[ERROR] stuck at converting sd 15 ckpt about everydream2trainer

[Dwuuds]

**Megachill ** commented Apr 6, 2023 •

Having the exact same issue. I cant find any documentation about it anywhere, and if I try to convert a different 1.5 model it throws the error "convert_original_stable_diffusion_to_diffusers.py: error: unrecognized arguments: epsilon"

[victorchall]

Someone submitted an update quite a while ago that attempts to order weight/bias params because it seems something out there must place them in state_dict in a different order. I believe this solved most of those issues. If there is a particular ckpt still causing errors please open a new issue with details.

[buzzz3d]

Running it a second time worked for me.
text_config_dict is provided which will be used to initialize CLIPTextConfig. The value text_config["id2label"] will be overriden.
E:\AI-Image-Models\EveryDream2trainer\venv\lib\site-packages\transformers\models\clip\feature_extraction_clip.py:28: FutureWarning: The class CLIPFeatureExtractor is deprecated and will be removed in version 5 of Transformers. Please use CLIPImageProcessor instead.
warnings.warn(

• Converted model to diffusers for training: ckpt_cache/sd_v1-5_vae

[victorchall]

I followed your guide and wanted to start the conversion of the 1_5 model but that is not working, I converted the sd_2-1 768 no problem whatsoever, but the 15 is stuck.

Left it running over night has not done anything.

Conversion command as follows:

python utils/convert_original_stable_diffusion_to_diffusers.py --scheduler_type ddim ^
--original_config_file v1-inference.yaml ^
--image_size 512 ^
--checkpoint_path sd_v1-5_vae.ckpt ^
--prediction_type epsilon ^
--upcast_attn False ^
--dump_path "ckpt_cache/sd_v1-5_vae"

The image isn't showing any error, that's normal for SD1.x because it always throws some deprecation warnings. But its just a warning, not an error.

I don't know why it would hang. Only theory is you dragged your mouse on the window and effectively paused processing. The whole thing should take moments.

If you want to use SD 1.5 use just resume_ckpt "panopstors/everydream" and it will just pull the preconverted one off huggingface.