docs

Lightning Module Enhanced documentation

Making pytorch-lightning more pythonic, whatever that means.

Wrapper on top of Lightning Module.

Some features:

• Adds some common patterns, such as creating attributes for optimizer, scheduler, metric, criterion, device.
• Adds model.summary() via torchinfo.
• Adding np_forward to pass numpy data and relatively automatically getting the correct device for a forward pass.
• Adds model.model_algorithm method for manual low level callback for a forward/loss/metrics pass.
• Adds MultiTrainer that wraps lightning's Trainer to train the same model N times and then seamlessly pick the best varaints. Used to minimize variance of training.

Examples:

• See model algorithm for how to use a model algorithm callback, for more complicated train or test semantics, that aren't a simple feed forward.
• See multi trainer test for MultiTrainer code.