Learning rate stepping option

[SkafteNicki]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
• Did you read the contributor guideline?
• Did you make sure to update the docs?
• Did you write any new necessary tests?
• If you made a notable change (that affects users), did you update the CHANGELOG?

What does this PR do?

Fixes #806
Fixes part of #827
Fixes #945

This PR will allow the user to choose the frequency at which the learning rate schedulers are called by supplying it as:

def configure_optimizer(self):
    optimizer = ...
    scheduler = ...
    return [optimizer], [[scheduler, freq]]

this will call scheduler.step() every freq training step. The user can still just write

    return [optimizer], [scheduler]

which (like now) will call scheduler.step() after every epoch. This way the changes should be backward compatible. This may not be the most intuitive way to do it and I am happy to discuss another way of doing it.

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

[williamFalcon]

@SkafteNicki awesome haha. Can you replace it with a dict instead of array? (#945)

Fixes #945

[williamFalcon]

@SkafteNicki mind rebase as well? :)

[SkafteNicki]

@williamFalcon will do the rebase and change structure to a dict within the next day or so. You are right that for future features it is probably best to have it in a dict like structure. I propose to have three fields:
{'scheduler': LRScheduler, 'interval': 'batch|epoch'`, 'frequency': integer}
where 'interval' is as default set to 'epoch' and 'frequency' is set to 1. Should probably still allow user to just input list of LRSchedulers for backward compatible.

[williamFalcon]

cool. do we need frequency?
@srush @ethanwharris

[ethanwharris]

This looks nice :) Personally I would opt for a batch / epoch option as suggested by @williamFalcon in #945

There may be some value to a frequency argument for iterable stuff (where there's no real notion of an epoch) - but perhaps the options could be 'epoch', 'batch', or int?

It should also work smoothly with the accumulate grads stuff, i.e. 'batch' should really mean call to .step

[srush]

Yeah we just need "step". That's become the standard way to do it. If you need "every x steps" that should be in the scheduler El mar., 25 de febrero de 2020 15:54, Ethan Harris <notifications@github.com> escribió:

…

This looks nice :) Personally I would opt for a batch / epoch option as suggested by @williamFalcon <https://github.com/williamFalcon> in #945 <#945> There may be some value to a frequency argument for iterable stuff (where there's no real notion of an epoch) - but perhaps the options could be 'epoch', 'batch', or int? It should also work smoothly with the accumulate grads stuff, i.e. 'batch' should really mean call to .step — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#941>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAIYKXE6NT6ZGS7YR5HUR3REWAQPANCNFSM4K3M2GYQ> .

[pep8speaks]

Hello @SkafteNicki! Thanks for updating this PR.

• In the file pytorch_lightning/core/lightning.py:

Line 786:101: E501 line too long (119 > 100 characters)

• In the file pytorch_lightning/trainer/training_io.py:

Line 34:101: E501 line too long (103 > 100 characters)
Line 35:101: E501 line too long (101 > 100 characters)
Line 37:101: E501 line too long (109 > 100 characters)
Line 70:101: E501 line too long (109 > 100 characters)

• In the file pytorch_lightning/trainer/training_loop.py:

Line 725:101: E501 line too long (101 > 100 characters)

• In the file tests/models/utils.py:

Line 85:101: E501 line too long (116 > 100 characters)

• In the file tests/trainer/test_optimizers.py:

Line 48:101: E501 line too long (110 > 100 characters)
Line 88:101: E501 line too long (114 > 100 characters)

Comment last updated at 2020-03-05 08:42:37 UTC

[Borda]

it becomes quite long... what about LR move to separate mixin, also tests to test_trainer_lr

[williamFalcon]

@SkafteNicki awesome! Let's rebase so we can get this merged :)

[williamFalcon]

@SkafteNicki fix tests?

[SkafteNicki]

@williamFalcon the failing test seems to be unrelated to the PR, tests were passing before the latest merge

[ethanwharris]

Ok, so merging master has caused some issues here, the callback test in test_trainer.py now exists in tests/trainer/test_callbacks.py

@SkafteNicki - the simplest thing to do would be to have a look at the diff and just remove the bits that weren't you - or revert the merge and do a rebase

[williamFalcon]

@SkafteNicki
awesome! very close.

have some failing GPU tests.

