It is a implementation of DDPM for 2D point cloud.
For dataset: We generate the dataset by ourself. You can generate the dataset by running the Dataset_generator.py.
This image is the sample image we generate.
For train: The code is in the Train.py file. The model we use is a simple MLP instead of U-net. The model is supposed to converge for about 1000 epoch. You can run this file for training.
For Infer: The training process is separated from inference process. After implemented Train.py, weight of the model "save.pt" will generated. You can choose a specific weight for inference in Infer.py. After infering process, we will also print the loss and the epoch for the training of the weight.
For Config: You can change the num_step and hyperparameters in Config.py file.
Reference: https://github.com/azad-academy/denoising-diffusion-model