auto learning rate finder fails AFTER optimization due to misconfiguration

[noamzilo]

🐛 Bug

Optimizing learning rate using auto_lr_find=True fails AFTER a learning rate was found with error

pytorch_lightning.utilities.exceptions.MisconfigurationException: When auto_lr_find is set to True, expects that modelormodel.hparamseither has fieldlrorlearning_rate that can overridden

AFTER having waited for several minutes.

This can be caught BEFORE the long run to save time.

Please reproduce using the BoringModel

I can just copy paste everything to here... the only lacking line is self.lr = 999 under model::__init__(), if it is not present the error happens, otherwise it doesn't happen.

Reproducing this a single time takes me 15 minutes (due to the lr finding process), but I still wanted to report this because it seems important and I don't have time right now for a full reproduction on the template.

trainer = Trainer(auto_lr_find=True) trainer.tune(model)

then for a while I see the progress bar

Finding best initial lr: 27%|██▋ | 27/100 [02:43<07:37, 6.27s/it]

and then the error after a lr has been found.

Expected behavior

pytorch_lightning.utilities.exceptions.MisconfigurationException: When auto_lr_find is set to True, expects that modelormodel.hparamseither has fieldlrorlearning_rate that can overridden

should hit BEFORE the long calculation.

Environment

• PyTorch Version (e.g., 1.0): 1.7.1
• OS (e.g., Linux): Windows
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source): -
• Python version: 3.7
• CUDA/cuDNN version: 101
• GPU models and configuration: single GPU NVIDIA quadro P4000
• Any other relevant information:-

[rohitgr7]

good catch! Mind send a PR with a fix?

[noamzilo]

good catch! Mind send a PR with a fix?

I actually never contributed to open source before :)

How does this work? I get assigned and there's a deadline?
is there a readme?

I want to do it!

[carmocca]

We have a basic guide here!

There's no deadline! Any contribution is welcome 😄

[noamzilo]

Can the same issue be done twice by some people w/o them knowing about it until the latter does a PR?

[carmocca]

Yes. For this reason, you might want to create the PR in draft mode while you are working on it, so others can know about it

https://github.blog/2019-02-14-introducing-draft-pull-requests/

[edenlightning]

Thanks for contributing @noamzilo! I have assigned this ticket to you, so other contributors know that this is already in progress. please let us know if you need any help!

[noamzilo]

How to link this post to the PR?
PR at:

#5638

currently my test doesn't pass due to lack of

pip install pytest-timeout (from https://stackoverflow.com/questions/19527320/how-can-i-limit-the-maximum-running-time-for-a-unit-test) in the test server.

Please allow this library, or advise how I should limit test time.

Thanks :)