Code from ICLR 2023 paper 'Imitating human behaviour with diffusion models' - https://arxiv.org/abs/2301.10677
This code currently only replicates the claw environment experiments of the paper. We plan to update this with other experiments soon.
- Install Python (ran on Python 3.9) and/or create a fresh environment
- Install requirements
pip install -r requirements.txt
As a lightweight entry-point to the code, we provide claw_mini_script.py, which trains a diffusion model and runs the Diffusion-X sampling procedure, on the claw dataset. The dataset must first be created locally, by running python make_dataset.py, taking about 1GB of space. claw_mini_script.py then runs in around 5 minutes, outputting code/figures/claw_mini_diffusion_eg.png. Note this is a smaller network than that used to generate the paper figures.
To recreate the claw machine figures in the paper, run ./run.sh or alternatively run each step separately:
python make_dataset.py
python train.py
python plot.py
This creates a figures directory containing the main claw machine result figures shown in the paper. This cycles through all methods, and uses the transformer denoising network, resulting in a run time >24 hours.
For nice labels when plotting, set IS_USE_LATEX=True, and ensure you have a valid tex/latex installation. To get it on Ubuntu run, sudo apt install texlive texlive-latex-extra texlive-latex-recommended dvipng cm-super msttcorefonts.
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.
Code is licensed under MIT, data and all other content is licensed under Microsoft Research License Agreement (MSR-LA). See LICENSE.