search

chaofengc / IQA-PyTorch

👁️ 🖼️ 🔥PyTorch Toolbox for Image Quality Assessment, including PSNR, SSIM, LPIPS, FID, NIQE, NRQM(Ma), MUSIQ, TOPIQ, NIMA, DBCNN, BRISQUE, PI and more...
https://iqa-pytorch.readthedocs.io/
Other
2.03k stars 176 forks source link

About retrain topiq model #109

Closed cyclopssssss closed 1 year ago

cyclopssssss commented 1 year ago

use your training options of TOPIQ on koniq dataset like below: python pyiqa/train.py -opt options/train/TOPIQ/train_TOPIQ_res50_koniq.yml and the result is bit lower than the benchmark you provided: M~SQJL9A@D%WN1FOV81Z1DO Could you give some advice about training the model?🥺 Thanks for your patience!

chaofengc commented 1 year ago

Thanks for your information. I have checked the training history and updated the loss function as below https://github.com/chaofengc/IQA-PyTorch/blob/81edc661f1ea13ccd7d0cec8a9518f377155f9da/options/train/TOPIQ/train_TOPIQ_res50_koniq.yml#L82-L88

You may refer to the paper Norm-in-Norm Loss with Faster Convergence and Better Performance for Image Quality Assessment for more details.

cyclopssssss commented 1 year ago

Thanks again for your reply!