New Features
- Add ONNX support for adaptive average pooling
- Add TINShift op (used in MMAction2)
- Add DepthwiseSeparableConvModule
- Support PyTorch 1.6
Improvements
- Allow log_config to be None when registering logger hooks
- Support unserializable values for JsonHandler
- Add a trouble shooting page for common issues related to installation and usage
- Change the model url from AWS to Aliyun
- Close the config file after opening to avoid warnings
Bug fixes
- Fix MMDataParallel in CPU mode
- Fix the val mode in logging
- Fix the API incompatibility of SAConv in PyTorch 1.6
- Fix PaviLoggerHook
- Fix the local rank when using slurm
- Fix the `set_epoch()` for customized BatchSampler
- Fix the parameter order mismatch of SyncBN
- Fix the logger config for IterBaseRunner
- Fix the typo for yaml file extension check in Config