search

sinzlab / sensorium

Code base for the SENSORIUM competition.
https://sensorium2022.net/
MIT License
59 stars 32 forks source link

Improving model-training notebook and 'trainers.py' #112

Open pollytur opened 2 years ago

pollytur commented 2 years ago
  1. The standard trainer will not stop and will not give any warning if the loss when to 'nan', it will just keep training.
  2. It would be nice to add in the example notebooks 'track_training': True', 'verbose': True for the trainer_config (as its crucially needed for other models experiments)
KonstantinWilleke commented 2 years ago

Hi - yes I agree that (1) is an important feature. Has the validation accuracy been a NaN value?

Same for (2), it'd be a great addition to make it more user-friendly. If you already developed your own tracking, feel free to issue a PR against the current main.

Thanks a lot again for your helpful suggestions.