Lucas Sloan

When you say slow training speed do you mean examples per second or epochs to convergence? I'm seeing slightly more examples per second than resnet50, but many more epochs to...

Got it - thanks for the info!

The linked example covers my needs. Thanks!

Using torch.autocast() makes sense to me, but doesn't seem to resolve the issue. The other PR implementing similar functionality seems to be getting around it by not using fp16 weights...

Fixed it by not unwrapping the unet.

Added wandb and tensorboard integration.

I've been testing this change, and it seems like it doesn't actually do the fine tuning. Possibly creating the pipeline overwrites the weights, losing the training progress? I'll do some...

Another feature I'd like to have is the ability to provide multiple prompts. However, when I added `action="append"` to the `--validation_prompt` argument, I get an error on this line: `accelerator.init_trackers("text2image-fine-tune",...

@sayakpaul if you could add the methods to EMAModel, that'd be great.

Figured out why the model wasn't training - I was using the `--enable_xformers_memory_efficient_attention` flag. Is that known not to work or do I have an issue with my set up?