Diffusion

My replication for diffusion model.

Environment

conda env create -f environment.yml conda activate diffusion

Data

run the code in extract_data.ipynb to download CIFAR-10 data.

Unconditional DDPM

cd unconditional_DDPM

Train

torchrun --nnodes=1 --nproc_per_node=2 train.py --config config.yaml --use_amp

Sample

torchrun --nnodes=1 --nproc_per_node=2 sample.py --config config.yaml --use_amp

Evaluation

python3 -m pytorch_fid unconditional_DDPM/2000_unconditional_DDPM/EMAgenerated_ep1999_ddpm_steps1000/pngs data/cifar10-pngs

FID: 3.2444050218018106

Class-condiitonal DDPM

cd Class_conditional_DDPM

Train

torchrun --nnodes=1 --nproc_per_node=2 --rdzv_endpoint=localhost:55000 train.py --config config.yaml --use_amp

Sample

torchrun --nnodes=1 --nproc_per_node=2 --rdzv_endpoint=localhost:55500 sample.py --config config.yaml --use_amp

Evaluation

python3 -m pytorch_fid Class_conditional_DDPM/2000_conditional_DDPM/EMAgenerated_ep1999_ddpm_steps1000/pngs data/cifar10-pngs

FID: 2.9473849136342665

Unconditional EDM

cd unconditional_EDM

Train

torchrun --nnodes=1 --nproc_per_node=2 --rdzv_endpoint=localhost:55001 train.py --config config.yaml --use_amp

Sample

torchrun --nnodes=1 --nproc_per_node=2 --rdzv_endpoint=localhost:55501 sample.py --config config.yaml --use_amp

Evaluation

python3 -m pytorch_fid unconditional_EDM/1200_unconditional_EDM/EMAgenerated_ep1199_edm_steps18/pngs data/cifar10-pngs

FID: 4.234571648031874

Class-condiitonal EDM

Train

torchrun --nnodes=1 --nproc_per_node=2 --rdzv_endpoint=localhost:55002 train.py --config config.yaml --use_amp

Sample

torchrun --nnodes=1 --nproc_per_node=2 --rdzv_endpoint=localhost:55502 sample.py --config config.yaml --use_amp

Evaluation

python3 -m pytorch_fid Class_conditional_EDM/1200_conditional_EDM/EMAgenerated_ep1199_edm_steps18/pngs data/cifar10-pngs

FID: 3.707334678054849

Unconditional Consistency Model

Class-conditional Consistency Model