialized():0 != true:1.] (at ../paddle/fluid/pir/dialect/operator/ir/ir_meta_tensor.cc:22)

内部相关同学在解决这个问题了~

changeyoung98 · 2023-12-25T02:18:04Z

HsigmoidLossOp::Vjp遇到报错，报错栈为

2023-12-20 03:24:01 I1219 19:24:00.857872 24887 pass.cc:38] --- detected [0] subgraphs!
2023-12-20 03:24:01 test_hsigmoid_op failed
2023-12-20 03:24:01  .............E..
2023-12-20 03:24:01 ======================================================================
2023-12-20 03:24:01 ERROR: test_check_grad (test_hsigmoid_op.TestHSigmoidOpWithCostumTreeWithoutBias)
2023-12-20 03:24:01 ----------------------------------------------------------------------
2023-12-20 03:24:01 Traceback (most recent call last):
2023-12-20 03:24:01   File "/workspace/Paddle/build/test/legacy_test/test_hsigmoid_op.py", line 490, in test_check_grad
2023-12-20 03:24:01     self.check_grad(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/op_test.py", line 2969, in check_grad
2023-12-20 03:24:01     self.check_grad_with_place(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/op_test.py", line 3288, in check_grad_with_place
2023-12-20 03:24:01     pir_grad = self._get_ir_gradient(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/op_test.py", line 3814, in _get_ir_gradient
2023-12-20 03:24:01     grad_inputs = ir_grad(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 1077, in grad
2023-12-20 03:24:01     input_grad = calc_gradient(outputs, inputs, grad_outputs, no_grad_set)
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 962, in calc_gradient
2023-12-20 03:24:01     input_to_inputgrad_map = calc_gradient_helper(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 901, in calc_gradient_helper
2023-12-20 03:24:01     append_backward_ops(
2023-12-20 03:24:01   File "/workspace/Paddle/build/python/paddle/autograd/ir_backward.py", line 759, in append_backward_ops
2023-12-20 03:24:01     input_grads = paddle.framework.core.call_vjp(
2023-12-20 03:24:01 ValueError: 
2023-12-20 03:24:01 
2023-12-20 03:24:01 0   paddle::dialect::VjpInterface::Vjp(pir::Operation*, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 1   paddle::dialect::VjpInterface::Model<paddle::dialect::HsigmoidLossOp>::Vjp(pir::Operation*, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 2   paddle::dialect::HsigmoidLossOp::Vjp(pir::Operation*, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<pir::OpResult, std::allocator<pir::OpResult> >, std::allocator<std::vector<pir::OpResult, std::allocator<pir::OpResult> > > > const&, std::vector<std::vector<pir::Value, std::allocator<pir::Value> >, std::allocator<std::vector<pir::Value, std::allocator<pir::Value> > > > const&, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 3   paddle::primitive::hsigmoid_loss_vjp(paddle::Tensor const&, paddle::Tensor const&, paddle::Tensor const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::Tensor const&, paddle::Tensor const&, int, bool, std::vector<std::vector<bool, std::allocator<bool> >, std::allocator<std::vector<bool, std::allocator<bool> > > > const&)
2023-12-20 03:24:01 4   std::tuple<paddle::Tensor, paddle::Tensor, paddle::Tensor> paddle::primitive::backend::hsigmoid_loss_grad<paddle::primitive::LazyTensor>(paddle::Tensor const&, paddle::Tensor const&, paddle::Tensor const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::optional<paddle::Tensor> const&, paddle::Tensor const&, paddle::Tensor const&, int, bool)
2023-12-20 03:24:01 5   paddle::dialect::hsigmoid_loss_grad(pir::Value const&, pir::Value const&, pir::Value const&, paddle::optional<pir::Value> const&, paddle::optional<pir::Value> const&, paddle::optional<pir::Value> const&, pir::Value const&, pir::Value const&, int, bool)
2023-12-20 03:24:01 6   paddle::dialect::HsigmoidLossGradOp::Build(pir::Builder&, pir::OperationArgument&, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, pir::Value, int, bool)
2023-12-20 03:24:01 7   paddle::dialect::IrMetaTensor::share_meta(phi::MetaTensor const&)
2023-12-20 03:24:01 8   phi::enforce::EnforceNotMet::EnforceNotMet(common::ErrorSummary const&, char const*, int)
2023-12-20 03:24:01 9   phi::enforce::GetCurrentTraceBackString[abi:cxx11](bool)
2023-12-20 03:24:01 
2023-12-20 03:24:01 ----------------------
2023-12-20 03:24:01 Error Message Summary:
2023-12-20 03:24:01 ----------------------
2023-12-20 03:24:01 InvalidArgumentError: The current MetaTensor is not initialized.
2023-12-20 03:24:01   [Hint: Expected meta_tensor.initialized() == true, but received meta_tensor.initialized():0 != true:1.] (at ../paddle/fluid/pir/dialect/operator/ir/ir_meta_tensor.cc:22)

