Remove opencv dependency from tests

[guarin]

We still have an old opencv dependency in the tests, see here: https://github.com/lightly-ai/lightly/blob/2b215aa5d19b50515f280c1b0181f2f135cbc93f/tests/data/test_VideoDataset.py#L11

We should refactor the test(s) to not rely on opencv and remove opencv from all requirements files.

[ishaanagw]

Hey @guarin, Can I work on this ?

[SauravMaheshkar]

Sure @ishaanagw give it a go 😃

[ishaanagw]

Can you please explain the task a little more please ? Do I have to replace the videoWriters from OpenCV to some other library ?

[guarin]

Can you please explain the task a little more please ? Do I have to replace the videoWriters from OpenCV to some other library ?

Yes the idea would be to replace cv2.VideoWriter with write_video from torchvision in lightly/tests/data/test_VideoDataset.py. Looking at it this could actually be a more tricky issue than anticipated. If you would like to work on it, could you try the following?

• Replace all cv2.VideoWriter with torchvision write_video calls instead
• Replace the numpy arrays containing the frames (see here) with tensors because write_video expects tensors
• Try to run the tests with pytest tests/data/test_VideoDataset.py

If you run into errors regarding video timestamps we most likely can't remove cv2.

[ishaanagw]

Getting this error - ValueError: Expected numpy array with ndim 3 but got 2 in self._test_video_dataset_non_increasing_timestamps()

Is this what you were talking about ?

[guarin]

Interesting, could you post the full stack trace?

Is this what you were talking about ?

No I believe the ValueError already happens before. With the timestamps I meant that depending on which library is used when creating a new video file, the timestamps stored in the file can be different. And I am not sure if write_video creates the exact same timestamps as cv2.VideoWriter.

[ishaanagw]

Stack Trace - tests/data/test_VideoDataset.py:238: in _test_video_dataset_non_increasing_timestamps self.create_dataset(n_videos=2, n_frames_per_video=5) tests/data/test_VideoDataset.py:83: in create_dataset out = torchvision.io.write_video(path, frame, n_frames_per_video) ../miniconda3/envs/lightly/lib/python3.10/site-packages/torchvision/io/video.py:131: in write_video frame = av.VideoFrame.from_ndarray(img, format="rgb24") av/video/frame.pyx:551: in av.video.frame.VideoFrame.from_ndarray

[guarin]

This looks like something is wrong with the dimensions of the numpy arrays. Each frame should have dimension (W, H, C) with C = 3. But somehow the array it got has only 2 dimensions instead of 3.

[SGCODEX]

Can I work on this issue as well @guarin

[ishaanagw]

I think we can remove it, I was able to remove cv2 dependencies but facing some test failure, these test fail with no changes as well.

Failed Test Cases - FAILED tests/data/test_VideoDataset.py::TestVideoDataset::test_video_dataset_dataloader__pyav - TypeError: cannot pickle '_thread.lock' object FAILED tests/data/test_VideoDataset.py::TestVideoDataset::test_video_dataset_non_increasing_timestamps__pyav - TypeError: cannot pickle 'VideoLoader' object

Do you think this is due to some local setup ?

[moli-debugger]

I would like to pick this up as well @guarin

[guarin]

Do you think this is due to some local setup ?

Could be, which operating system are you using? You can also already create a PR and push it, then we can check if it is due to the local setup or not.

@SGCODEX @moli-debugger Its best if one person works on an issue at the time. Please have a look at our other issues for contribution opportunities.

[moli-debugger]

@SGCODEX , please pick up this task . I will continue work on the other issues.

CC: @guarin

[ishaanagw]

Could be, which operating system are you using?

Mac OS

You can also already create a PR and push it, then we can check if it is due to the local setup or not.

Created PR for this - https://github.com/lightly-ai/lightly/pull/1679