Add max pool op (with index) #4461

chengduoZH · 2017-09-28T02:45:07Z

chengduoZH · 2017-09-30T07:33:05Z

I've commented out the check_grad in max pooling because the current method of testing can not effectively test the gradient of max pooling. But it ensures that the gradients computed by GPU and CPU are equal.

… Add_maxpool_withIdx_only

typhoonzero · 2017-10-09T05:42:50Z

paddle/operators/CMakeLists.txt

@@ -75,6 +75,12 @@ function(op_library TARGET)
        file(APPEND ${pybind_file} "USE_OP(reduce_sum);\n")
    endif()

+    if ("${TARGET}" STREQUAL "pool_with_index_op")
+        set(pybind_flag 1)
+        # It's enough to just adding one operator to pybind


This comment does not make clear why need add another name to this operator. See above comments, like:

# reduce_op contains several operators

typhoonzero · 2017-10-09T05:47:05Z

paddle/operators/math/pooling.cc

@@ -458,6 +458,233 @@ template class Pool3dGradFunctor<
    platform::CPUPlace, paddle::operators::math::MaxPoolGrad<double>, double>;
 template class Pool3dGradFunctor<
    platform::CPUPlace, paddle::operators::math::AvgPoolGrad<double>, double>;
+
+template <typename T>
+class MaxPool2dWithIndexFunctor<platform::CPUPlace, T> {


Need to add some comments about these functors, like "Input order NCHW or NHWC"

Please take a glance at https://google.github.io/styleguide/cppguide.html#Function_Comments

Done. Thx !!!

typhoonzero · 2017-10-09T06:10:44Z

paddle/operators/pool_with_index_op.cc

+namespace paddle {
+namespace operators {
+
+int OutputSizeMaxPool(int input_size, int filter_size, int padding,


This function can be a inline. And, this function can be of more general use for conv and pool operators.

Done.
This function is frequently called in conv2/3d, conv2/3d_cudnn, pool2/3d and pool2/3_cudnn.
I think we should convert this function to functor, and move to namespace math.

typhoonzero · 2017-10-09T06:12:24Z

paddle/operators/pool_with_index_op.cc

+
+namespace ops = paddle::operators;
+
+REGISTER_OP(maxPool2dWithIndex, ops::MaxPoolWithIndexOp,


Seems operator names are commonly xxx_xxx_op but not "camal case".

typhoonzero

Some minor comments.

typhoonzero · 2017-10-09T13:42:24Z

paddle/operators/CMakeLists.txt

+    if ("${TARGET}" STREQUAL "pool_with_index_op")
+        set(pybind_flag 1)
+        # It's enough to just adding one operator to pybind
+        file(APPEND ${pybind_file} "USE_OP(max_pool2d_with_index);\n")


Also need to add max_pool3d_with_index

Just adding one operator of pool_with_index_op is ok. And unit test(max_pool2d_with_index, max_pool3d_with_index) passed.

typhoonzero · 2017-10-09T13:49:50Z

paddle/operators/pool_with_index_op.cc

+                   "X(Input) of Pooling should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput(framework::GradVarName("X")),
+        "X@GRAD(Input@GRAD) of MaxPoolWithIndexOpGrad should not be null.");


@GRAD suffix is for internal use, may confuse users if this message came out.

I approve with you. If X@GRAD is null, it must be an internet error. But I think it is helpful to developer.

typhoonzero · 2017-10-09T13:50:03Z

paddle/operators/pool_with_index_op.cc

+        "ksize",
+        "Pooling size(height, width) of pooling operator."
+        "If globalPooling = true, ksize is ignored and need not be "
+        "specified.");  // TODO(Add checker)


Seems you need to configure pre-commit on your local environment before running git commit, so the style check can automatically prints some style errors.

In this case, "TODO" comments should be like: https://google.github.io/styleguide/cppguide.html#TODO_Comments

Thanks, this style of writing is really nonstandard.
I have configured pre-commit on my local environment, but it also passed.

typhoonzero · 2017-10-09T13:53:27Z

paddle/operators/pool_with_index_op.cc

+        "image.");
+    AddOutput("Out",
+              "The output tensor of pooling operator."
+              "The format of output tensor is also NCDHW."


