Spaces:
Sleeping
Sleeping
| """Training — losses, schedules, callbacks, and the trainer. | |
| The notebook computes loss + masked accuracy inside the model's ``train_step``; | |
| we keep that structure for parity but expose the loss function and callbacks | |
| as standalone modules so they can be unit-tested and reused. | |
| losses.py ``masked_sparse_categorical_crossentropy`` (baseline) + | |
| ``label_smoothed_crossentropy`` + ``build_loss`` | |
| schedules.py ``WarmupCosineDecay`` + ``build_learning_rate`` | |
| callbacks.py ``default_callbacks(config)`` — early stopping + checkpoint | |
| trainer.py ``Trainer.fit()`` — wraps compile + fit + history serialization | |
| """ | |
| from captioning.training.callbacks import default_callbacks | |
| from captioning.training.losses import ( | |
| build_loss, | |
| label_smoothed_crossentropy, | |
| masked_sparse_categorical_crossentropy, | |
| ) | |
| from captioning.training.schedules import WarmupCosineDecay, build_learning_rate | |
| from captioning.training.trainer import Trainer | |
| __all__ = [ | |
| "Trainer", | |
| "WarmupCosineDecay", | |
| "build_learning_rate", | |
| "build_loss", | |
| "default_callbacks", | |
| "label_smoothed_crossentropy", | |
| "masked_sparse_categorical_crossentropy", | |
| ] | |