The pretrained model weights can be downloaded from Google Drive.
After downloading, place the .pt file in the models/ directory.
The training code for these models is availble at https://github.com/gabrielraya/ddpm-mnist.git The repository follows a similar structure to the previously mentioned one.
Images are generated from utils/Generate_Batch.py.
The posterior is computed iteratively using the compute_posterior_DNG(...) function present in Generate_Batch.py (Alg. 1). For numerical stability the posterior is clamped between p_min and p_max.
Make sure you have older checkpoints of the diffusers library such to work with the pretrained unconditional CIFAR10 model from https://huggingface.co/google/ddpm-cifar10-32: pip install diffusers==0.18.2, huggingface_hub==0.25.0
Then generate the desired samples using for example: python main.py --N_tot 32 --N_batch 32 --to_remove_class 0 --guidance_type 'dynamic_negative_guidance' --guidance_scale 5. --prior 0.01 --Temp 0.2 --offset 2e-4 --p_max 0.8