"The format of output tensor is also NCDHW."
"Where N is batch size, C is "
"the number of channels, D, H and W is the depth, height and "
"width of image."

Can put these explains to AddComment, so here can be more simpler.

typhoonzero

LGTM++

chengduoZH changed the title ~~Add max pool op (with index)~~ [WIP] Add max pool op (with index) Sep 28, 2017

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 1cb037f to 8a97324 Compare September 28, 2017 03:33

Xreki added the OpPorting label Sep 28, 2017

chengduoZH force-pushed the Add_maxpool_withIdx_only branch 4 times, most recently from e4e6964 to d987c44 Compare September 28, 2017 08:48

Add max pool with index

6326c40

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 2149e40 to 6326c40 Compare September 28, 2017 09:31

chengduoZH requested review from hedaoyuan and QiJune September 29, 2017 11:35

chengduoZH force-pushed the Add_maxpool_withIdx_only branch 3 times, most recently from 2d72558 to d907eba Compare September 29, 2017 12:24

chengduoZH changed the title ~~[WIP] Add max pool op (with index)~~ Add max pool op (with index) Sep 29, 2017

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from d907eba to 5b606f3 Compare September 30, 2017 06:05

fix code format and some bug

bee95fc

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 5b606f3 to bee95fc Compare September 30, 2017 07:06

fix unit test

6fc4480

remove conflict

2ed56df

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 59f20d5 to f5e625a Compare September 30, 2017 07:48

chengduoZH requested review from luotao1 and NHZlX September 30, 2017 07:59

fix kernel func

bb33c2b

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 42b9a34 to bb33c2b Compare September 30, 2017 09:00

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

14b2c98

… Add_maxpool_withIdx_only

typhoonzero reviewed Oct 9, 2017

View reviewed changes

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 232077d to 811cf0b Compare October 9, 2017 11:32

follow coments

fcfce48

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 811cf0b to fcfce48 Compare October 9, 2017 11:36

typhoonzero reviewed Oct 9, 2017

View reviewed changes

follow coments

6c6474c

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from c93063a to 6c6474c Compare October 10, 2017 03:26

fix atomicAdd -> CudaAtomicAdd

e21e564

chengduoZH force-pushed the Add_maxpool_withIdx_only branch 2 times, most recently from 5308f84 to 68b0508 Compare October 10, 2017 08:29

Add code comments

36da825

chengduoZH force-pushed the Add_maxpool_withIdx_only branch from 68b0508 to 36da825 Compare October 10, 2017 10:55

typhoonzero approved these changes Oct 10, 2017

View reviewed changes

chengduoZH merged commit 0f1d3af into PaddlePaddle:develop Oct 10, 2017

chengduoZH added this to Done in Port Operators Oct 19, 2017

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add max pool op (with index) #4461

Add max pool op (with index) #4461

chengduoZH commented Sep 28, 2017 •

edited

chengduoZH commented Sep 30, 2017 •

edited

typhoonzero Oct 9, 2017

chengduoZH Oct 9, 2017

typhoonzero Oct 9, 2017

chengduoZH Oct 9, 2017

typhoonzero Oct 9, 2017

chengduoZH Oct 9, 2017

typhoonzero Oct 9, 2017

chengduoZH Oct 9, 2017

typhoonzero left a comment

typhoonzero Oct 9, 2017

chengduoZH Oct 10, 2017

typhoonzero Oct 9, 2017

chengduoZH Oct 10, 2017

typhoonzero Oct 9, 2017

chengduoZH Oct 10, 2017 •

edited

typhoonzero Oct 9, 2017

chengduoZH Oct 10, 2017

typhoonzero left a comment


		namespace ops = paddle::operators;

		REGISTER_OP(maxPool2dWithIndex, ops::MaxPoolWithIndexOp,

Add max pool op (with index) #4461

Add max pool op (with index) #4461

Conversation

chengduoZH commented Sep 28, 2017 • edited

chengduoZH commented Sep 30, 2017 • edited

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

typhoonzero 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

chengduoZH Oct 10, 2017 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

typhoonzero left a comment

Choose a reason for hiding this comment

chengduoZH commented Sep 28, 2017 •

edited

chengduoZH commented Sep 30, 2017 •

edited

chengduoZH Oct 10, 2017 •

edited