Neural Network course project

This project was done as a course project for CS-852, Fall-2024, UNH to better understand diffusion policy.

In the first step for sanity checking I let the diffusion model to generate some simple data. Then I use the same model to generate actions as a diffusion policy.

Quick start: Diffusion Model Sanity check

my_swissroll.ipynb

Diffusion Policy

Diffusion policy reimplementation for the course project of the Neural Network course.

Based on the paper: Visuomotor Policy Learning via Action Diffusion

Installation

• For pushT environment install the original diffusion policy codebase from here

Data download

• download the pusht data and put it inside the data folder

wget https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip

• convert to hdf5 format: Use the zarr2hdf5.ipynb notebook

How to run the code

• Option 1: jupyter notebook

train_example.ipynb jupyter 

• Option 2: train.py script

python train.py --hdf5_file data/pusht/pusht_v7_zarr_206.hdf5

File structure

File name	Description
my_swissroll.ipynb	Sanity check for the diffusion model
train.py	Main script to train the model
train_example.ipynb	Jupyter notebook with the training example
myddpm.py	DDPM scheduler and sampler
noise_predictor_model.py	UNet model
pos_encoding.py	Sinusoidal positional encoding
film.py	Fiture wise Linear Modulation (FiLM)
vision_model.py	Vision encoder
pusht_data_utils.py	Utility functions from original diffusion policy codebase
data/pusht/pusht_v7_zarr_206.hdf5	pusht data in hdf5 format
zarr2hdf5.ipynb	Convert zarr to hdf5 format