What is the difference between callbacks.model_checkpoint and callbacks.metrics_over_trainsteps_checkpoint?

[motcapbovit]

Hi, it's me again. Thanks for your earlier fast response! As I have been fine-tuning interpolation model since you released the code, I found it hard to understand the main idea of callbacks.model_checkpoint and callbacks.metrics_over_trainsteps_checkpoint in your code.

I observe that there are distinct specified steps for each, and they're all seemed to be a model checkpoint. So just want to ask you what is the difference between them, and if I change the every_n_train_steps what could happen?

Many thanks!

[Doubiiu]

Hi. The checkpoint saved in model_checkpoint will be replaced after each every_n_train_steps (thus only one checkpoint will be left), while that in metrics_over_trainsteps_checkpoint will be saved for every every_n_train_steps.

[motcapbovit]

Oh! So, if I load the model from the model_checkpoint or the metrics_over_train_steps checkpoint, it's the same, right?

[Doubiiu]

yeap