callbacks.py

import logging
import os

from pytorch_lightning.callbacks import ModelCheckpoint


logger = logging.getLogger(__name__)


def get_checkpoint_callback(output_dir, metric):
    """Saves the best model by validation EM score."""
    if metric == "rouge2":
        exp = "{val_avg_rouge2:.4f}-{step_count}"
    elif metric == "bleu":
        exp = "{val_avg_bleu:.4f}-{step_count}"
    elif metric == "em":
        exp = "{val_avg_em:.4f}-{step_count}"
    else:
        raise NotImplementedError(
            f"seq2seq callbacks only support rouge2 and bleu, got {metric}, You can make your own by adding to this function."
        )

    checkpoint_callback = ModelCheckpoint(
        filepath=os.path.join(output_dir, exp),
        monitor=f"val_{metric}",
        mode="max",
        save_top_k=3,
        period=0,  # maybe save a checkpoint every time val is run, not just end of epoch.
    )
    return checkpoint_callback