raindrop313

Hi, Thank you for sharing this code with us. However, I was confused with the axial rotary embeddings in rotary_embedding_torch.py file. " elif freqs_for == 'pixel': freqs = torch.linspace(1., max_freq...

I have modified the model of GLEAN, which can be trained normally and will periodically make inference on the validation set, but when the inference is completed, an error is...

TLC在平均池化中的kernel_size

2

感谢你的分享，我对于平均池化的局部性有一些疑问。 您的工作应该是想要减小推理和训练的不一致性，当推理阶段图片大于训练的图片时，根据我对论文的理解，对于平均池化，会进行一个局部的池化，应该用卷积的方式完成。例如训练用64*64的patch，推理用256*256的图片，那么应该池化核应该为96（64*1.5），但是您开源的代码计算方法如下： “”“ self.kernel_size[0] = x.shape[2]*self.base_size[0]//self.train_size[-2] self.kernel_size[1] = x.shape[3]*self.base_size[1]//self.train_size[-1] ”“” 由于base_size是train_size的1.5倍，那么池化核不是永远都是输入特征图x的1.5倍吗，好像和local的思想不同。 也许是我理解有误，恳请指点一二，不胜感激

At 24x super resolution, using matlab's downsampling and upsampling does not give 1024 inputs, but I see that the paper releases results for X24, how should I run the model...

Hello， I see that there are only the training models of SOTS in your Google Drive, can you post the training model of O-Haze? thanks.