Enrich ConvShift to support sequence data input #2133

pkuyym · 2017-05-15T03:30:05Z

luotao1

建议将修改逻辑放进matrix函数里面，现在修改的和原有matrix里的重复太多。代码可读性不好。

luotao1 · 2017-05-15T04:53:13Z

paddle/gserver/layers/ConvShiftLayer.cpp

+    outV->circularConv(*inV0, *inV1);
+  } else {
+    circularConvSeq();
+  }


将circularConvSeq函数挪到matrix.cpp里面，这个函数的接口应该类似circularConv。即
outV->circularConv(*inV0, *inV1, *seqstartposition);

这样200-207行，可以修改为

if (inLayer0.sequenceStartPositions !=nullptr) { } else { }

去掉isSeqType这个函数。

luotao1

添加单测

luotao1 · 2017-05-15T04:53:41Z

paddle/gserver/layers/ConvShiftLayer.cpp

    outG->circularConvDerivative(*outG, *inV0, *inV1, *inG0, *inG1);
  } else {
-    CHECK(!inG0 || !inG1) << "Not supported";
+    circularConvSeqDerivative();


这里修改同forward函数

luotao1 · 2017-05-15T04:55:16Z

paddle/gserver/layers/ConvShiftLayer.cpp

+  real* outV = out->getData();
+
+  int leftCtxLen = (width1 - 1) / 2;
+  for (size_t x = 0; x < numSeqs - 1; x++) {


for循环最后的x++统一改成++x，下同

luotao1 · 2017-05-15T04:56:26Z

paddle/gserver/layers/ConvShiftLayer.cpp

+        int index = i + j - leftCtxLen;
+        index = (index + curSeqWidth) % curSeqWidth;
+        int outVRowOffset = i / width0;
+        int outVColOffset = i % width0;


113和114行只和i有关，应该放在109行的for循环里面，不要放在110行的for循环里面，这样会多算很多次。

luotao1 · 2017-05-15T04:58:17Z

paddle/gserver/layers/ConvShiftLayer.cpp

+    for (size_t i = 0; i < curSeqWidth; i++) {
+      for (size_t j = 0; j < width1; ++j) {
+        int index = i + j - leftCtxLen;
+        index = (index + curSeqWidth) % curSeqWidth;


111行和112行可以写在一块。

luotao1 · 2017-05-15T04:59:49Z

paddle/gserver/layers/ConvShiftLayer.cpp

+  CHECK_EQ(height1, inG1->getHeight());
+  CHECK_EQ(width1, inG1->getWidth());
+  CHECK_EQ(height0, outG->getHeight());
+  CHECK_EQ(width0, outG->getWidth());


cheack_eq有点多，可以只保留143,148,149三行

这个check是必要的，梯度、输入等的维度必须严格一致

luotao1 · 2017-05-15T05:01:41Z

paddle/gserver/layers/ConvShiftLayer.cpp

+  real* inV0 = in0->getData();
+  real* inV1 = in1->getData();
+  real* inGV0 = inG0->getData();
+  real* inGV1 = inG1->getData();


命名的时候，GV都可以改成G，下同。

延伸了Matrix函数的命名方式，GV代表真实的矩阵，感觉比较合适

luotao1 · 2017-05-15T05:01:56Z

paddle/gserver/layers/ConvShiftLayer.cpp

+        int inGV0RowOffset = index / width0;
+        int inGV0ColOffset = index % width0;
+        int outGVRowOffset = i / width0;
+        int outGVColOffset = i % width0;


修改意见同forward

lcy-seso

请先保证单测通过。

lcy-seso · 2017-05-15T07:37:37Z

paddle/gserver/layers/ConvShiftLayer.cpp

@@ -69,122 +66,12 @@ bool ConvShiftLayer::init(const LayerMap& layerMap,
  return true;
 }



22 ~ 42行的注释也需要对应地修改。

已经移到Matrix中

lcy-seso · 2017-05-15T07:39:55Z

paddle/gserver/layers/ConvShiftLayer.cpp

@@ -69,122 +66,12 @@ bool ConvShiftLayer::init(const LayerMap& layerMap,
  return true;
 }

