advimman / lama

🦙 LaMa Image Inpainting, Resolution-robust Large Mask Inpainting with Fourier Convolutions, WACV 2022
https://advimman.github.io/lama-project/
Apache License 2.0
8.14k stars 864 forks source link

Cannot fine tune CelebaHq model checkpoint #268

Open aman-autophoto-ai opened 1 year ago

aman-autophoto-ai commented 1 year ago

Hello, I am unable to load celebaHq model.ckpt for fine tuning purpose. It show the below warning even after using "+trainer.kwargs.resume_from_checkpoint=LaMa_models/lama-celeba-hq/lama-fourier/models/best.ckpt" flag for using pretrained checkpoints.

UserWarning: No checkpoint file exists at resume_from_checkpoint. Start from scratch

Kindly help me resolve this issue.

Best regards, Aman

aman-autophoto-ai commented 1 year ago

Got the solution for the above problem. But now it is showing another error:

lama_error_model_loading

aman-autophoto-ai commented 1 year ago

@windj007 Hello, I am currently working on a project where i need to fine tune my model, and i believe that using the LAMA inpainting CelebaHQ model weights files as a starting point could be greatly benefit my research. I kindly request if you would be willing to share the complete CelebaHQ model weight files (both Gen&Disc) with me.

Looking forward for your help and support.

best regards, Aman Gupta

yuji-byun commented 1 year ago

Got the solution for the above problem. But now it is showing another error:

lama_error_model_loading

Hi! How did you solve this problem? I have same error message now

amangupta2303 commented 1 year ago

@HappyYuji

Hello, I have addressed the "UserWarning: No checkpoint file exists at resume_from_checkpoint. Start from scratch" error by using the absolute path for the weight files.

Regarding the error depicted in the image, I discovered that the pretrained LAMA model does not include discriminator weight files. Instead, it only contains generator weights in the .ckpt file. This means we cannot fine-tune the pretrained model on our dataset. The "best.ckpt" model weights are solely intended for inference purposes.

Furthermore, I believe the authors have only provided a complete checkpoint (both discriminator and generator weights) for the Big LAMA variant.

yuji-byun commented 1 year ago

@aman-autophoto-ai Thank you! Then I also hope that the authors provide a complete checkpoint. Have a nice day.

aman-autophoto-ai commented 1 year ago

+1