0.4.3 (2024-12-10)

• Allow loss-type to be specified (allowing 0.4.0, 0.4.1 or 0.4.2 to be reproduced).

Full Changelog: 0.4.2...0.4.3

0.4.2 (2024-12-03)

• Lower early stopping threshold
• Set alpha_gdl to 1 to give the same scale as MAE
• Step the optimiser once per set of predictions

0.4.1 (2024-12-02)

• Add gradient difference loss
• Set default GDL weighting to 10
• Fix checkpoint loading bugs

0.4.0 (2024-11-29)

• Better logging of meta information during training
• Use size of input data to construct model
• Do not encode image size into model parameters (for better flexibility)

0.3.5 (2024-11-27)

• Checkpoint at the end of each epoch as well as each batch
• Add early stopping of epoch if improvements are not seen

0.3.4 (2024-11-16)

• Better logging of model parameters
• Checkpointing after validation
• Checkpoint after fixed number of batches (instead of truncating epochs)
• Updated hyperparameter defaults

0.3.3 (2024-11-16)

• Speed up training following profiling

0.3.2 (2024-11-14)

• Use Lightning for training/validation

0.3.1 (2024-11-13)

• Revert changes to run all forecasts in forward() as this requires too much memory
• Construct time-embedding on-the-fly to simplify interface
• General memory saving through tensor deletion
• Protect against input NaNs in prediction which propagated through the predictions

0.3.0 (2024-11-12)

• Perform all forecasting steps before updating loss
• Use same encoder for X and y