How can I get attention maps from transformer?

[amoeba04]

Thank you for releasing your great work!

I'm trying to follow your work, but I face an error below:

Traceback (most recent call last):
  File "train_flux_lora.py", line 467, in <module>
    finetune(args)
  File "train_flux_lora.py", line 403, in finetune
    upper_loss, t_enc_ddpm = calculate_upper_loss(args, batch, compute_text_embeddings, text_encoders, tokenizers, transformer, noise_scheduler_copy, batch["prompts"], vae, criteria, negative_guidance=args.negative_guidance, weight_dtype=weight_dtype, neg_prompts=str(args.prompt_b), start_guidance=3, ddim_steps=28, lamb1=float(args.lamb1), lamb2=float(args.lamb2))
  File "EraseAnything/utils/calc_loss.py", line 374, in calculate_upper_loss
    z, latent_image_ids = latent_sample(transformer,
  File "miniconda3/envs/sd3/lib/python3.8/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "EraseAnything/utils/esd_utils.py", line 91, in latent_sample
    noise_pred, attn_maps = transformer(
ValueError: not enough values to unpack (expected 2, got 1)

I cannot find options like output_attentions=True in FluxTransformer2DModel, so I wonder how did you implement these lines.

[tomguluson92]

You need to install my version of diffusers, not public one.

[amoeba04]

Thank you for quick answer.

However, now I meet below error:

Traceback (most recent call last):
  File "train_flux_lora.py", line 43, in <module>
    import diffusers
  File "EraseAnything/diffusers/src/diffusers/__init__.py", line 5, in <module>
    from .utils import (
ModuleNotFoundError: No module named 'diffusers.utils'

I think there is no utils in your version of diffusers ;(
If I'm wrong, please advice me...

[amoeba04]

I think I succeed to run train.sh by copying utils and schedulers from diffusers 0.32.0 into your diffusers directory.
If you have any recommendation instead of this, please advice me!
Thank you!

[tomguluson92]

I think you can just submit a PR and I will me the e it. Thanks!

[kojunseok0217]

How can you run train.sh? When using diffusers in this repo, I got many errors for running inference.py. Is that ok for train.sh??

[kojunseok0217]

I have one more question about training. Which dataset did you use??