ermongroup / alignflow Goto Github PK

According to your code of flow2flow which is for unpaired image to image translation, you just shuffle the data before training. I think you need to shuffle the target images in every epoch.

Thanks for your nice work. To run the code, I suggest to correct some typos as bellows:

source activate cflw . --> . source activate aflw

And comment the line

I would like to ask you some question related to the project:

1. Your code implemented "Glow: Generative Flow with Invertible 1×1 Convolutions" , Have you evaluate the performance of generator with different architectures: Glow and RealNVP? I did not find the use case of Glow in the project.

2. I tried to apply cycleflow to the mri brain image with single channel input (gray image). However, the model looks collapse or failure to learning the mapping between domain A and B. The details between regions in the syn. image become blurring, while the cycleGAN still works well. Any suggestion to fix it? Thanks in advance

How to train the model for Unsupervised Domain Adaptation?
How to set the code for the task of Unsupervised Domain Adaptation?
Thank you very much.

Hi, Thanks a lot for sharing your code!

could you please share the args for training Maps dataset?

I have tried with resize_shape='96,96', crop_shape="64,64", lambda_mle=1e-5, but I can only get MSE_src2tgt=0.024, MSE_tgt2src=0.161, is there something wrong?

BTW, how do you evaluate? I noticed that you used nn.BatchNorm2d instead of BatchNormStats, did you use the running stats or use the input batch stats?