monkey-patching wandb.watch when training a fastai RCNN learner with …

[FraPochetti]

…the WandbCallback

This one addresses issue #490

[codecov]

Codecov Report

Merging #524 into master will decrease coverage by 0.15%.
The diff coverage is 50.00%.

@@            Coverage Diff             @@
##           master     #524      +/-   ##
==========================================
- Coverage   86.63%   86.47%   -0.16%     
==========================================
  Files          91       91              
  Lines        2184     2189       +5     
==========================================
+ Hits         1892     1893       +1     
- Misses        292      296       +4     

Flag	Coverage Δ
unittests	86.47% <50.00%> (-0.16%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
icevision/models/rcnn/fastai/learner.py	82.60% <50.00%> (-17.40%)	⬇️
icevision/core/mask.py	86.82% <0.00%> (-0.30%)	⬇️
icevision/models/rcnn/mask_rcnn/model.py	100.00% <0.00%> (ø)
icevision/models/rcnn/faster_rcnn/model.py	100.00% <0.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update be2c4ba...c266ea3. Read the comment docs.

[lgvaz]

Looks good! Just a question before we merge this, if instead of patching we use log=None, is the bug solved?

[FraPochetti]

if instead of patching we use log=None, is the bug solved?

Nope. That just doesn't log the gradients.

[lgvaz]

Okay, makes a lot of sense!!

So, the final thing before we merge, can you add a logger.warning if if len(is_wandb) == 1:?

It would be nice to open a new issue in icevision to show your findings, you can just copy and paste the messages you sent me on discord, explaining that torchvision rcnn models are having a problem with the wandb hooks and the best quick solution we found is to simply don't add the hooks

Then in the warning message you could say something like "Wandb quickfix implemented, for more info check issue #number_of_issue"

[FraPochetti]

Sure will do!
Shall I just document my findings here or create a new issue from scratch?

[FraPochetti]

I added a final comment to this issue describing what we did.
Is this a sufficient level of detail?

[lgvaz]

The level of detail is perfect!

Open another issue because the original one contains useless information and will be closed by this PR. It's easier to have a direct and clear issue for further reference =)

[FraPochetti]

Done.
Issue created and PR updated.

[lgvaz]

Looks perfect! Thanks SO MUCH again Francesco, as I said before, the amount of line changes doesn't reflect at all the work put into this ❤️