yeonghyeon / dgm-tf Goto Github PK

Hi! I am trying to run the code on colab but unfortunately, there is a memory leak during training. Since the default version of TensorFlow of colab is v2, I have to set it back into v1 so that it runs successfully. Could you please provide the versions of the packages of your environment for this repo? Thanks in advance!

Hi, thank you for your great work! And I have some questions and need your help.

I found that in your implementation, the encoder and decoder have different structures compared to the original DGM paper. Basically, it has 2 conv layers followed by a pooling layer every time. Is it because the mnist data have relatively small sizes and don't need too many layers?

I'm currently trying to use DGM to generate residue maps on CXR images with the size of 224*224(same as in the paper). And I've modified the depth of the model and in/out channels. It has successfully output the parameters for flow 1, 2 and 3. But when it starts training, it has the following bugs.

In the encoder, we have modified it to this:

And in the decoder, we have modified to this:

I really like your work and want to see how it works on my customized dataset. Could you give me some instructions on modifying the model? I'm really a novice to this.
I'll appreciate your response.