add unfold op (new op)#17944
Merged
SunGaofeng merged 4 commits intoPaddlePaddle:developfrom Jun 11, 2019
Merged
Conversation
test=develop
SunGaofeng
changed the title
heavengate
reviewed
Jun 10, 2019
paddle/fluid/operators/unfold_op.cc
Outdated
| This Operator is used to extract sliding local blocks from a batched input tensor, also known | ||
| as im2col when operated on batched 2D image tensor. For each block under the convolution filter, | ||
| all element will be rearranged as a column. While the convolution filter silding over the input | ||
| feature map, a series of such columns will be formed. |
Contributor
There was a problem hiding this comment.
give same examples or formula
Contributor
Author
There was a problem hiding this comment.
Yes, the formula is given in the python api
paddle/fluid/operators/unfold_op.cc
Outdated
| strides.size(), kernel_sizes.size(), | ||
| "The dims of strides shold be the same with that of kernel_sizes. " | ||
| "But recieved dims(strides: %u) != dims(kernel_sizes: %u).", | ||
| strides.size(), dilations.size()); |
Member
There was a problem hiding this comment.
strides.size(), dilations.size() -> strides.size(), kernel_sizes.size()
heavengate
reviewed
Jun 10, 2019
| auto& dev_ctx = ctx.template device_context<DeviceContext>(); | ||
|
|
||
| math::SetConstant<DeviceContext, T> set_zero; | ||
| set_zero(dev_ctx, input_grad, static_cast<T>(0)); |
Contributor
There was a problem hiding this comment.
下面for循环里input_grad会有点赋不到值么，如果没有的话这里应该可以不用初始化吧
Contributor
Author
There was a problem hiding this comment.
在Functor col2im里面，计算input_grad的时候会用的是累加，所以需要先把input_grad清零。
heavengate
reviewed
Jun 10, 2019
| std::unique_ptr<framework::OpDesc> op(new framework::OpDesc()); | ||
| op->SetType("unfold_grad"); | ||
| op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); | ||
| op->SetInput("X", Input("X")); |
Contributor
Author
There was a problem hiding this comment.
需要input(X)的dims，然后加上UnfoldGradOpNoNeedBufferVarsInference声明不需要input(X)的数据，这样在计算grad时只会保留input(X)的dims而不用保留数据。
wanghaoshuang
previously approved these changes
Jun 10, 2019
| @@ -0,0 +1,102 @@ | |||
| # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. | |||
| This is for test on unfold Op | ||
| """ | ||
|
|
||
| def init_data(self): |
| dkernel_h = self.dilations[0] * (self.kernel_sizes[0] - 1) + 1 | ||
| dkernel_w = self.dilations[1] * (self.kernel_sizes[1] - 1) + 1 | ||
| out_height = (self.input_height + self.paddings[0] + self.paddings[2] - | ||
| dkernel_h) / self.strides[0] + 1 |
SunGaofeng
dismissed stale reviews from heavengate and wanghaoshuang
via
e658cb6
wanghaoshuang
previously approved these changes
Jun 10, 2019
qingqing01
reviewed
Jun 10, 2019
python/paddle/fluid/layers/nn.py
Outdated
|
|
||
| def unfold(x, kernel_sizes, strides=1, paddings=0, dilations=1): | ||
| """ | ||
| **unfold** |
python/paddle/fluid/layers/nn.py
Outdated
| return output | ||
|
|
||
|
|
||
| def unfold(x, kernel_sizes, strides=1, paddings=0, dilations=1): |
paddle/fluid/operators/unfold_op.cc
Outdated
|
|
||
| int dkernel_h = dilations[0] * (kernel_sizes[0] - 1) + 1; | ||
| int dkernel_w = dilations[1] * (kernel_sizes[1] - 1) + 1; | ||
| int conv_out_height = |
Contributor
There was a problem hiding this comment.
conv_out_height -> out_height remove prefix conv_
paddle/fluid/operators/unfold_op.h
Outdated
| 1; | ||
| int output_width = | ||
| (input_dims[3] + paddings[1] + paddings[3] - dkernel_w) / strides[1] + | ||
| 1; |
Contributor
There was a problem hiding this comment.
这些代码重复了3次，conv里有抽取一些公用的函数吧
Contributor
Author
There was a problem hiding this comment.
另外写了一个inline fucntion， 没有用conv_op.h里面的计算conv output 尺寸的函数，那里面默认为padding左右是一样的，在unfold里面padding可以设置成上下左右四个值都不一样的。
test=develop
qingqing01
approved these changes
Jun 11, 2019
SunGaofeng
added a commit
to SunGaofeng/Paddle
that referenced
this pull request
Jun 11, 2019
* add unfold op test=develop * fix divide bug in python3 when calculating output width and height test=develop * add name=None in python api, move redundant code into inline function * try to trigger ci for this code test=develop
SunGaofeng
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is for unfold op (new op), which is also called im2col for batched 2D image tensor. Details are in the following description: