Lightning-AI/pytorch-lightning

trainer.test() with given checkpoint logs last epoch instead of checkpoint epoch

Open

#20,052 创建于 2024年7月5日

在 GitHub 查看
 (1 评论) (1 反应) (0 负责人)Python (26,687 star) (3,233 fork)batch import
bughelp wantedrepro needed

描述

Bug description

Testing from a given checkpoint leads to logging the epoch number of the last checkpoint instead of the checkpoint specified:

trainer = Trainer(..., max_epochs=10)
lightning_module = MyLightningModule(...)
datamodule = MyDatamodule()

trainer.fit(lightning_module , datamodule=datamodule)

trainer.test(lightning_module , datamodule=datamodule, ckpt_path="last")     # <-- ok: logs correct epoch and step
ckpt_path="/.../checkpoints/epoch=2-step=396.ckpt"
trainer.test(lightning_module , datamodule=datamodule, ckpt_path=ckpt_path)  # <-- incorrect: logs last epoch and step

The second test logs epoch 10 instead of epoch 2. Similarly, the step number of the second test is incorrect.

What version are you seeing the problem on?

v2.2.1

贡献者指南

技术栈
python
领域
backendobservability
议题类型
bug
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-3 hours
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
fresh
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
Basic knowledge of PyTorch LightningUnderstanding of checkpoint loading
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
70
研究方向
The bug is in the trainer.test() method where logging uses the global epoch/step instead of the checkpoint's epoch/step. Likely the fix is in pytorch lightning/trainer/trainer.py or the logging callback (pytorch lightning/callbacks/model checkpoint.py). Check how ckpt path is used to restore state and ensure that the logger's current epoch and global step are set from the checkpoint. The issue has clear reproduction steps and no assignees, making it a good candidate for a first time contributor.