Add dilations for conv2d and optimize conv2d code #5472

chengduoZH · 2017-11-08T06:18:01Z

Add dilations for im2col(ColFormat::kCFO).
Fix conv2d doc (Add dilation)
Refine conv2d_op unit test.(Add dilation)
Refine conv_op for filter size:(1,1), (when filter_size, paddings, strides, dilations are (1,1), (0, 0), (1,1), (1,1), conv_op just likes fc_op)
Add dilations for vol2col(ColFormat::kCFO).
Fix conv3d doc (Add dilation)
Refine conv3d_op unit test.(Add dilation)

qingqing01 · 2017-11-14T09:51:05Z

paddle/operators/conv_op.h

+    filter_1 &= (static_cast<int>(filter_dim[j]) == 1);
+    strides_1 &= (strides[j] == 1);
+    padding_0 &= (paddings[j] == 0);
+    dilation_1 &= (dilations[j] == 1);


&= -> &&= ?

没有这种表示&&=，我改成了strides_1 = strides_1 && (strides[j] == 1)

qingqing01 · 2017-11-14T09:54:22Z

paddle/operators/conv_op.h

-          vol2col(context.device_context(), in_slice, col, strides[0],
-                  strides[1], strides[2], paddings[0], paddings[1],
-                  paddings[2]);
+    if (!not_expand) {


! not expand = expand, the logic is a little complex,

How about rename NotExpand to IsExpand? Then return True, means that it needs to expand, ortherwise, not expand.

qingqing01 · 2017-11-14T10:35:56Z

paddle/operators/conv_op.h

+      for (int i = 0; i < batch_size; i++) {
+        Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
+        Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
+        for (int g = 0; g < groups; g++) {


for (int i = 0; i < batch_size; i++) { // .... for (int g = 0; g < groups; g++) { if(!IsExpand) { ShareDataWith(); } else if () { im2col(); } else if () { im2vol(); } } }

qingqing01 · 2017-11-14T10:36:53Z

paddle/operators/conv_op.h

+
+            math::matmul<Place, T>(context.device_context(), filter_slice, true,
+                                   out_grad_slice, false, T(1.0), &col_matrix,
+                                   T(0.0));


The code structure is same as above.

qingqing01 · 2017-11-14T11:30:58Z

paddle/operators/math/im2col.cc

+                          1,
+                      col_width,
+                      "col_width and padding(padding_left, padding_right) are "
+                      "inconsistent.");


写functor的时候，我也在考虑functor里面是否还有必要再次check shape的正确性，因为要么是在Op里计算得到的，要么InferShape里也已经check过。

是的，我也有想过这些，比如在Op里面检测过了，在GradOp中就不用检测了吧

qingqing01 · 2017-11-14T11:37:02Z

paddle/operators/math/im2col.cc

-                  int padding_down, int padding_left, int padding_right) {
+                  int dilation_h, int dilation_w, int stride_height,
+                  int stride_width, int padding_up, int padding_down,
+                  int padding_left, int padding_right) {


Maybe std::vector<int>& dilations, std::vector<int>& strides, std::vector<int>& paddings are short. And the op also uses std::vector<int>.

qingqing01 · 2017-11-14T14:03:21Z

paddle/operators/conv_op.h

+  int output_size = (input_size + padding_up + padding_down -
+                     (dilation * (filter_size - 1) + 1)) /
+                        stride +
+                    1;


const int dkernel = dilation * (filter_size - 1) + 1; const int output_size = (input_size + padding_up + padding_down - dkernel)/stride + 1;

qingqing01 · 2017-11-14T14:04:33Z

paddle/operators/math/vol2col.cc

+                          1,
+                      output_width,
+                      "input_width and output_width are "
+                      "Mismatching.");


Same as above, whether it needs to check again?

I think we can write in this way first, and discuss it later. Because other functors also have similar problem.

qingqing01 · 2017-11-14T14:07:23Z

paddle/operators/math/vol2col.cu

+                     output_height +
+                 h_col) *
+                    output_width +
+                w_col;


data_col_index的计算太长，不容易看清楚。一些计算可以移到各自循环里。

把这个公式分成了两个，现在可能会好一点

… refine_im2col

chengduoZH force-pushed the refine_im2col branch 4 times, most recently from 10101ea to 27805c2 Compare November 8, 2017 06:53

qingqing01 added the OpPorting label Nov 8, 2017

chengduoZH force-pushed the refine_im2col branch 3 times, most recently from 607e0e8 to 520dec7 Compare November 8, 2017 10:04

add dilation for im2col

97e9dd7

chengduoZH force-pushed the refine_im2col branch from 520dec7 to 97e9dd7 Compare November 8, 2017 10:08

chengduoZH changed the title ~~add dilations for im2col~~ add dilations for conv2d Nov 8, 2017

fix conv2d doc

b6f9ba4

chengduoZH requested review from hedaoyuan, NHZlX, qingqing01 and typhoonzero November 8, 2017 12:45

chengduoZH changed the title ~~add dilations for conv2d~~ Add dilations for conv2d and optimize conv2d code Nov 9, 2017

chengduoZH force-pushed the refine_im2col branch from b125570 to bd73642 Compare November 9, 2017 05:47

refine conv2d for filter size:(1,1)

21ce704

chengduoZH force-pushed the refine_im2col branch 2 times, most recently from b2d5245 to caf24f3 Compare November 9, 2017 08:39

refine unit test (Add dilation)

93551bd

chengduoZH force-pushed the refine_im2col branch from caf24f3 to 93551bd Compare November 10, 2017 01:51

Add dilation for vol2col

271fc9c

chengduoZH force-pushed the refine_im2col branch 2 times, most recently from a5bbe8d to 3e60b6b Compare November 10, 2017 05:55

fix unit test (conv3d)

7d73b8f

chengduoZH force-pushed the refine_im2col branch 2 times, most recently from a3e15fd to 7d73b8f Compare November 14, 2017 11:34

qingqing01 reviewed Nov 14, 2017

View reviewed changes

chengduoZH added 2 commits November 15, 2017 12:49

follow comments

356d695

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

e25bfc7

… refine_im2col

chengduoZH force-pushed the refine_im2col branch 2 times, most recently from 1607de4 to 8fffa9e Compare November 15, 2017 07:21

fix ContextProjectFunctor parameter order

31dc019

chengduoZH force-pushed the refine_im2col branch from 8fffa9e to 31dc019 Compare November 15, 2017 11:00

chengduoZH added 2 commits November 15, 2017 19:57

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

f91422e

… refine_im2col

remove conflict

00e0881

qingqing01 approved these changes Nov 15, 2017

View reviewed changes

chengduoZH merged commit 4fc9f55 into PaddlePaddle:develop Nov 15, 2017

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add dilations for conv2d and optimize conv2d code #5472

Add dilations for conv2d and optimize conv2d code #5472

chengduoZH commented Nov 8, 2017 •

edited

Loading

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

qingqing01 Nov 14, 2017

chengduoZH Nov 15, 2017

Add dilations for conv2d and optimize conv2d code #5472

Add dilations for conv2d and optimize conv2d code #5472

Conversation

chengduoZH commented Nov 8, 2017 • edited Loading

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

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

chengduoZH commented Nov 8, 2017 •

edited

Loading