Add support for ReduceLROnPlateau about lightning

[LarryThermo]

Would you consider providing (or point me to an example) a simple but complete example of using ReduceLROnPlateau. Thanks, Lars

[Ir1d]

Also, at this moment https://github.com/williamFalcon/pytorch-lightning/blob/master/pytorch_lightning/trainer/trainer.py#L958 lr_scheduler.step() is called at the end of the epoch, while some scheduler (eg: https://pytorch.org/docs/master/optim.html#torch.optim.lr_scheduler.OneCycleLR) should be called at the end of the batch

[williamFalcon]

@terkelbo good suggestion. can you propose a clean way of doing this? maybe it can be merged with #29

[Ir1d]

I think we can directly adjust lr in the optimizer as keras, which means we don't need ReduceLROnPlateau as a metric.

Specifically, perhaps we can consider adding a hook in pytorch_lightning/root_module/root_module.py like optimizer_step, so that this function can be exposed to callbacks?

[williamFalcon]

@lr1d good point. you can just adjust the LR in the current callback also (optimizer_step). or in any of the other callbacks

[williamFalcon]

closing because this should be implicitly supported since we can pass a ReduceLROnPlateau object... nothing we need to do, this is standard PyTorch functionality

[Ir1d]

When using the pytorch object, it requires to pass the metric as a param. for exmaple lr_scheduler.step(val_loss). How do you overwrite the origianl schedulers in thie way?

[stolpa4]

Please, I can't follow. Why the issue is closed? Do you suggest using some hooks like optimizer_step and implement reducing on plateau scheduling by hand?

[terkelbo]

I’m not sure either how to pass the ReduceLROnPlateau object as it needs the metric argument as pointed out by @Ir1d. @williamFalcon would it be possible that you give an example of how to use this scheduler with pytorch lightning?

[kvhooreb]

It feels rather dirty to me, but you can save the loss in your pl.LightningModule's training_step() method, and then use it in the optimizer_step method. I can't verify whether it works as expected right now though.

def training_step(self, batch, batch_nb):
        # REQUIRED
        x, y = batch
        y_hat = self.forward(x)
        self.loss = F.mse_loss(y, y_hat)
        return {'loss': self.loss}

def configure_optimizers(self):
        # REQUIRED
        # can return multiple optimizers and learning_rate schedulers
        self.optimizer = torch.optim.Adam(self.parameters(), lr=self.lrinit)
        self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, mode='min', 
                                                           factor=0.1, patience=10, 
                                                           verbose=False, threshold=0.0001, 
                                                           threshold_mode='rel', cooldown=0, 
                                                           min_lr=0, eps=1e-08)
        return self.optimizer

def optimizer_step(self, epoch_nb, batch_nb, optimizer, optimizer_i):
        """
        Do something instead of the standard optimizer behavior
        :param epoch_nb:
        :param batch_nb:
        :param optimizer:
        :param optimizer_i:
        :return:
        """
        # Sometimes (if closure is not None), this step method should return the loss. Now its None
        loss = optimizer.step()
        # So let's use self.loss, set in training_step, for LR scheduling
        self.scheduler.step(self.loss)
        # clear gradients
        optimizer.zero_grad()

Edit: This will not work as intended, since the optimizer step is called after every batch.

[williamFalcon]

we could modify the scheduler step to take in the loss when it needs it? i have to look at this more carefully though

[kvhooreb]

That would be the best solution indeed, but then you'd need a way to figure out which ones need the loss when calling step. I'm not sure how to do that at this moment.

For now, I solved it as follows, which may help people until there is a better solution:

• Use the on_epoch_start callback on your LightningModule to initialize an empty list
• In every training_step call, append the loss to the list
• In the on_epoch_end callback, process the list to get the averga loss, call scheduler.step(mean_loss) and clear the list

[williamFalcon]

From PT docs.
https://pytorch.org/docs/stable/optim.html

Prior to PyTorch 1.1.0, the learning rate scheduler was expected to be called before the optimizer’s update; 1.1.0 changed this behavior in a BC-breaking way. If you use the learning rate scheduler (calling scheduler.step()) before the optimizer’s update (calling optimizer.step()), this will skip the first value of the learning rate schedule. If you are unable to reproduce results after upgrading to PyTorch 1.1.0, please check if you are calling scheduler.step() at the wrong time.

[williamFalcon]

Either way, @kvhooreb i added the .step(epoch) fix. Let me know if this works for you.

[shawwn]

Yes, a simple example would be great please.