search

ming053l / DRCT

Accepted by New Trends in Image Restoration and Enhancement workshop (NTIRE), in conjunction with CVPR 2024.
MIT License
157 stars 11 forks source link

difference between DRTC-L and DRCT-XL #15

Open Ar0Kim opened 3 months ago

Ar0Kim commented 3 months ago

스크린샷 2024-06-30 183500 Hi, thank you for sharing your project. I have a question about the differences between DRTC-L and DRCT-XL. In the paper, it is mentioned that DRCT-L is also pretrained on the ImageNet dataset, but on the GitHub page, it is indicated as "pretrained on ImageNet" only under DRCT-XL. Secondly, to train DRCT-L according to your paper, for SRX4, should I first use train_DRCT-L_SRx4_ImageNet_from_scratch.yml to train and obtain the pretrained model, and then use train_DRCT-L_SRx4_finetune_from_ImageNet_pretrain.yml for the final training? image

ming053l commented 3 months ago

hi, the difference between Large and X-Large version is the depth of RDG. XL version have 14 RDG, L have 12.

Secondly, to train DRCT-L according to your paper, for SRX4, should I first use train_DRCT-L_SRx4_ImageNet_from_scratch.yml to train and obtain the pretrained model, and then use train_DRCT-L_SRx4_finetune_from_ImageNet_pretrain.yml for the final training?

Yes, you need to pre-train DRCT on large scale dataset ('imagenet') and then fine-tune it on your own dataset like DF2K

Ar0Kim commented 3 months ago

thank you for the reply. it's so helpful :)