[Prim][PIR] dropout forward sink #59176

kevincheng2 · 2023-11-20T12:22:11Z

PR types

New features

PR changes

Others

Description

prim dropout sink c++ and add uniform in primitive.h

Comparison of Python and C++ Program：
prim in Python

{
 (%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"x",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[8,16,32,64],stop_gradient:[false]} : () -> pd_op.tensor<8x16x32x64xf32>
 (%1) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)8,(Int64)16,(Int64)32,(Int64)64]} : () -> pd_op.tensor<4xi64>
 (%2) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)0} : () -> pd_op.tensor<1xf32>
 (%3) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)1} : () -> pd_op.tensor<1xf32>
 (%4) = "pd_op.uniform" (%1, %2, %3) {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),seed:(Int32)0,stop_gradient:[true]} : (pd_op.tensor<4xi64>, pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%5) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)0.5} : () -> pd_op.tensor<1xf32>
 (%6) = "pd_op.greater_equal" (%4, %5) {stop_gradient:[true]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<8x16x32x64xb>
 (%7) = "pd_op.cast" (%6) {dtype:(pd_op.DataType)float32,stop_gradient:[true]} : (pd_op.tensor<8x16x32x64xb>) -> pd_op.tensor<8x16x32x64xf32>
 (%8) = "pd_op.multiply" (%0, %7) {stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%9) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[8,16,32,64],stop_gradient:[true],value:(Float)0.5} : () -> pd_op.tensor<8x16x32x64xf32>
 (%10) = "pd_op.divide" (%8, %9) {stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%11) = "pd_op.cast" (%7) {dtype:(pd_op.DataType)uint8,stop_gradient:[true]} : (pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xu8>
 (%12) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)1} : () -> pd_op.tensor<1xf32>
 (%13) = "pd_op.full_like" (%10, %12) {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%14, %15) = "pd_op.divide_grad" (%8, %9, %10, %13) {axis:(Int32)-1,stop_gradient:[false,false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>, <<NULL TYPE>>
 (%16, %17) = "pd_op.multiply_grad" (%0, %7, %14) {axis:(Int32)-1,stop_gradient:[false,false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>, <<NULL TYPE>>
 (%18) = "pd_op.fetch" (%10) {col:(Int32)0,name:"fetch0",stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%19) = "pd_op.fetch" (%16) {col:(Int32)1,name:"fetch1",stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
}

prim in C++

{
 (%0) = "pd_op.data" () {dtype:(pd_op.DataType)float32,name:"x",place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[8,16,32,64],stop_gradient:[false]} : () -> pd_op.tensor<8x16x32x64xf32>
 (%1) = "pd_op.full_int_array" () {dtype:(pd_op.DataType)int64,place:(pd_op.Place)Place(cpu),stop_gradient:[true],value:[(Int64)8,(Int64)16,(Int64)32,(Int64)64]} : () -> pd_op.tensor<4xi64>
 (%2) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)0} : () -> pd_op.tensor<1xf32>
 (%3) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)1} : () -> pd_op.tensor<1xf32>
 (%4) = "pd_op.uniform" (%1, %2, %3) {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),seed:(Int32)0,stop_gradient:[true]} : (pd_op.tensor<4xi64>, pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%5) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[8,16,32,64],stop_gradient:[true],value:(Float)0.5} : () -> pd_op.tensor<8x16x32x64xf32>
 (%6) = "pd_op.greater_equal" (%4, %5) {stop_gradient:[true]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xb>
 (%7) = "pd_op.cast" (%6) {dtype:(pd_op.DataType)float32,stop_gradient:[true]} : (pd_op.tensor<8x16x32x64xb>) -> pd_op.tensor<8x16x32x64xf32>
 (%8) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[8,16,32,64],stop_gradient:[true],value:(Float)0.5} : () -> pd_op.tensor<8x16x32x64xf32>
 (%9) = "pd_op.multiply" (%0, %7) {stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%10) = "pd_op.divide" (%9, %8) {stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%11) = "pd_op.cast" (%7) {dtype:(pd_op.DataType)uint8,stop_gradient:[true]} : (pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xu8>
 (%12) = "pd_op.full" () {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(cpu),shape:(pd_op.IntArray)[1],stop_gradient:[true],value:(Float)1} : () -> pd_op.tensor<1xf32>
 (%13) = "pd_op.full_like" (%10, %12) {dtype:(pd_op.DataType)float32,place:(pd_op.Place)Place(undefined:0),stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%14, %15) = "pd_op.divide_grad" (%9, %8, %10, %13) {axis:(Int32)-1,stop_gradient:[false,false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>, <<NULL TYPE>>
 (%16, %17) = "pd_op.multiply_grad" (%0, %7, %14) {axis:(Int32)-1,stop_gradient:[false,false]} : (pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>, pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>, <<NULL TYPE>>
 (%18) = "pd_op.fetch" (%10) {col:(Int32)0,name:"fetch0",stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
 (%19) = "pd_op.fetch" (%16) {col:(Int32)1,name:"fetch1",stop_gradient:[false]} : (pd_op.tensor<8x16x32x64xf32>) -> pd_op.tensor<8x16x32x64xf32>
}

