Important note on tests:
- they must be smart: improving coverage just for the sake of coverage isn’t a goal. We also run the full doc notebooks as a test before a release, which has probably a really good coverage, but it doesn’t necessarily catch subtle bugs
- they must be fast: no downloading torchvision models, no training apart from the 4 current integration tests (one in vision, one in text, one in tabular and one in collab) or be marked as slow (but we’d prefer fast tests).
- they should abide by the style guide of fastai
Thanks a lot for helping!