troyfeng116/hf-trainer

[troyfeng116]

No description provided.

[troyfeng116]

For training:

• hf_trainer.py + eval_helpers.py

For inference:

• hf_infer.py

[niansong1996]

As an example, this is how PyTorchLightning (PL) works:

PL Trainer -> DataModule
-> Dataset
-> PLModule
-> Seq2Seq Model
-> validation_step
-> ...

[troyfeng116]

• See hf_infer.py for HF Trainer basic usage, with documentation

The two ways to inject custom validation behavior into HF Trainer loop I found are 1) compute_metrics and 2) TrainerCallbacks (potentially 3: override methods on transformers.Trainer class directly)

• In eval_helpers, see attempts to use compute_metrics for training (did not get to work)
• In hf_infer and hf_train, see attempt to use TrainerCallback (this works)
• In hf_infer at bottom of file, see attempt to override Trainer class methods (this did not work, but discussed overriding behavior farther up inheritance tree to match with generation)

In general, HF trainer seems to be less well-documented for text generation.