fix bug in roi pooling forwarding with cpu

[apple2373]

I implemented fast r-cnn https://github.com/apple2373/chainer-simple-fast-rnn, and realized roi pooling with cpu does not work. It always outputs only one roi when forwarding. I just copied two lines from GPU implementation. Also I am not sure backward function is also wrong or not because I only use forwarding now.

[okuta]

Thank you for your PR.
Would you write the test?

[apple2373]

Unfortunately I am not familiar with software testing or software engineering. It seems that roi pooling already has the test: https://github.com/apple2373/chainer/blob/c4b5082e3b0d2973152022aecefb07486b806f5e/tests/chainer_tests/functions_tests/pooling_tests/test_roi_pooling_2d.py

I want to ask inputs from @wkentaro who is the developer of this layer. I guess this is the easy mistake to understand for him.

[wkentaro]

The code looks good to me.
But I'm not sure how to test this.. the test_roi_pooling_2d.py already has test_forward_cpu_gpu_equal and it passed unexpectedly..

[mitmul]

I came across the same error.

@apple2373 As for the test, how about modifying this line from N = 4 to N = 3 here?: pooling_tests/test_roi_pooling_2d.py#L17
The number of images in a minibatch and number of rois are incidentally same. That's why the test was passed I think, so please consider to try this modification.

[t-abe]

mitmul's idea looks good for testing this PR.
@apple2373 Could you modify the test as he mentioned?

[apple2373]

I really do not know about software testing. Is it ok just to modify the test_roi_pooling_2d.py as mitmul mentioned, and add it to this PR?

[t-abe]

Yes. You can add commits into fix_roi_pooling_cpu branch, then this PR is automatically updated.

[apple2373]

I did it.

[t-abe]

Added test failed.
Probably we should fix same bug in backward.
https://github.com/apple2373/chainer/blob/cbfd52bf0d9b26aa8451c4dbf65d991b2b0ed824/chainer/functions/pooling/roi_pooling_2d.py#L188

[apple2373]

I modified cpu backward in the same way.

[t-abe]

Thank you.
The test still failed. I am going to check the code.
Could you do it too?
You can run a test by $ nosetests -v tests/chainer_tests/function_tests/pooling_tests/....py.

[apple2373]

I am not sure what is going on but the test will end up "RuntimeError: CUDA environment is not correctly set up". Maybe test cannot be run only with CPU. I'll try to run on GPU server later.

[t-abe]

You can run tests only on cpu by giving -a '!gpu' option to nosetests.
And you can see the test results here: https://travis-ci.org/pfnet/chainer/jobs/141322278 .

[apple2373]

Thank you. I was able to reproduce test failure on my laptop, but I do not think I can fix it because I do not fully understand the implementation.

[t-abe]

I fixed the bug and open PR #1352.