search

victoresque / pytorch-template

PyTorch deep learning projects made easy.
MIT License
4.7k stars 1.08k forks source link

Setting 'early_stop: 0' does not disable it #84

Closed Andry-Bal closed 3 years ago

Andry-Bal commented 3 years ago

In the README it is said about early stopping that 'This feature can be turned off by passing 0'. This is, however, will not work. self.early_stop will be set to 0 in: https://github.com/victoresque/pytorch-template/blob/85c55356547245b0718291019dfed89385e75e0f/base/base_trainer.py#L34 and later during training this condition will be true if current epoch was worse than previous: https://github.com/victoresque/pytorch-template/blob/85c55356547245b0718291019dfed89385e75e0f/base/base_trainer.py#L91 leading to the output: Validation performance didn't improve for 0 epochs. Training stops.

SunQpark commented 3 years ago

Thank you for reporting this, @Andry-Bal. I believe simply adding a conditional statement after line 34 will fix this.

self.mnt_best = inf if self.mnt_mode == 'min' else -inf
self.early_stop = cfg_trainer.get('early_stop', inf)
if self.early_stop <= 0:
    self.early_stop = inf

I'll make a PR after doing some test. (or I'll appreciate if you make one)

Sorry for late response on this and other issue(#79) you reported. That's because the version of template I'm currently using for most of my projects is on hydra_DDP branch instead of master branch.

Andry-Bal commented 3 years ago

No problem! Thank you for working on this project, it's really helpful.