如何在PyTorch Lightning中记录每次保存的Checkpoint路径

方法1：自定义带日志的Checkpoint回调 #

直接继承ModelCheckpoint，重写核心保存方法，在Checkpoint完成保存后立即记录路径：

from pytorch_lightning.callbacks import ModelCheckpoint
import logging

logger = logging.getLogger(__name__)

class LoggingModelCheckpoint(ModelCheckpoint):
    def _save_checkpoint(self, trainer, filepath):
        # 先执行父类的保存逻辑
        super()._save_checkpoint(trainer, filepath)
        # 记录Checkpoint完整路径
        logger.info(f"Checkpoint saved to: {filepath}")

# 替换原Checkpoint回调
checkpoint_callback = LoggingModelCheckpoint(
    filename="fa_classifier_{epoch:02d}",
    every_n_epochs=val_every_n_epochs,
    save_top_k=-1,
)

# Trainer初始化保持不变
trainer = Trainer(
    callbacks=[checkpoint_callback],
    default_root_dir=checkpoints_path,
    check_val_every_n_epoch=val_every_n_epochs,
    max_epochs=max_epochs,
    gpus=1
)

方法2：新增独立回调读取已保存路径 #

如果不想修改原Checkpoint逻辑，可以添加一个单独的回调，在验证阶段结束后读取已保存的Checkpoint路径：

from pytorch_lightning.callbacks import Callback, ModelCheckpoint
import logging

logger = logging.getLogger(__name__)

class CheckpointPathLogger(Callback):
    def on_validation_end(self, trainer, pl_module):
        # 找到Trainer中的Checkpoint回调实例
        checkpoint_cb = next(cb for cb in trainer.callbacks if isinstance(cb, ModelCheckpoint))
        # 读取最新保存的Checkpoint路径
        if checkpoint_cb.saved_checkpoints:
            latest_path = checkpoint_cb.saved_checkpoints[-1]
            logger.info(f"Latest checkpoint saved to: {latest_path}")

# 初始化原Checkpoint回调和日志回调
checkpoint_callback = ModelCheckpoint(
    filename="fa_classifier_{epoch:02d}",
    every_n_epochs=val_every_n_epochs,
    save_top_k=-1,
)
path_logger = CheckpointPathLogger()

# 同时传入两个回调到Trainer
trainer = Trainer(
    callbacks=[checkpoint_callback, path_logger],
    default_root_dir=checkpoints_path,
    check_val_every_n_epoch=val_every_n_epochs,
    max_epochs=max_epochs,
    gpus=1
)

说明 #

• 方法1的时机最精准，在Checkpoint完成写入后立即触发日志，不会有延迟。
• 方法2通过监听验证结束事件，利用ModelCheckpoint自带的saved_checkpoints属性获取历史保存路径，适合无需改动原Checkpoint逻辑的场景。

内容的提问来源于stack exchange，提问作者Gulzar