-
Notifications
You must be signed in to change notification settings - Fork 150
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
Adds new CI, Covers Multiple OS, Python Versions and Packaging #226
Conversation
Codecov Report
@@ Coverage Diff @@
## master #226 +/- ##
=======================================
Coverage 80.17% 80.17%
=======================================
Files 101 101
Lines 1897 1897
=======================================
Hits 1521 1521
Misses 376 376
Flags with carried forward coverage won't be shown. Click here to find out more. Continue to review full report at Codecov.
|
I was too able to reproduce this error on my windows Machine |
@lgvaz can you fix up the windows build issue of efficient det repo ? Once that is solved and I fix up the packaging We can then add these CI changes. I will fix up the |
@oke-aditya This is very very hard for me to solve, because, well.. I don't have windows hahahhah Does windows support multi-gpu after all? |
Windows does support We are installing the same package in Linux and MacOs as It is also CPU only. It should work. Microsoft provides nice support for windows with pytorch now so it should be fixed. |
I'm facing a very small issue with Packaging install CI. Current scenario: - Possible solutions: - Other possible solution: - It's very strange as same workflow runs for PyTorch Lightning. And it doesn't run for us. |
So Windows CI error is linked to pytorch/pytorch#42095. Another option is mark these tests that fail on windows for now and skip them on windows OS. This is slightly poor technique as its somewhere in between of being supported or not. Windows users have option of docker container with WSL2, code on windows; train on Linux VMs, This isn't our issue anyways coz if PyTorch doesn't support it. We cannot do it either. |
AAh the install bug was fixed with |
Again the build failed coz we tried to |
Nice, we are getting there. New issue. We import PyTorch Lightning. Note we do not package this with our repo. It is not in our requirements as well, it is in additional requirements. So when we build the wheel, we do not package this dependency or take care that it is installed. The same strategy that of HuggingFace as well. They do not package PyTorch or Tensorflow or take care it is installed. Again this error arises from nasty imports folder, which is causing lot of concern.
Fixes possible.
Note that egg or wheel file will always install minimal version of mantisshrimp. |
|
Tests run on all. |
Okay. So the final broken things are Tests on Windows. This time it is not due to the distributed issue. It is due to refactor of COCO metrics. Windows we use different COCO Api repository. @lgvaz can you have a look it seems small dtype conversion error. |
Maybe this is the cause rbgirshick/py-faster-rcnn#481. We are using an upgraded version of numpy but the coco API needs a lower one ? The COCO APi is always painful and a bottleneck. |
Currently we don't have any test to let us know if it is fixed. So I guess better will be to merge this and try to fix it up. It will be easier to fix as then we will know if Windows CI passes. |
Following tests fail only on Windows
@lgvaz @ai-fast-track as we said Windows support is experimental. And reason is PyCoco Tools and PyTorch both have its limitations. Suggestions to Proceed: -
|
Does this ( C.2- Installing cocoapi in Windows:pycoco cannot be installed using the command above (see issue-185 in the cocoapi repository). We are using this workaround:
|
I am using the same COCO API on workflow file. Still I'm getting error. You can have look at workflow here Also in a few days. We need to bump to PyTorch 1.6 The reason is COCO API is very old and maybe there is some internal numpy typecast issue from our side or from this github repo itself. This is a liability to be carried forward. |
Bumping all the tests to PyTorch 1.6 with the new relase today, |
Wow. Additional trouble now. We can't bump to PyTorch 1.6. @lgvaz this needs multi-headed attention now (pun). Problem is with Ross's EfficientDet tests. Also, a test of fastai is failing.
from To avoid this from next time. We should run CI tests on PyTorch current as well as nightly. This was not very expected. But this increases our test count 2x. Which will now be too many. The Fastai error is
arising from learner.fine_tune() A fix for now maybe to ship with PyTorch 1.5 and try to cover 1.6. |
K converting this to draft PR. Will raise 2 new PRs one for CI and one for the PyPI release. |
Do let me know if we need to delete this branch |
With the current change I added CI for more OS systems. It is good idea to build on multiple OS before publishing.
This technique is taken from PyTorch Lightning.
I have added CI for releasing to both PyPi test server and PyPi official. This is again taken from Lightning
We need a test to build package as well. And test PyPi releasing with a dummy release.