DDPM

PaddlePaddle reproduction of Denoising Diffusion Probabilistic Models (DDPM)

Denoising Diffusion Probabilistic Models (DDPM)

1. Introduction

This is a PaddlePaddle reproduction of Denoising Diffusion Probabilistic Models (DDPM) and Denoising Diffusion Implicit Models (DDIM). We trained the diffusion model with original-designed UNet on CIFAR10 (32X32) and simplified verison (FPN, according to Keras-DDPM) on CelebA-HQ (128X128). The code of this reproduction can be found in link.

CelebA-HQ (DDPM) generated samples

2. DDIM & Progressive Generation

CelebA-HQ (DDIM, stride 10) generated samples
CelebA-HQ (DDPM) progressive generation

3. Training

3.1 Datasets

  • CIFAR10

  • CelebA-HQ (please download and place it in datasets folder)

3.2 Train from scratch

  • CIFAR10

      # train with default
  python run_cifar10.py

    note: training on Tesla V100 16G GPU with default configuration (800k iters) requires about 100 hours.

  • CelebA-HQ

      # train with default
  python run_celeba_hq.py

  # the sampling procedure in training is time-consupming, you can use '--sample_interval 5000' to reduce the frequency

    note: training on Tesla V100 16G GPU with default configuration (1000k iters) requires about 120 hours.

3.3 Resume training

# CIFAR10
python run_cifar10.py --resume_ckpt_path '<CHECKPOINT_PATH>' --resume True

# CelebA-HQ
python run_celeba_hq.py --resume_ckpt_path '<CHECKPOINT_PATH>' --resume True

4. Sampling

4.1 Pre-trained Weights

Baidu Yun, password: tew7

4.2 DDPM

  • CIFAR10

      # 1. convert ckpt to pretrained weights
  python run_cifar10.py --pretrained_path './ckpts/ddpm_cifar10_i800000_ema.pdparam' --resume_ckpt_path './ckpts/cifar10_ckpt_iter_800000.pdparam' --mode 'convert'

  # 2. sample 400 images (20X20)
  python run_cifar10.py --pretrained_path './ckpts/ddpm_cifar10_i800000_ema.pdparam' --num_images '400-20-20' --mode 'denoise'

  # 3. sample 16 images with progressive generation
  python run_cifar10.py --pretrained_path './ckpts/ddpm_cifar10_i800000_ema.pdparam' --num_images '16-4-4'  --mode 'denoise' --sample_postfix 'ddpm_cifar10_sequence' --sequence True

  • CelebA-HQ

      # 1. convert ckpt to pretrained weights
  python run_celeba_hq.py --pretrained_path './ckpts/ddpm_celeba_hq_i1000000_ema.pdparam' --resume_ckpt_path './ckpts/celeba_hq_ckpt_iter_1000000.pdparam' --mode 'convert'

  # 2. sample 64 images (8X8)
  python run_celeba_hq.py --pretrained_path './ckpts/ddpm_celeba_hq_i1000000_ema.pdparam' --num_images '64-8-8' --mode 'denoise' --batchsize 64

  # 3. sample 8 images with progressive generation
  python run_celeba_hq.py --pretrained_path './ckpts/ddpm_celeba_hq_i1000000_ema.pdparam' --num_images '8-4-2'  --mode 'denoise' --sample_postfix 'ddpm_celeba_hq_sequence' --sequence True

4.3 DDIM

note: you need to convert the weights if you want to use the checkpoints (see 2.1)

  • CIFAR10

      # 1. sample 400 images (20X20)
  python run_cifar10_ddim.py --pretrained_path './ckpts/ddpm_cifar10_i800000_ema.pdparam' --num_images '400-20-20'

  • CelebA-HQ

      # 1. sample 64 images (8X8)
  python run_celeba_hq_ddim.py --pretrained_path './ckpts/ddpm_cifar10_i1000000_ema.pdparam' --num_images '64-8-8' --batchsize 64

5. References