last.ckpt is a soft link of best.ckpt

[Jinbo-Hu]

Bug description

When I save the last.ckpt, the last.ckpt is symlink of the best.ckpt. I have already set model_checkpoint.save_last=True.

What version are you seeing the problem on?

v2.1, master

How to reproduce the bug

No response

Error messages and logs

# Error messages and logs here please

Environment

Current environment

#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow):
#- PyTorch Lightning Version (e.g., 1.5.0):
#- Lightning App Version (e.g., 0.5.2):
#- PyTorch Version (e.g., 2.0):
#- Python version (e.g., 3.9):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):
#- Running environment of LightningApp (e.g. local, cloud):

More info

No response

cc @carmocca @awaelchli

[awaelchli]

The was a new feature in 2.1, not a bug. You can find some discussion here: #18995
Sorry if it came unexpected.

[Jinbo-Hu]

But how do I save the latest or newest epoch? I just upgraded the version from 2.0.4 to 2.1.3.

[Jinbo-Hu]

But how do I save the latest or newest epoch? I just upgraded the version from 2.0.4 to 2.1.3.

automatically

[Eleven1Liu]

In my case, when I set model_checkpoint.save_last=True, the last checkpoint links to best_model.ckpt.
However, best_model.ckpt is different from last.ckpt. In my understanding, best_model.ckpt is the best model you get by validation metric. last.ckpt is the model of the last epoch. They are not the same.

The second error we got was when we use

trainer.fit(model, train_loader)

without val_loader (training for a fixed number of epochs without validation)

We got OSError: [Error 40] Too many levels of symbolic links: '...' because last.ckpt is linked to last.ckpt.
I think this PR is helpful for us.

Or should we wait for the latest release of the following code:

ModelCheckpoint(save_last='copy')

We've upgraded our lightning version to 2.1.3 by pip install but still got the symbolic link version.
Thanks!

[bfs18]

It is note fixed in 2.1.4