This code provides the necessary tools for converting a tensorflow-trained proGAN into a pytorch-based model with the same structure, so that the user can then run Gandissect on their model.

In order to perform the conversion, the user should first create a new directory with their name of choice inside of netdissect, and place the .pkl file of their model in that directory.

The user should then run

python netdissect/translate.py --id <name of directory>

The conversion code will then output a .pth file inside of the directory you made. If you'd like to then dissect this converted version of your model, the next step is to copy or move that .pth file into models/karras, and then follow the instructions listed in the original ganddisect repo.

Much of the code in this project was borrowed from the following sources.

All files except the ones listed below are from https://github.com/CSAILVision/gandissect.

The following files are from https://github.com/llSourcell/Progressive_GANs:

• netdissect/dataset.py
• netdissect/legacy.py
• netdissect/misc.py
• netdissect/networks.py
• netdissect/tfutil.py

The following file is from https://github.com/NVlabs/stylegan/blob/master/config.py:

• netdissect/config.py

The following file is originally from https://gandissect.csail.mit.edu/code/convert_to_pytorch.py, after which I made some modifications.

• netdissect/translate.py

David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Joshua B. Tenenbaum, William T. Freeman, Antonio Torralba. GAN Dissection: Visualizing and Understanding Generative Adversarial Networks, Proceedings of the International Conference on Learning Representations (ICLR), 2019.