Note

We simplify the training logic, which don't follow config-based and function-based code. We place all codes out of the function except public code blocks.

File Tree

• loss
  • the folder for loss. We only place our used label smoothing loss here.
• models
  • the folder for models. Place your new model file here, then rename the main class as "Model". The runner would automatically load the model with configuration (parameter: args)
• TPM_MES_lara.py
  • the main runner.
    • private block: YOU ONLY NEED TO CHANGE DATASET NAME & MODEL NAME HERE.
    • public block: load yaml config & find available cuda. We only utilize one GPU.
    • public block: build dataset and build loss
    • public block: train/eval controller
• README.md

Acknowledgement

Our code is inspired from:

• STGCN code from @mmaction2
• C2FTCN code from @C2F-TCN
• Metric code from @ASRF Much appreciate for their efforts.

Reference

If you found this repo useful, please cite our arxiv version:

not uploaded yet