Turn off validation if val_percent_check=0

[kuynzereb]

As was suggested by @williamFalcon in #536 (comment) val_percent_check=0 should turn off the validation loop. But now it will not work because of

 self.num_val_batches = max(1, self.num_val_batches)

So I suggest to fix it. Moreover I suggest to make more thorough processing of train_percent_check and val_check_interval:

1. We should require all *_percent_check and val_check_interval to be in the range [0.0; 1.0].
2. Final num_val_batches can be equal to 0 that will effectively disable validation.
3. Final num_train_batches and num_test_batches should be at least 1. (See also num_training_batches rounds down, causing 0 batches count #631)
4. Final val_check_interval should be at least 1.
5. The user can try to turn off validation by setting val_check_interval to a big value. Maybe in that case we should print a helpful message that validation can be turned off by setting val_percent_check=0.

Any thoughts?

[awaelchli]

One question, why should *_percent_check be in range [0, 1]. You mean [0, 100]? As the name suggests, it should be a percentage.

[kuynzereb]

Yeah, you are right, the naming is not correct. It is not real percents but just fractions. So now it is assumed that 1. means to use the whole dataset.

[cmpute]

This behavior seems to be absent in the documentation?

[Borda]

@cmpute mind sending a PRs with docs?

[miccio-dk]

Sorry to exhume this old issue, but this feature doesn't seem to work any longer. These are my trainer arguments:

min_epochs: 1
max_epochs: 1
max_steps: 1000
num_sanity_val_steps: 0
overfit_batches: 1
log_every_n_steps: 1
val_percent_check: 0

I'm using this settings to completely avoid the validation dataset, and just overfit a single batch of data over 1000 training steps.

[ananthsub]

@miccio-dk this is a very old issue - I'd recommend creating a new one as the project has changed significantly since this was filed.

you may also be interested in this: #10888