- Added unit tests to TrainData helper methods. - Added option to enable/disable progress in PyTorch model training loop - Added abstract base class for TrainCallbacks - Added abstract base class and simple printer for StepCallbacks - Added a callback to be called after each batch to the training loop
0.3.5
- Exponential smoothing per-batch loss also for validation errors - Reset learnable parameters of activation functions - Added more neural-network building blocks
0.3.4
- Exponential smoothing per-batch loss for reporting in training progress
0.3.3
- Updated docstring
0.3.2
- Added progress bar for train (and test) loss evaluation - Add options to always step the optimizer after each batch - Added norm-first option to (repeated) skip-connection blocks - Added Lazy concatenation option for tensors along first dimension.
0.3.1
- Removed old parent directory creation logic for checkpoint file - Added dropout option to activated block - Toggle loss between train and eval in PyTorch training loop - Added cross-entropy loss with label smoothing on and off for train and eval