DDPM Pytorch Implement

A pytorch DDPM implementation, simple to train and test on your dataset

Requirement

python >= 3.10

torch >= 2.0

albumentations

numpy

Quick Start

Training

First prepare your dataset.For simplest method, put your images(.jpg or .png format, rgb images) into data/images, or you can organize your images as pytorch image folder, and change dataset_path in config.yaml to your folder.
Then you can change options in config.yaml according to your preference, like batch_size, gpus, image resolutions etc.
After prepared data and set configs, just run "python train.py" to train your DDPM model, images generated after training each epoch will be saved in exp/generated by default.Model weights and optimizer state will be saved as exp/model.pth and exp/training_params.pt respectively after training is finished.

Testing(Generating Images)

After the training process is finished, you can simply run "python test.py" to generate images from trained model, generated images will be saved in exp/generated_test by default.

Generated Examples

Here are some real and generated examples of konachan anime faces dataset.Resolution of example images have been resized to 128x128 considering training cost.

Images from dataset