- Implement
StateSerializable
interface for training loops and training steps to support training loop checkpoint. - Implement evaluation loop.
- Support mixed-precision training.
- Improve logging.
v0.1.3 - 2023-12-22
- Not passing metrics to SimpleTrainingStep won't cause initialization error. (#26)
- Displaying scientific notation when metrics values becoming too large or too small. (#29)
v0.1.2 - 2023-11-27
- Support dataloaders without
__len__
implementation. (#21)
- Different verbose levels for training loops'
fit()
function. (#20)
v0.1.1 - 2023-11-11
- Implement distributed training loop. (#12)
v0.1.0 - 2023-11-03
TrainingLoop
instances can be initialized directly.
They receives an instance of TrainingStep
specifying the logic of training/validating a model.
- Breaking: Remove
SimpleTrainingLoop
, replace it withSimpleTrainingStep
. (#11) - Remove
tensorboard
dependency.
- Add usage section in README.
- Add an example for training conditional GAN on Fashion MNIST with custom loop.
v0.0.3 - 2023-10-25
- References to both training loop and model instances in callbacks.
v0.0.2 - 2023-10-24
- Update progress bar after each epoch finishes.
- Cuda training not working.
v0.0.1 - 2023-10-24
Initial Release.