About trainset setting

[zhangchen98]

I had a little trouble reproducing the results using the training sets mentioned in the paper (LSDIR and FFHQ 10k), and the model I trained (12501 step) was not as good as your released model in the background of some images. I followed the same training set and batchsize, but I found that there is a "--highquality_dataset_txt_paths" configuration option in the training command, and I don't see which training images should be put here. Can you help me?

[csslc]

Thanks for your question. We observe that a high-quality training dataset can significantly enhance the semantic information of the SR outputs. Therefore, the high-quality data are selected based on no-reference metrics, such as MUSIQ. Specifically, in the LSDIR dataset, images with a MUSIQ score higher than 76 are filtered and used as high-quality training data. 'preset/gt_selected_path.txt' records the paths of the high-quality training data using get_path.

[Albert2X]

Thanks for your question. We observe that a high-quality training dataset can significantly enhance the semantic information of the SR outputs. Therefore, the high-quality data are selected based on no-reference metrics, such as MUSIQ. Specifically, in the LSDIR dataset, images with a MUSIQ score higher than 76 are filtered and used as high-quality training data. 'preset/gt_selected_path.txt' records the paths of the high-quality training data using get_path.

It seems that you set --prob 0.1 to ensure that most of the images come from the high-quality training dataset, is that correct?
In addition, why does the code in dataset.py use randomly generated indices to select images?