此问题已解决：#60264 ，可拉取最新代码进行测试。

test/legacy_test/test_l1_loss.py

MarioLulab

LGTM

0x45f

LGTM，关于函数名的修改可以另提一个PR进行修改~

test/legacy_test/test_smooth_l1_loss.py

…Loss/SmoothL1Loss/KLDivLoss into pir (PaddlePaddle#58708)

add 5 loss

c7249f5

luotao1 mentioned this pull request Nov 6, 2023

新IR Python API适配升级 #58067

Closed

DrRyanHuang added 4 commits November 6, 2023 07:49

sm_l1_loss static program && rm @test_with_pir_api of test_errors && …

e592ddb

…add t_w_pir_api

Program -> paddle.static.Program

6693972

add program_guard to run_static

f4fe3ee

rm .name in_pir_mode

5ddbbf3

0x45f reviewed Nov 17, 2023

View reviewed changes

test/legacy_test/test_l1_loss.py Outdated Show resolved Hide resolved

test/legacy_test/test_hsigmoid_op.py Outdated Show resolved Hide resolved

test/legacy_test/test_hsigmoid_op.py Outdated Show resolved Hide resolved

test/legacy_test/test_bce_loss.py Outdated Show resolved Hide resolved

DrRyanHuang added 2 commits November 17, 2023 07:44

fix some questions

8dca4cd

NumpyArrayInitializer in pir mode

be51458

DrRyanHuang mentioned this pull request Nov 21, 2023

[WeeklyReports] 2023.11.08~2023.11.21 周报汇总 PFCCLab/Camp#77

Closed

21 tasks

Merge pull request #22 from PaddlePaddle/develop

f42beed

as

Merge branch 'develop' into loss

6d6e502

DrRyanHuang mentioned this pull request Dec 6, 2023

[WeeklyReports] 2023.11.22~2023.12.05 周报汇总 PFCCLab/Camp#102

Closed

20 tasks

DrRyanHuang and others added 5 commits December 19, 2023 14:53

Merge branch 'develop' into loss

86b79bd

Merge branch 'develop' into loss

ce56553

revert assign

c400c22

revert assign

0b1b9f3

fix MetaTensor is not initialized

aae45ce

This comment was marked as resolved.

Sign in to view

Merge branch 'develop' into loss

9906a26

This comment was marked as resolved.

Sign in to view

DrRyanHuang added 3 commits December 25, 2023 07:26

fix codestyle and can't find kernel fp16 using cpu

c62d385

add @

e393ada

Merge branch 'develop' into loss

f54f5a5

MarioLulab reviewed Dec 26, 2023

View reviewed changes

test/legacy_test/test_l1_loss.py Outdated Show resolved Hide resolved

test/legacy_test/test_l1_loss.py Outdated Show resolved Hide resolved

add in_pir_mode branch

c0acad3

MarioLulab approved these changes Dec 26, 2023

View reviewed changes

0x45f approved these changes Dec 26, 2023

View reviewed changes

test/legacy_test/test_smooth_l1_loss.py Show resolved Hide resolved

test/legacy_test/test_smooth_l1_loss.py Show resolved Hide resolved

test/legacy_test/test_smooth_l1_loss.py Show resolved Hide resolved

test/legacy_test/test_smooth_l1_loss.py Show resolved Hide resolved

0x45f merged commit a3a3466 into PaddlePaddle:develop Dec 26, 2023
29 checks passed

DrRyanHuang deleted the loss branch December 26, 2023 07:46

Wanglongzhi2001 pushed a commit to Wanglongzhi2001/Paddle that referenced this pull request Jan 7, 2024

【PIR API adaptor No.1、20、103、104、120】 Migrate L1Loss/BCELoss/HSigmoid…

fd8e00c

…Loss/SmoothL1Loss/KLDivLoss into pir (PaddlePaddle#58708)

DrRyanHuang mentioned this pull request Jan 24, 2024

[pir] fix frexp datatype #61087

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

【PIR API adaptor No.1、20、103、104、120】 Migrate L1Loss/BCELoss/HSigmoidLoss/SmoothL1Loss/KLDivLoss into pir #58708

【PIR API adaptor No.1、20、103、104、120】 Migrate L1Loss/BCELoss/HSigmoidLoss/SmoothL1Loss/KLDivLoss into pir #58708

DrRyanHuang commented Nov 6, 2023 •

edited

DrRyanHuang commented Nov 24, 2023

MarioLulab commented Dec 19, 2023 •

edited

This comment was marked as resolved.

0x45f commented Dec 22, 2023

changeyoung98 commented Dec 25, 2023 •

edited

This comment was marked as resolved.

MarioLulab left a comment

0x45f left a comment

【PIR API adaptor No.1、20、103、104、120】 Migrate L1Loss/BCELoss/HSigmoidLoss/SmoothL1Loss/KLDivLoss into pir #58708

【PIR API adaptor No.1、20、103、104、120】 Migrate L1Loss/BCELoss/HSigmoidLoss/SmoothL1Loss/KLDivLoss into pir #58708

Conversation

DrRyanHuang commented Nov 6, 2023 • edited

PR types

PR changes

Description

DrRyanHuang commented Nov 24, 2023

MarioLulab commented Dec 19, 2023 • edited

This comment was marked as resolved.

0x45f commented Dec 22, 2023

changeyoung98 commented Dec 25, 2023 • edited

This comment was marked as resolved.

MarioLulab left a comment

Choose a reason for hiding this comment

0x45f left a comment

Choose a reason for hiding this comment

DrRyanHuang commented Nov 6, 2023 •

edited

MarioLulab commented Dec 19, 2023 •

edited

changeyoung98 commented Dec 25, 2023 •

edited