-
Notifications
You must be signed in to change notification settings - Fork 2.4k
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.
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
Fix deformable_convolution #6393
Fix deformable_convolution #6393
Conversation
@DingZhangIntel Could you please add new tests? |
|
I have added a single test case in the latest commit, but I don't know whether this is what you want. If it is, I think it is a simple example to explain why I made these changes. If not, please contact me at any your time, I will give more details about this PR. |
@DingZhangIntel Looks like new tests are failing. |
cead7cc
to
7fc16e5
Compare
I just made a little changes in the origin test file, but it still fails on CI. I think it's a little weird. I will very appreciate it if you can check these changes at your convenience. I will keep looking into this issue, thanks. |
As I see the CPU plugin cannot create a network for your tests:
|
I just tested it in my local openvino repo and found that if i set the parameter group equals 2, the test will fail. And if I set the parameter group equals 1, the test will work. please check. |
I tested it again and found that if I set the parameter group greater than 1, the test will fail. Please check. |
I changed the tests back and added a single test case to explain why I made these changes in ref implementation, and the single test case will produce the mismatch results comparing with MXNet. The error will occur on line 256 of the origin ref implementation, the offset of the variable |
@itikhono Please have a look. |
hello @DingZhangIntel, As understood, you encounted 2 issues:
|
* Fix deformable_convolution * Fix unused-variable * Change a test case * Change tests back and add single test case
* Fix deformable_convolution * Fix unused-variable * Change a test case * Change tests back and add single test case
* Fix deformable_convolution * Fix unused-variable * Change a test case * Change tests back and add single test case
I think there is a error about the reference implementation of deformable_convolution.
The two parameters group and deformable_group is independent, and there are details about the parameter deformable_group in MXNet
Suppose that the shape of input data is (1,6,5,5), the shape of filter is (2,3,3,3), the shape of the offset is (1,54,3,3), group is 2, deformable_group is 3. And an error will occur on line 256 when group_idx equals 1.