/pytorch-transformers-template

This is a template for fast prototype relying on transformers, hydra, fairscale and deepspeed, etc.

Primary LanguagePythonMIT LicenseMIT

pytorch-transformers-template

This is a template for fast prototype relying on transformers, hydra, fairscale and deepspeed, etc.

Update Notes

2023/02/11

  • Unify torch vanilla FSDP training with fairscale FSDP into single trainer.
  • Add some post processors as example for reference.
  • Add a Trie based logits processor as example.
  • Optimize the evaluator.
  • Add trainer supporting multiple training dataset.
  • Add support to different learning schedulers.

TODO List

  • Multiple training set.
  • Use wandb to replace naive tensorboard.
  • Resume training from checkpoint.
    • Different setup for different engines, e.g., fairscale and deepspeed, or vanilla pytorch.