Failed to configure_optimizers from dictionary without lr_scheduler field presented

[alexeykarnachev]

🐛 Bug

Optimizer is failed to be configured from the dictionary without lr_sheduler field.
Consider an example of the Module configure_optimizers method:

        def configure_optimizers(self):
            config = {
                'optimizer': torch.optim.SGD(params=self.parameters(), lr=1e-03)
            }
            return config

Then, we run a simple trainer:

    trainer_options = dict(default_save_path=tmpdir, max_epochs=1)
    trainer = Trainer(**trainer_options)
    _ = trainer.fit(model)

And we fail with an error:

UnboundLocalError: local variable 'lr_schedulers' referenced before assignment

I believe, that the reason is that lr_schedulers local variable is not determined here:
https://github.com/PyTorchLightning/pytorch-lightning/blob/8dd9b80d7a192117783195a748ddce6c33d556f3/pytorch_lightning/trainer/optimizers.py#L36-L42

I think, it could be fixed like this:

        # single dictionary
        elif isinstance(optim_conf, dict):
            optimizer = optim_conf["optimizer"]
            lr_scheduler = optim_conf.get("lr_scheduler", [])
            if lr_scheduler:
                lr_schedulers = self.configure_schedulers([lr_scheduler])
            else:
                lr_schedulers = []
            return [optimizer], lr_schedulers, []

To Reproduce

Steps to reproduce the behavior:

1. Create a simple Module with configure_optimizers which looks like above.
2. Run the fit Trainer method with the model.
3. See error

Code sample

https://gist.github.com/alexeykarnachev/c61a5b1ca3bf876e19b4547eeb9f42dc

Expected behavior

I suppose, that such the configuration: {"optimizer": ...}, without "lr_scheduler" must be a valid one, and this error must not be occurred.

Environment

OS: Linux
architecture: 64bit
processor: x86_64
python: 3.7.6
version: #97~16.04.1-Ubuntu SMP Wed Apr 1 03:03:31 UTC 2020

pytorch-lightning: 0.7.3rc1

[williamFalcon]

yeah, agreed that dict should work without the scheduler. mind submitting a PR?