paddle-bot · 2023-11-20T12:22:20Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

cyber-pioneer · 2023-11-29T06:25:02Z

paddle/fluid/primitive/composite/composite.h

+    bool fix_seed) {
+  auto org_dtype = x.dtype();
+  bool upscale_in_train = false;
+  if (mode.compare("upscale_in_train") == 0) {


直接用 == 判断吧，这里是区分了大小写的，api侧已经check过了

cyber-pioneer · 2023-11-29T06:26:29Z

paddle/fluid/primitive/composite/composite.h

+        auto zero = full<T>(phi::vectorize(x.dims()), 0.0, org_dtype);
+        return std::make_tuple(x * zero, cast<T>(zero, phi::DataType::UINT8));
+      } else {
+        auto ans = divide<T>(x * mask, ones_p);


divide -> /

kevincheng2 added 7 commits November 16, 2023 08:53

dropout op sink

9bdaedd

update code

6e388bd

dropout sink

b9c5f0e

Merge branch 'develop' into dropout_prim_pir

9ae7cc6

test dropout op

67c279b

prim dropout sink

682afa7

pirm dropout sink

6d8f015

paddle-bot bot added the contributor External developers label Nov 20, 2023

kevincheng2 added 2 commits November 20, 2023 12:23

remove dropout in python

caefe00

merge code

18e149f

kevincheng2 mentioned this pull request Nov 24, 2023

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

Closed

21 tasks

kevincheng2 added 4 commits November 27, 2023 06:55

merge code

78028cc

Merge branch 'develop' into dropout_prim_pir

77424b5

merge code

1c63254

merge code

930608f

cyber-pioneer reviewed Nov 29, 2023

View reviewed changes

cyber-pioneer approved these changes Nov 29, 2023

View reviewed changes

cyber-pioneer merged commit ee63c32 into PaddlePaddle:develop Nov 29, 2023
30 checks passed

kevincheng2 deleted the dropout_prim_pir branch January 5, 2024 07:42

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Prim][PIR] dropout forward sink #59176

[Prim][PIR] dropout forward sink #59176

kevincheng2 commented Nov 20, 2023 •

edited

Loading

paddle-bot bot commented Nov 20, 2023

cyber-pioneer Nov 29, 2023

kevincheng2 Nov 29, 2023

cyber-pioneer Nov 29, 2023

kevincheng2 Nov 29, 2023

[Prim][PIR] dropout forward sink #59176

[Prim][PIR] dropout forward sink #59176

Conversation

kevincheng2 commented Nov 20, 2023 • edited Loading

PR types

PR changes

Description

paddle-bot bot commented Nov 20, 2023

cyber-pioneer Nov 29, 2023

Choose a reason for hiding this comment

kevincheng2 Nov 29, 2023

Choose a reason for hiding this comment

cyber-pioneer Nov 29, 2023

Choose a reason for hiding this comment

kevincheng2 Nov 29, 2023

Choose a reason for hiding this comment

kevincheng2 commented Nov 20, 2023 •

edited

Loading