Fix train_text_to_image_qat.py

[AlexanderDokuchaev]

Changes

• Fix incorrect format of openvino/stable-diffusion/requirements.txt
• Change argument name of Accelerator from logging_dir to project_dir, logging_dir deprecated since 0.18
• Remove second call accelerator.prepare for unet

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[AlexKoff88]

As far as I understood model export gets broken after this change so it needs to be fixed before merging this.

[fxmarty]

Thank you!

I can't test those changes as I get an error at this point

optimum-intel/examples/openvino/stable-diffusion/train_text_to_image_qat.py

Line 648 in 0a80e20

for _, batch in tqdm(zip(range(args.opt_init_steps), itertools.islice(dataloader, 0, args.opt_init_steps))):

in the script, telling me that the collate_fn from the dataloader is not pickable. Have you had the issue @AlexKoff88 ?

[fxmarty]

Is there anything blocking for pytorch 2.0?

[AlexanderDokuchaev]

Did not check it.

[AlexKoff88]

Thank you!

I can't test those changes as I get an error at this point

optimum-intel/examples/openvino/stable-diffusion/train_text_to_image_qat.py

Line 648 in 0a80e20

for _, batch in tqdm(zip(range(args.opt_init_steps), itertools.islice(dataloader, 0, args.opt_init_steps))):

in the script, telling me that the collate_fn from the dataloader is not pickable. Have you had the issue @AlexKoff88 ?

I am looking into it. There are multiple issues introduced by updates in transformers and accelerate libraries. Need to highly refactor the script.

[AlexKoff88]

Some issues in this PR should be fixed after merging this PR: #397