_____________________________________________________________________________________________ test_amp_gpu_ddp _____________________________________________________________________________________________

tmpdir = local('/tmp/pytest-of-waf251/pytest-45/test_amp_gpu_ddp0')

    def test_amp_gpu_ddp(tmpdir):
        """Make sure DDP + AMP work."""
        if not tutils.can_run_gpu_test():
            return

        tutils.reset_seed()
        tutils.set_random_master_port()

        hparams = tutils.get_hparams()
        model = LightningTestModel(hparams)

        trainer_options = dict(
            default_save_path=tmpdir,
            show_progress_bar=True,
            max_epochs=1,
            gpus=2,
            distributed_backend='ddp',
            precision=16
        )

>       tutils.run_model_test(trainer_options, model)

tests/test_amp.py:81:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
tests/models/utils.py:88: in run_model_test
    trainer.hpc_save(save_dir, logger)
pytorch_lightning/trainer/training_io.py:431: in hpc_save
    checkpoint = self.dump_checkpoint()
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <pytorch_lightning.trainer.trainer.Trainer object at 0x7f677f977bd0>

    def dump_checkpoint(self):
        checkpoint = {
            'epoch': self.current_epoch + 1,
            'global_step': self.global_step + 1,
        }

        if self.checkpoint_callback is not None and self.checkpoint_callback is not False:
            checkpoint['checkpoint_callback_best'] = self.checkpoint_callback.best

        if self.early_stop_callback is not None and self.checkpoint_callback is not False:
            checkpoint['early_stop_callback_wait'] = self.early_stop_callback.wait
            checkpoint['early_stop_callback_patience'] = self.early_stop_callback.patience

        # save optimizers
        optimizer_states = []
        for i, optimizer in enumerate(self.optimizers):
            optimizer_states.append(optimizer.state_dict())

        checkpoint['optimizer_states'] = optimizer_states

        # save lr schedulers
        lr_schedulers = []
        for scheduler in self.lr_schedulers:
>           lr_schedulers.append(scheduler['scheduler'].state_dict())
E           TypeError: 'CosineAnnealingLR' object is not subscriptable

[williamFalcon]

@SkafteNicki any chance you can finish this in the next few hours so we can merge and release?

[SkafteNicki]

@williamFalcon, sorry for the delay, been really busy the last couple of days.
The latest change should fix the gpu test.

The failing checks all seems to be course by (nothing to do with this PR)
urllib.error.HTTPError: HTTP Error 403: Forbidden
which has something to do with MNIST downloading I guess

[Borda]

The failing checks all seems to be course by (nothing to do with this PR)
urllib.error.HTTPError: HTTP Error 403: Forbidden
which has something to do with MNIST downloading I guess

Yeah, it is happening from yesterday...

[SkafteNicki]

@Borda anything I can do?

[Borda]

can we make the monitoring "val_loss" as a parameter?

[Borda]

@Borda anything I can do?

I guess it is fine (maybe check las Wills comment)... Great job again!
We will try to do something around the downloading datasets... :/

[SkafteNicki]

@Borda I have changed code such that users can also pass a monitor key, which ReduceLROnPlauteau schedulers are then conditioned on. As default it is set to val_loss.

[williamFalcon]

______________________________________________________________________________________ test_optimizer_return_options _______________________________________________________________________________________

    def test_optimizer_return_options():
        tutils.reset_seed()

        trainer = Trainer()
        model, hparams = tutils.get_model()

        # single optimizer
        opt_a = torch.optim.Adam(model.parameters(), lr=0.002)
        opt_b = torch.optim.SGD(model.parameters(), lr=0.002)
        optim, lr_sched = trainer.init_optimizers(opt_a)
        assert len(optim) == 1 and len(lr_sched) == 0

        # opt tuple
        opts = (opt_a, opt_b)
        optim, lr_sched = trainer.init_optimizers(opts)
        assert len(optim) == 2 and optim[0] == opts[0] and optim[1] == opts[1]
        assert len(lr_sched) == 0

        # opt list
        opts = [opt_a, opt_b]
        optim, lr_sched = trainer.init_optimizers(opts)
        assert len(optim) == 2 and optim[0] == opts[0] and optim[1] == opts[1]
        assert len(lr_sched) == 0

        # opt tuple of lists
        scheduler = torch.optim.lr_scheduler.StepLR(opt_a, 10)
        opts = ([opt_a], [scheduler])
        optim, lr_sched = trainer.init_optimizers(opts)
        assert len(optim) == 1 and len(lr_sched) == 1
>       assert optim[0] == opts[0][0] and \
            lr_sched[0] == dict(scheduler=scheduler, interval='epoch',
                                frequency=1, reduce_on_plateau=False)
E       AssertionError: assert (Adam (\nParame...ght_decay: 0\n) == Adam (\nParame...ght_decay: 0\n)
E         -Adam (\n
E         -Parameter Group 0\n
E         -    amsgrad: False\n
E         -    betas: (0.9, 0.999)\n
E         -    eps: 1e-08\n
E         -    initial_lr: 0.002\n
E         -    lr: 0.002\n...
E
E         ...Full output truncated (12 lines hidden), use '-vv' to show and {'frequency':...': False, ...} == {'frequency':...7f1c2c0acc90>}
E         Omitting 4 identical items, use -vv to show
E         Left contains 1 more item:
E         {'monitor': 'val_loss'}
E         Full diff:
E           {
E            'frequency': 1,
E            'interval': 'epoch',...
E
E         ...Full output truncated (5 lines hidden), use '-vv' to show)

tests/test_gpu_models.py:123: AssertionError

[SkafteNicki]

@williamFalcon fixed the failing test, so all checks are passing now