* Refactor some binarizer modules
* Training framework
* Acoustic model training
* Restore config values
* Disable `shuffle_batches`
* Support saving weights only
* Fix metric candidates
* Support console color and wrap empty objects
* Support specifying log dir
* `dask.compute` everything at once to improve perf
* Support nested schedulers
* Support auto resuming from latest checkpoint
* Remove some default values
* Check for weights_only
* Support more flexible optimizer settings
* Fix stuck in DDP (probably)
* Use default monitor candidates
* Support `ReduceLROnPlateau` scheduler
* Require sync with validation for ReduceLROnPlateau
* Print metrics
* Add suggestion to change coverage check option
* Support fine-tuning and parameter freezing
* Move registries
* Edit message
* Fix high memory usage and epoch syncing
* Add `build_xxx_dataset` methods
* Simplify augmentation index
* Fix typo: compact -> compat
* Add rank in file pattern to avoid conflict
* `torch.load` with weights_only=True
* Change to rank_zero_info
* sync_dist=True
* Fix `rank_zero_only.rank` needs to be set before use
* Try to optimize message
* Try to optimize message
* Try to optimize message
* Fix config check failure
* Fix overlapping points on TensorBoard when accumulate_grad_batches > 1
* Support EMA (experimental)
* Rename file
* Rename embedding
* Add comments and type hints
* Clean up and re-organize code
* Rename file
* Rename `used` to `enabled`
* Remove redundant wrapper method
* Fix voicing extraction
* Add augmentation flag and check
* Update tensorboard logging
yqzhishen
9f86135c3c