The model trained on my own dataset gives an error when using inference_resshift.py for inference.

[zwt811]

I used autoencoder_vq_f4.pth as the VQGAN model to train my own dataset, but during inference, I encountered the error KeyError: 'vae'.

[zsyOAOA]

For the task of deblurring, I finetuned the autoencoder with LoRA, which requires the key "vae". Please carefully read the code of this part.

https://github.com/zsyOAOA/ResShift/blob/journal/sampler.py#L103

[Mark-Dou]

For the task of deblurring, I finetuned the autoencoder with LoRA, which requires the key "vae". Please carefully read the code of this part.

https://github.com/zsyOAOA/ResShift/blob/journal/sampler.py#L103

It seems the config file for deblurring task is set 'tune_decoder' is True rather than lora_tune_decoder is True, besides, the 'save_ckpt' also do not save the fine-tuned decoder part, so that raise the error when inference with custom datasets re-training.

[Mark-Dou]

Also, the re-training part do not contain any logic for fine-tuning the decoder part, while set all params in autoencoder to False

[zsyOAOA]

I have updated the configs for fine-tuning the decoder. FYI. @Mark-Dou