Meta Upscale Module about meta-sr-pytorch

[tetterl]

meta.py has a bug on line https://github.com/XuecaiHu/Meta-SR-Pytorch/blob/master/model/meta.py#L99 since we always feed the scaling factor of 2. That might be the reason that it doesn't work.

[tetterl]

metardn.py implements Meta-RDN from the paper. I'm not fully sure if meta.py implements BicuConv or Meta-Bicu from the paper, I guess one of those two is missing. Maybe @XuecaiHu can clarify it?

[XuecaiHu]

@Auth0rM0rgan @tetterl sorry for the late reply. meta.py is not the meta-bicu/ bicuconv. In meta.py, we directly concatenate the output feature(by upsampling s times) and pos_mat instead of using the pos_mat to predict weights for each pixel. But we find it doesnot work. It is a method we donot mention in our paper.

[XuecaiHu]

We donot provide the implementation of metabicu and bicuconv. It is easy to implement them by reference to meta.py

[XuecaiHu]

thank you, i will fix the bug later and retrain our model.