Implementation of DDPM (https://arxiv.org/abs/2006.11239) in Pytorch with class conditioning.
Model architecture:
- Class conditioned on label.
- UNet with 4 downsample, 4 upsample blocks. Self-attention at the 16x16 resolution and bottleneck layer.
- Pre-activation for all resnet blocks.
accelerate launch train.py
Generate samples and render a grid of images into one image.
python sample_grid.py --checkpoint_path=<path> --rows=8 --cols=8
python sample_grid.py --checkpoint_path=<path> --cols=8 --num_classes=10 --labels_list="airplane,automobile,bird,cat,deer,dog,frog,horse,ship,truck"
Save a large number of images to a destination folder for running eval metrics.
python sample_images.py --checkpoint_path=<path> --output_dir=<dir> --num_images=50000 --batch_size=2500