torch-training-loop

Fixed

• Not passing metrics to SimpleTrainingStep won't cause initialization error. (#26)
• Displaying scientific notation when metrics values becoming too large or too small. (#29)

Changed

• Support dataloaders without __len__ implementation. (#21)

Added

• Different verbose levels for training loops' fit() function. (#20)

Added

• Implement distributed training loop. (#12)

TrainingLoop instances can be initialized directly.
They receive an instance of TrainingStep specifying the logic of training/validating a model.

Removed

• Breaking: Remove SimpleTrainingLoop, replace it with SimpleTrainingStep.
  (#11)
• Remove tensorboard dependency.

Added

• Add usage section in README.
• Add an example for training conditional GAN on Fashion MNIST with custom loop.

Callbacks will now have references to both training loop and model instances.

• Fix: Cuda training not working;
• Add: Update progress bar after an epoch finishes.

Initial release.