v0.3.0

Tsunami v0.3.0

Diff since v0.2.0

v0.3.0

Breaking changes:

• fit! returns nothing instead of a FitState object. The FitState object can be accessed via trainer.fit_state.
• on_before_pullback has been removed. Use on_train_batch_start instead.
• on_*_batch_start now receives the batch on device.
• Some of the hooks now take more inputs.

Highlights:

• Now Tsunami uses MLDataDevices.DeviceIterator to wrap dataloaders for more efficient device memory management.

• training_step, validation_step, and test_step can now return a named tuple
  for flexibility. One of the fields of the named tuple should be loss which is used to compute the loss value.

Merged pull requests:

• deprecate return fit_state (#86) (@CarloLucibello)
• more tests for enzyme (#87) (@CarloLucibello)
• bfloat16 initial support (#90) (@CarloLucibello)
• add a flow matching example (#91) (@CarloLucibello)
• changes for v0.3 (#94) (@CarloLucibello)
• update examples + fix callbacks (#96) (@CarloLucibello)

Closed issues:

• support BFloat16 (#69)
• do not return fit_state from fit! (#85)
• use DeviceIterator (#88)
• missing BFloat16 support (#89)
• allow train_step to return a named tuple (#93)