-bool ConvShiftLayer::isSeqType() {


请先加单测保证 test_layer_grad 一定可以通过。计算程序，硬看很难保证计算结果的正确。

pkuyym · 2017-05-15T08:23:53Z

@luotao1 谢谢review，卷积逻辑已merge到matrix，现在ConvShift层应该比较清爽，对sequence和non-sequence的卷积逻辑已经并在一起

…2132

luotao1 · 2017-05-23T08:23:32Z

paddle/gserver/tests/LayerGradUtil.cpp

+        data.value->add(-0.5);
+        if (testLayerName != "prelu") {
+          data.value->sigmoid(*data.value);
+        }


406-408行可以去掉，这里不会用prelu这个layer

… fix-2132

luotao1 · 2019-02-01T04:43:57Z

感谢您给PaddlePaddle贡献代码。由于Paddle V1/V2版本已不再维护，相关代码也已从develop分支上删除，因此关闭您的PR，欢迎您向Paddle最新版-Fluid贡献代码。
Thanks for contributing to PaddlePaddle! Since V1/V2 will not be maintained anymore, and related codes have been deleted from develop branch as well, we close this PR. Welcome to contribute to Fluid——the latest version of PaddlePaddle.

pkuyym requested review from lcy-seso and luotao1 May 15, 2017 03:30

preliminary implementation

6adf4ac

pkuyym force-pushed the fix-2132 branch from 6e0b9a3 to 6adf4ac Compare May 15, 2017 03:35

pkuyym mentioned this pull request May 15, 2017

Enrich ConvShift to support sequence data input #2132

Closed

luotao1 requested changes May 15, 2017

View reviewed changes

luotao1 reviewed May 15, 2017

View reviewed changes

lcy-seso requested changes May 15, 2017

View reviewed changes

merge convolution logic into class Matrix

8cd2222

pkuyym force-pushed the fix-2132 branch from 05145d9 to 8cd2222 Compare May 15, 2017 08:19

yangyaming added 2 commits May 16, 2017 18:12

merge convolution logic into class Matrix

25cdee6

add unitest

167ceb5

pkuyym requested a review from qingqing01 May 17, 2017 03:27

Merge branch 'fix-2132' of https://github.com/pkuyym/Paddle into fix-…

cec53b9

…2132

pkuyym force-pushed the fix-2132 branch from ab1abe1 to cec53b9 Compare May 17, 2017 08:44

luotao1 reviewed May 23, 2017

View reviewed changes

yangyaming and others added 4 commits May 24, 2017 14:37

remove sigmoid activation

aa4ac87

Remove INPUT_SEQUENCE_MNUM_DATA type in LayerGradUtil

a4e5e66

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

490e35d

… fix-2132

Use SELF_DEFINE_DATA to build unit test for ConvShiftLayer.

eeccac1

luotao1 closed this Feb 1, 2019

heavengate pushed a commit to heavengate/Paddle that referenced this pull request Aug 16, 2021

fit for fcos (PaddlePaddle#2133)

d48b3ff

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Enrich ConvShift to support sequence data input #2133

Enrich ConvShift to support sequence data input #2133

pkuyym commented May 15, 2017

luotao1 left a comment

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 left a comment

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 May 15, 2017

pkuyym May 16, 2017

luotao1 May 15, 2017

pkuyym May 16, 2017

lcy-seso left a comment

lcy-seso May 15, 2017

pkuyym May 16, 2017

lcy-seso May 15, 2017

pkuyym May 17, 2017

pkuyym commented May 15, 2017

luotao1 May 23, 2017

pkuyym May 24, 2017

luotao1 commented Feb 1, 2019

		@@ -69,122 +66,12 @@ bool ConvShiftLayer::init(const LayerMap& layerMap,
		return true;
		}

Enrich ConvShift to support sequence data input #2133

Enrich ConvShift to support sequence data input #2133

Conversation

pkuyym commented May 15, 2017

luotao1 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

luotao1 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lcy-seso left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

pkuyym commented May 15, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

luotao1 commented Feb 1, 2019