v1.2.0
List of changes:
- Update template for compatibility with lightning v1.5 and pytorch v1.10
- General documentation improvements
- Move
LICENSE
toREADME.md
- Add manual resetting of metrics at the end of every epoch, to make sure no one makes hard to spot calculation mistakes
- Add experiment mode to all experiment configs
- Improve logging paths for experiment mode
- Add
MaxMetric
to model, for computation of best so far validation accuracy - Add
RichProgressBar
to default callbacks for the pretty formatted progress bar - Get rid of the trick for preventing auto hparam logging, since lightning now supports it with
self.save_hyperparameters(logger=False)
- Add
self.save_hyperparameters()
to datamodule since lightinng now supports it - Deprecate Apex support since native pytorch mixed-precision is better
- Deprecate bash script for conda setup since installation commands change too often to maintain it
- Change
trainer.terminate_on_nan
debug option totrainer.detect_anomaly
for compatibility with lightning v1.5 - Specify model and datamodule during
trainer.test()
, for compatibility with lightning v1.5 - Remove
configs/trainer/all_params.yaml
- Make hyperparameter optimization compatible with lightning v1.5
- Specify that EarlyStopping patience is counted in validation epochs and not in training epochs.
- Add a new way for accessing datamodule attributes to the
README.md
- Make debug mode automatically set the level of all command-line loggers to
DEBUG
- Make debug mode automatically set the trainer config to
debug.yaml
- Add generator seed to prevent test data leaking to train data in
datamodule.setup()
when seed is not set up - Move Dockerfile to
dockerfiles
branch - Modifiy
configs/trainer/debug.yaml
to enable some debug options - Remove unused
if config.get("debug"):
inextras
Special thanks for PRs to: @CharlesGaydon, @eungbean, @gscriva