BasicSR
BasicSR copied to clipboard
Some Trainning & Testing issue
当我用DF2K训练SwinIR x4时,这是我的成绩,“# psnr: 32.6784 Best: 32.6911 @ 340000 iter”,训练50w个iteration, 我总是达不到论文里面介绍的在Set5上面的32.9(PSNR) , 然后就是能不能更新完全SwinIR的test文件,RCAN的Test文件,RCAN下面的train是x2的,到test就是x4了。 就自己弄的就是会报错,官方更新总是很不错的!! 谢谢你们的框架
hi can you reimplememt the same results on x2 task by using swinir?
I am sorry...... whatever I set the resolution of LR images as 48 or 64, the result is still hard to reach the value of Paper, whether the training set is DIV2K or DF2K using SwinIR; I did a lot of experiments to reimplement SwinIR training, X2 or X4 by scratch, the result is still lowering the value of Paper about 0.3db(Set5 PSNR). And training Transformer Models like SwinIR and HAT, the Result of training SR_X4 from scratch is lower than pre_train SR_x2, can you get my point?
0.3db is too much, i got the result drops nearly 0.05db.
当我用DF2K训练SwinIR x4时,这是我的成绩,“# psnr: 32.6784 Best: 32.6911 @ 340000 iter”,训练50w个iteration, 我总是达不到论文里面介绍的在Set5上面的32.9(PSNR) , 然后就是能不能更新完全SwinIR的test文件,RCAN的Test文件,RCAN下面的train是x2的,到test就是x4了。 就自己弄的就是会报错,官方更新总是很不错的!! 谢谢你们的框架
您好,我也遇见了和您一样的问题,甚至结果还不如您,请问您解决了吗,期待您的回复
你可以将swinirmodel类的test方法按官方库里的改一下,我用的48的尺寸的输入得到的指标基本和官方库测出来的一样,32.72