Skip to content

Enhance concat op to support empty input. #17015

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jerrywgz merged 3 commits into from
May 5, 2019
Merged

Enhance concat op to support empty input. #17015

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
02fcdda
enhance_concat, test=develop
jerrywgz Apr 22, 2019
821dd88
refine code & add case for unittest, test=develop
jerrywgz Apr 24, 2019
ad10ce4
refine code, test=develop
jerrywgz Apr 30, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 11 additions & 3 deletions paddle/fluid/operators/concat_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,9 @@ class ConcatKernel : public framework::OpKernel<T> {
if (axis == 0 && ins.size() < 10) {
size_t output_offset = 0;
for (auto* in : ins) {
if (!in || in->numel() == 0UL) {
continue;
}
auto in_stride = framework::stride_numel(in->dims());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis,
Expand All @@ -45,9 +48,13 @@ class ConcatKernel : public framework::OpKernel<T> {
output_offset += in_stride[axis];
}
} else {
std::vector<framework::Tensor> inputs(ins.size());
std::vector<framework::Tensor> inputs;
for (size_t j = 0; j < ins.size(); ++j) {
inputs[j] = *ins[j];
if (ins[j] || ins[j]->numel() > 0) {
inputs.push_back(*ins[j]);
} else {
continue;
}
Copy link
Copy Markdown
Contributor

@qingqing01 qingqing01 Apr 23, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more. 

if (ins[j] || in->numel() > 0) {
  inputs.push_back(*ins[j]);
} else {
  continue;
}

Copy link
Copy Markdown
Contributor Author

@jerrywgz jerrywgz Apr 24, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done, but here should be || or &&?

Copy link
Copy Markdown
Contributor

@qingqing01 qingqing01 Apr 25, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, should be &&

Copy link
Copy Markdown
Contributor

@qingqing01 qingqing01 Apr 26, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@jerrywgz please to fix code

Copy link
Copy Markdown
Contributor Author

@jerrywgz jerrywgz Apr 30, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done

}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatFunctor<DeviceContext, T> concat_functor;
Expand Down Expand Up @@ -82,7 +89,8 @@ class ConcatGradKernel : public framework::OpKernel<T> {
// get output tensor that the name is not kEmptyVarName
std::vector<framework::Tensor*> outputs;
for (size_t j = 0; j < outs.size(); ++j) {
if (out_var_names[j] != framework::kEmptyVarName) {
if (out_var_names[j] != framework::kEmptyVarName &&
outs[j]->numel() != 0UL) {
outs[j]->mutable_data<T>(ctx.GetPlace());
outputs.push_back(outs[j]);
} else {
Expand Down
11 changes: 11 additions & 0 deletions python/paddle/fluid/tests/unittests/test_concat_op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,5 +64,16 @@ def test_check_grad(self):
pass


class TestConcatOp4(TestConcatOp):
def init_test_data(self):
self.x0 = np.random.random((2, 3, 4, 5)).astype('float32')
self.x1 = np.random.random((2, 3, 4, 5)).astype('float32')
self.x2 = np.random.random((0, 3, 4, 5)).astype('float32')
self.axis = 0

def test_check_grad(self):
pass
Copy link
Copy Markdown
Contributor

@qingqing01 qingqing01 Apr 25, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why pass check_grad ?

Copy link
Copy Markdown
Contributor Author

@jerrywgz jerrywgz Apr 25, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because it will meet fetch error when there is a null grad



if __name__ == '__main__':
unittest.main()