Training Denoising Diffusion Probability Model (DDPM) on CIFAR-10

This is a fork of github repo: https://github.com/zoubohao/DenoisingDiffusionProbabilityModel-ddpm- I modify the code to make it easier to run and add some comments to make it easier to understand. I also trained the model on CIFAR-10 dataset and upload the pretrain weight, in Checkpoints folder. CIFAR-10 dataset is in CIFAR10

To provide a vivid example of diffusion process, I also add a part to demonstrate the diffusion process of DDPM. The result is in Imgs

How to run

0. Install the requirements

pip install torch torchvision numpy tqdm

1. Train the model

train the model without guidence:

python Main.py --state train

it takes about 3 minutes to train one epoch on a single V100 GPU. It takes 10 hours to train 200 epochs. The ckeckpoint used to generate image is ckpt_150_.pt

train the model with guidence:

python MainCondition.py --state train

see Main.py and MainCondition.py for more details about the parameters.

2. Generate images

python Main.py --state eval --sampled_dir ./out --save_weight_dir ./Checkpoints/ --test_load_weight ckpt_150_.pt

generate image with intermediate steps:

python Main.py --state eval --sampled_dir ./out --test_load_weight ckpt_150_.pt --save_middle_result True

generate images with guidence:

python MainCondition.py --state eval --sampled_dir ./out --test_load_weight ckpt_59_.pt

result and demostration

see NoGuidence and Guidence for the generated images.

1. No guidence

2. Guidence

3. Diffusion process

the following part is the original README.md

DenoisingDiffusionProbabilityModel

This may be the simplest implement of DDPM. I trained with CIFAR-10 dataset. The links of pretrain weight, which trained on CIFAR-10 are in the Issue 2.

If you really want to know more about the framwork of DDPM, I have listed some papers for reading by order in the closed Issue 1.

Lil' Log is also a very nice blog for understanding the details of DDPM, the reference is "https://lilianweng.github.io/posts/2021-07-11-diffusion-models/#:~:text=Diffusion%20models%20are%20inspired%20by,data%20samples%20from%20the%20noise."

HOW TO RUN

1. You can run Main.py to train the UNet on CIFAR-10 dataset. After training, you can set the parameters in the model config to see the amazing process of DDPM.
1. You can run MainCondition.py to train UNet on CIFAR-10. This is for DDPM + Classifier free guidence.

Some generated images are showed below:

1. DDPM without guidence:

1. DDPM + Classifier free guidence:

Kevin-lkw / DenoisingDiffusionProbabilityModel