- Refactor using zouqi==1.0.8 to comply with the python type hint. - `elapsed_time` is calculated by default. - YAML configuration for training is now dumped into the run folder.
1.0.7
Changes:
- Add discard_optimizer_state_dict option to discard optimizer if the optimizer state_dict does not match. - Fix model scheduling bugs when at the first step of validation. - Prevent deep copy from schedulers.
1.0.6
- Fix CLI empty value. - Add parser for optional arguments. - Add new command "ls" - Inheritable batch size - Change model train/eval state by default after switch mode - Fix epoch loading. - Support recovery from OOM. - Fix blinking progress bar. - Allow loading checkpoint 0.pth
1.0.5
- Support mode switching so that one can do train and validation in the same run. - Change cli from zouqi to tzq for short. - Fix bugs.