mmaction

This code is based on MMAction which supports modular design and high efficiency. Our TPN would be merged into the latest MMAction in the future.

Here we briefly introduce the structure of this codebase:

• apis: contains the launcher of the whole codebase and intializer of distributed training environment.
• core: contains multiple hooks for evaluation e.g. calculating the Top-1/Top-5 accuracy.
• datasets: contains rawframes_dataset and transform for training.
• losses: contains kinds of CrossEntropy loss.
• models: contains recognizers and various submodules of network e.g. backbone, neck,and head under models/tenons

Such modular design helps us quickly and easily conduct experiments with different modules.