Could I get the code for experiment in CellGAN

[zhang373]

Thanks a lot for your research and kindness to share the GAN code. I wonder if I could get the code for experiments shown in your paper? Especially the code for proving that CellGan can better augment dataset in experiment 3.3.

[zhang373]

I would be really happy if I can also get the code for other code for experiment. Again, thanks a lot for sharing the training code!

[ZhenrongShen]

Thanks for your interest in our work! For experiment 3.3, we trained the downstream classifiers with the combined real and synthetic data using MMClassification (https://github.com/open-mmlab/mmclassification). I would recommend that you refer to their documentation to adapt the code for the replication or your specific use case.

[zhang373]

Thanks a lot for your kind reply! I will try to use the code! I have another question. Which repo does you use for LDM? Have you made any changes to the LDM code? How is category control implemented?

[ZhenrongShen]

The LDM official repository (https://github.com/CompVis/latent-diffusion) provides configurations for image generation using class labels. But for the sake of implementation convenience, I chose to develop my own LDM code following the tutorial at https://nn.labml.ai/diffusion/index.html. Given the constraints of GPU memory and dataset size, I made adjustments to LDM implementation by reducing the channels in both the autoencoder and the denoising U-Net. Both KL- and VQ-reg autoencoders with f=8 are trained and tested on our dataset. We found that the KL-reg autoencoder performed slightly better, and thus used it in LDM implementation. For category control, I adopted a similar approach to CellGAN, which first produces class embeddings from one-hot class labels using an MLP and then controls the denoising U-Net via cross-attention layers.