Denoising Diffusion Probabilistic Models

PyTorch Re-Implementation of Denoising Diffusion Probabilistic Models

Plot sampled epoch images - just run pic_extract.ipynb

train_vae.py is used for vae-denoising task as a baseline compared to ddpm.

• Datasets
  • CIFAR10 https://www.cs.toronto.edu/~kriz/cifar.html
  • Mnist https://yann.lecun.com/exdb/mnist/
  • CelebA https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Requirements

pip install -r requirements.txt

Training

• Take CelebA for example:

  python ./ddpm/train.py --dataset CelebA 
  

• Overwrite arguments

  python ./ddpm/train.py --dataset CelebA --epochs 50 --channels 3 --sample_epoch 5
  
  

FID Evaluate

make sure install package pytorch-fid first

pip install pytorch-fid

cd ./ddpm
python fid.py /real/image/dir /generated/image/dir 

[Optional] Overwrite argument

  --sample_epoch 1 --epoch 10 --batch_size 20

Reference：

• Denoising Diffusion Probabilistic Models](https://arxiv.org/abs/2006.11239)

• Official TensorFlow implementation](https://github.com/hojonathanho/diffusion)

• Denoising Diffusion Probabilistic Models (DDPM) | https://nn.labml.ai/diffusion/ddpm/index.html

• labmlai | https://github.com/labmlai/annotated_deep_learning_paper_implementations

• CelebA Dataset | https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

• U-Net model for Denoising Diffusion Probabilistic Models (DDPM) | https://nn.labml.ai/diffusion/ddpm/unet.html

  Todo

  -Eval "fake image" step 51000/800000 (unfinished training) FID=87,65