Support New Methods and Fix Bugs (V0.2.4)

[Lupin1998]

Supporting new methods:

1. Support new backbone architectures (LITv2).
2. Refactor code structures weight initialization in various network modules (using BaseModule in mmcv).
3. Refactor code structures of openmixup.models.utils.layers to support more network structures.

Fixing bugs:

1. Fix bugs that cause degenerate performances of pure Transformer backbones (DeiT and Swin) in OpenMixup. The main reason might be the old version of auto_fp16 and DistOptimizerHook implementations, since PyTorch=>1.6.0 has better support of fp16 training than mmcv.
2. Fix the bug of ViT fine-tuning for MIM methods (e.g., MAE, SimMIM). The original MIMVisionTransformer in openmixup.models.mim_vit has frozen all the backbone parameters during fine-tuning.
3. Fix the weight initialization of Transformer-based architectures (e.g., ViT, Swin) to reproduce the train-from-scratch performance. Update weight initialization and fp16 training settings in relevant config files.

[Lupin1998]

OpenMixup v0.2.4 is now released, which supports various Transformer architectures and fixes bugs.