Fangneng Zhan

@panzhang0212 do you have any idea about this error? thanks.

> Hi, I run the code with resolution 128, but do not find any problem in the code. What's your command? if you run with 128, you need to add...

Hi @UdonDa , thanks for your interest. The code inherits from the implementation of CoCosNet which includes other information for building correspondence. It is indeed somewhat different from the description...

We run the cocostuff experiments in a separated repo, which may not match this repo perfectly. I have uploaded the data processing code, training and testing scripts of cocostuff to...

Hi, how about the size of ref_? You can print the size of ref_ to check if it matches with f_div_C.

Hi, do you change parameters setting in training scripts? As the learned correspondences start to converge to constant, you can delete --use_coordconv or decrease --ctx_w to be smaller (e.g., 0.5)...

Hi, 1. of course, the model can be applied to other datasets. You should add the python file of the corresponding dataset, e.g., data/ade20k_dataset.py, and modify the processing code here...

你好，通过X和Z之间最优传输，可以得到一个矩阵（Transport Plan），然后通过这个矩阵特征Z会向X对齐得到aligned features (图中的特征热力图)。aligned features会和原始X一起输入到生成网络当中（绿色部分）。映射后的特征就是简单的把transport plan和特征Z做个外积。

> 因为我暂时没有在网上找到有关矩阵外积的定义，只找到一些向量外积的定义，所有不太清楚您这里所谓的外积是指什么意思？ 可以看下cocosnet 文章里面的公式（3）和（4），https://openaccess.thecvf.com/content_CVPR_2020/papers/Zhang_Cross-Domain_Correspondence_Learning_for_Exemplar-Based_Image_Translation_CVPR_2020_paper.pdf

The provided weight has been checked before uploading, the FID score of UNITE paper should be reproduced with the weights. You should compute the FID between testing set and generated...