Summary of issues related to Diffusion program debugging

As a powerful generative model, the Diffusion Model has "invaded" various low-level visual tasks. The existing Diffusion Models are all implemented based on Stable Diffusion, and some problems will be encountered when reproducing them. This blog post summarizes some bugs I encountered during the reappearance process and their solutions. Hope that helps.

1. from pytorch_lightning.utilities.distributed import rank_zero_only

Detailed description of the problem:

from pytorch_lightning.utilities.distributed import rank_zero_only
ModuleNotFoundError: No module named 'pytorch_lightning.utilities.distributed'

Solution: put

from pytorch_lightning.utilities.distributed import rank_zero_only

change into:

from pytorch_lightning.utilities.rank_zero  import rank_zero_only

2. parser = Trainer.add_argparse_args(parser)

Detailed description of the problem:

parser = Trainer.add_argparse_args(parser)
AttributeError: type object 'Trainer' has no attribute 'add_argparse_args'

Reason: Because pytorch-lightening 2.x is not compatible with 1.x.

Solution: Downgrade pytorch-lightning , you can try

pip install pytorch-lightning==1.9.4