Why no 'unet.train()' after log_validation in train_text_to_image.py

[wikiwen]

In diffusers/examples/text_to_image/train_text_to_image.pyline

As model will be "eval()" default in log_validation, why no unet.train() to restore after log_validation

[tolgacangoz]

Hi @wikiwen,
I guess log_validation() creates a new and independent pipeline by using accelerator.unwrap_model() and then removes it at the end. So, as I understand, it doesn't intervene previous stack's unet; in other words, there is no need to do unet.train() after log_validation() because there was no unet.eval() in the first place.

[wikiwen]

Thanks for your reply, May be