Skip to content

Can't repoduce your result

Hi,

Thank you for your work on this project. I recently tried to reproduce the results of your model using the provided checkpoint (MoE-DiffIR_DACLIP.ckpt) and configuration file (MoE-DiffIR_DACLIP.yaml) on the LIVE1 test dataset you provided for the JPEG task. I evaluated the results using the IQA-PyTorch command.

However, my results were quite different from those reported in your paper. Below are the results I got:

Image

For comparison, the corresponding results from your paper are shown here:

Image

Additionally, while loading the checkpoint, I encountered some unexpected keys in the prompt module, as shown in the following log:

Restored from MoE-DiffIR_DACLIP.ckpt with 0 missing and 5 unexpected keys
Unexpected Keys: ['cond_stage_model.adapter2.fc.0.weight', 'cond_stage_model.adapter2.fc.2.weight', 'structcond_stage_model.task_embedding', 'structcond_stage_model.conv_middle.0.weight', 'structcond_stage_model.conv_middle.0.bias']

I would appreciate it if you could help me identify whether I might have made a mistake during the setup or if the provided checkpoint might not be the correct one. If the checkpoint is correct, would you be able to share the evaluation code or provide additional guidance on achieving the reported results?