greatlog
commented
10 months ago 在DDP训练的时候有个bug忘记改了，已经提交新的commit https://github.com/greatlog/RealDAN/commit/d447dfa89c64378dcc007098a5e44e79f9d268ae 现在应该没问题了。
在yml设置里，datasets/train 字段里有一个 ratio https://github.com/greatlog/RealDAN/blob/d447dfa89c64378dcc007098a5e44e79f9d268ae/codes/config/RealDAN/options/train/dan_edsr_ema.yml#L24 设置为一个比较大的数（示例是256）也能解决这个问题。
Arcananana
作者您好，我尝试重训DAN在real-world合成数据集上4x超分的结果，但是我发现训练很快收敛，并且在验证集上的psnr逐渐降低。 前100000次迭代的验证结果如下： 23-09-26 01:19:10.436 - INFO: <epoch: 94, iter: 5,000, Average sccores: psnr: 22.921618; ssim: 0.604642; 23-09-26 02:05:25.968 - INFO: <epoch:188, iter: 10,000, Average sccores: psnr: 23.417599; ssim: 0.623735; 23-09-26 02:51:48.937 - INFO: <epoch:283, iter: 15,000, Average sccores: psnr: 23.514685; ssim: 0.629169; 23-09-26 03:37:47.865 - INFO: <epoch:377, iter: 20,000, Average sccores: psnr: 23.547827; ssim: 0.631476; 23-09-26 04:22:14.282 - INFO: <epoch:471, iter: 25,000, Average sccores: psnr: 23.571558; ssim: 0.631415; 23-09-26 05:07:12.161 - INFO: <epoch:566, iter: 30,000, Average sccores: psnr: 23.587416; ssim: 0.631748; 23-09-26 05:53:09.046 - INFO: <epoch:660, iter: 35,000, Average sccores: psnr: 23.586239; ssim: 0.631668; 23-09-26 06:38:48.869 - INFO: <epoch:754, iter: 40,000, Average sccores: psnr: 23.589043; ssim: 0.632581; 23-09-26 07:25:00.470 - INFO: <epoch:849, iter: 45,000, Average sccores: psnr: 23.580803; ssim: 0.632208; 23-09-26 08:11:06.986 - INFO: <epoch:943, iter: 50,000, Average sccores: psnr: 23.565691; ssim: 0.632091; 23-09-26 08:57:02.359 - INFO: <epoch:1037, iter: 55,000, Average sccores: psnr: 23.557306; ssim: 0.631869; 23-09-26 09:43:12.951 - INFO: <epoch:1132, iter: 60,000, Average sccores: psnr: 23.540745; ssim: 0.631238; 23-09-26 10:29:18.248 - INFO: <epoch:1226, iter: 65,000, Average sccores: psnr: 23.531569; ssim: 0.630892; 23-09-26 11:15:16.303 - INFO: <epoch:1320, iter: 70,000, Average sccores: psnr: 23.525421; ssim: 0.630781; 23-09-26 12:01:29.084 - INFO: <epoch:1415, iter: 75,000, Average sccores: psnr: 23.514370; ssim: 0.630370; 23-09-26 12:47:46.191 - INFO: <epoch:1509, iter: 80,000, Average sccores: psnr: 23.503069; ssim: 0.630095; 23-09-26 13:33:51.775 - INFO: <epoch:1603, iter: 85,000, Average sccores: psnr: 23.487091; ssim: 0.629482; 23-09-26 14:20:05.914 - INFO: <epoch:1698, iter: 90,000, Average sccores: psnr: 23.481306; ssim: 0.629495; 23-09-26 15:06:08.774 - INFO: <epoch:1792, iter: 95,000, Average sccores: psnr: 23.476807; ssim: 0.629153; 23-09-26 15:51:55.213 - INFO: <epoch:1886, iter: 100,000, Average sccores: psnr: 23.469683; ssim: 0.629166;
而您提供的预训练模型测得的psnr值为23.82,请问有什么我遗漏的地方吗？我的配置文件使用的是RealDAN/codes/config/RealDAN/options/train/dan_edsr_ema.yml，仅修改了文件路径。