How bgu-cs-vil/cludi works
CLUDI is a research implementation for clustering image embeddings with self-supervised diffusion. The documented workflow obtains pretrained-model embeddings with the external TEMI pipeline, then runs cludi.py to train student/teacher models and produce cluster assignments evaluated with clustering metrics. Internally, diff_utils.py owns diffusion noising and sampling calls. The sampled excerpts establish the main training and sampling links; other utility connections are limited to references in the source index, so the graph does not infer deeper data flow for them.
Main entry points: cludi.py Training entry point · train_utils.py Training utilities · denoising_diffusion_pytorch.py Gaussian diffusion · attend.py Attention kernels
Components 15
Clustering workflow
- Image embeddings
- Training entry point
cludi.py - Cluster evaluation
eval_utils.py
Model and diffusion runtime
- Student teacher models
vit_new.py - Gaussian diffusion
denoising_diffusion_pytorch/denoising_diffusion_pytorch.py - Diffusion sampling ops
diff_utils.py - Attention kernels
denoising_diffusion_pytorch/attend.py - 2D Karras U-Net
denoising_diffusion_pytorch/karras_unet.py - 1D Karras U-Net
denoising_diffusion_pytorch/karras_unet_1d.py - 3D Karras U-Net
denoising_diffusion_pytorch/karras_unet_3d.py - Diffusion variants
denoising_diffusion_pytorch
Training support
- Image augmentations
dino_utils.py - Training utilities
train_utils.py - Clustering losses
loss_utils.py - Logging utilities
log_utils.py
Connections 15
- Image embeddings → Training entry point: provides features
README.md - Training entry point → Student teacher models: trains models
cludi.py - Training entry point → Diffusion sampling ops: runs sampling
cludi.py - Diffusion sampling ops → Gaussian diffusion: noises and samples
diff_utils.py - Training utilities → Student teacher models: imports model
train_utils.py - Training utilities → Gaussian diffusion: imports diffusion
train_utils.py - Training entry point → Cluster evaluation: imports evaluation
cludi.py - Training entry point → Training utilities: imports utilities
cludi.py - Training entry point → Clustering losses: imports losses
cludi.py - Training entry point → Image augmentations: imports transforms
cludi.py - Training entry point → Logging utilities: imports logging
cludi.py - 2D Karras U-Net → Attention kernels: uses attention
denoising_diffusion_pytorch/karras_unet.py - 3D Karras U-Net → Attention kernels: uses attention
denoising_diffusion_pytorch/karras_unet_3d.py - 1D Karras U-Net → Attention kernels: uses attention
denoising_diffusion_pytorch/karras_unet_1d.py - Diffusion variants → Gaussian diffusion: extends diffusion APIs
denoising_diffusion_pytorch/__init__.py