Zhengxiong Luo

Thanks for your reply. I still cannot reimplement the reported results of CycleSR by myself. And I seem to have found the reason: the [div2k_unpair.py](https://github.com/huawei-noah/vega/blob/master/zeus/datasets/common/div2k_unpair.py) actually is not "unpaired". According...

Thanks. I know how to preprocess the datasets, and I have no doubt that the code can produce PSNR results better than 25 dB. The question is that the code...

Thanks for your reply and explanation. CycleSR is definitely an excellent method for unpaired SR. I am working on it because I believe it is promising. Yet, I still cannot...

One more question. According to https://github.com/huawei-noah/vega/blob/1bba6100ead802697e691403b951e6652a99ccae/examples/data_augmentation/cyclesr/cyclesr.yml#L19 and https://github.com/huawei-noah/vega/blob/1bba6100ead802697e691403b951e6652a99ccae/examples/data_augmentation/cyclesr/cyclesr.yml#L28, the training batch_size is 64, and imgs_per_gpu is 4. So do you use 16 GPUs to train CycleSR?

可以参考 [Bayesian Image Super-resolution](https://proceedings.neurips.cc/paper/2002/file/88bfcf02e7f554f9e9ea350b699bc6a7-Paper.pdf)，[Efficient Marginal Likelihood Optimization in Blind Deconvolution](https://ieeexplore.ieee.org/document/5995308)

You need to change the architecture of DAN at the same time. When the Restorer and Estimator are created, you need to set [in_nc=1](https://github.com/greatlog/DAN/blob/53f9b7010ba9ee662b2abb0b33c8df3939829a9f/codes/config/DANv1/models/modules/dan_arch.py#L159)

谢谢关注，我们近期整理，会尽快更新。

noise生成时有一些参数需要自己指定，比如指定spatial = False时，一般需要指定 mix=True，这样生成的noise还是 [B, 3, 32,32]，或者直接指定spatial=True（详细过程参考deg_arch.py）。提供的yml默认设置可能有些问题，感谢提醒，我回头修改一下。

deg下面是单独训练PDM的配置，optim_deg=True, optim_sr=False。因为validation是针对SR做的，训练deg的时候没有添加netSR，所以validation没法做。你可以把val_freq改成和 niter 一样大，这样就不会在训练时做 validation了。

除了使用PDM帮助训练超分模型，我没想到其他办法可以很好的衡量PDM的性能，因为PDM没有真正意义上的GT。训练好的PDM可以单独运行，大图进，小图出，但是训练过程中没法validation。如果大家有什么好的建议欢迎提PR。