Awesome

Planning with Diffusion   

Training and visualizing of diffusion models from Planning with Diffusion for Flexible Behavior Synthesis.

The main branch contains code for training diffusion models and planning via value-function guided sampling on the D4RL locomotion environments. The kuka branch contains block-stacking experiments. The maze2d branch contains goal-reaching via inpainting in the Maze2D environments.

Updates

• 12/09/2022: Diffuser (the RL model) has been integrated into 🤗 Diffusers (the Hugging Face diffusion model library)! See these docs for how to run Diffuser using their pipeline.
• 10/17/2022: A bug in the value function scaling has been fixed in this commit. Thanks to Philemon Brakel for catching it!

Quickstart

Load a pretrained diffusion model and sample from it in your browser with scripts/diffuser-sample.ipynb.

Installation

conda env create -f environment.yml
conda activate diffuser
pip install -e .

Using pretrained models

Downloading weights

Download pretrained diffusion models and value functions with:

./scripts/download_pretrained.sh

This command downloads and extracts a tarfile containing this directory to logs/pretrained. The models are organized according to the following structure:

└── logs/pretrained
    ├── ${environment_1}
    │   ├── diffusion
    │   │   └── ${experiment_name}
    │   │       ├── state_${epoch}.pt
    │   │       ├── sample-${epoch}-*.png
    │   │       └── {dataset, diffusion, model, render, trainer}_config.pkl
    │   ├── values
    │   │   └── ${experiment_name}
    │   │       ├── state_${epoch}.pt
    │   │       └── {dataset, diffusion, model, render, trainer}_config.pkl
    │   └── plans
    │       └── defaults
    │           ├── 0
    │           ├── 1
    │           ├── ...
    │           └── 149
    │
    ├── ${environment_2}
    │   └── ...

The state_${epoch}.pt files contain the network weights and the config.pkl files contain the instantation arguments for the relevant classes. The png files contain samples from different points during training of the diffusion model. Within the plans subfolders, there are the results of 150 evaluation trials for each environment using the default hyperparameters.

hopper-medium-replay-v2        | defaults   | logs/pretrained/hopper-medium-replay-v2/plans      | 150 scores
    93.6 +/- 0.37
hopper-medium-v2               | defaults   | logs/pretrained/hopper-medium-v2/plans             | 150 scores
    74.3 +/- 1.36
hopper-medium-expert-v2        | defaults   | logs/pretrained/hopper-medium-expert-v2/plans      | 150 scores
    103.3 +/- 1.30
walker2d-medium-replay-v2      | defaults   | logs/pretrained/walker2d-medium-replay-v2/plans    | 150 scores
    70.6 +/- 1.60
walker2d-medium-v2             | defaults   | logs/pretrained/walker2d-medium-v2/plans           | 150 scores
    79.6 +/- 0.55
walker2d-medium-expert-v2      | defaults   | logs/pretrained/walker2d-medium-expert-v2/plans    | 150 scores
    106.9 +/- 0.24
halfcheetah-medium-replay-v2   | defaults   | logs/pretrained/halfcheetah-medium-replay-v2/plans | 150 scores
    37.7 +/- 0.45
halfcheetah-medium-v2          | defaults   | logs/pretrained/halfcheetah-medium-v2/plans        | 150 scores
    42.8 +/- 0.32
halfcheetah-medium-expert-v2   | defaults   | logs/pretrained/halfcheetah-medium-expert-v2/plans | 150 scores
    88.9 +/- 0.25

\definecolor{tblue}{HTML}{1F77B4}
\definecolor{tred}{HTML}{FF6961}
\definecolor{tgreen}{HTML}{429E9D}
\definecolor{thighlight}{HTML}{000000}
\newcolumntype{P}{>{\raggedleft\arraybackslash}X}
\begin{table*}[hb!]
\centering
\small
\begin{tabularx}{\textwidth}{llPPPPPPPPr}
\toprule
\multicolumn{1}{r}{\bf \color{black} Dataset} & \multicolumn{1}{r}{\bf \color{black} Environment} & \multicolumn{1}{r}{\bf \color{black} BC} & \multicolumn{1}{r}{\bf \color{black} CQL} & \multicolumn{1}{r}{\bf \color{black} IQL} & \multicolumn{1}{r}{\bf \color{black} DT} & \multicolumn{1}{r}{\bf \color{black} TT} & \multicolumn{1}{r}{\bf \color{black} MOPO} & \multicolumn{1}{r}{\bf \color{black} MOReL} & \multicolumn{1}{r}{\bf \color{black} MBOP} & \multicolumn{1}{r}{\bf \color{black} Diffuser} \\ 
\midrule
Medium-Expert & HalfCheetah & $55.2$ & $91.6$ & $86.7$ & $86.8$ & $95.0$ & $63.3$ & $53.3$ & $\textbf{\color{thighlight}105.9}$ & $88.9$ \scriptsize{\raisebox{1pt}{$\pm 0.3$}} \\ 
Medium-Expert & Hopper & $52.5$ & $\textbf{\color{thighlight}105.4}$ & $91.5$ & $\textbf{\color{thighlight}107.6}$ & $\textbf{\color{thighlight}110.0}$ & $23.7$ & $\textbf{\color{thighlight}108.7}$ & $55.1$ & $103.3$ \scriptsize{\raisebox{1pt}{$\pm 1.3$}} \\ 
Medium-Expert & Walker2d & $\textbf{\color{thighlight}107.5}$ & $\textbf{\color{thighlight}108.8}$ & $\textbf{\color{thighlight}109.6}$ & $\textbf{\color{thighlight}108.1}$ & $101.9$ & $44.6$ & $95.6$ & $70.2$ & $\textbf{\color{thighlight}106.9}$ \scriptsize{\raisebox{1pt}{$\pm 0.2$}} \\ 
\midrule
Medium & HalfCheetah & $42.6$ & $44.0$ & $\textbf{\color{thighlight}47.4}$ & $42.6$ & $\textbf{\color{thighlight}46.9}$ & $42.3$ & $42.1$ & $44.6$ & $42.8$ \scriptsize{\raisebox{1pt}{$\pm 0.3$}} \\ 
Medium & Hopper & $52.9$ & $58.5$ & $66.3$ & $67.6$ & $61.1$ & $28.0$ & $\textbf{\color{thighlight}95.4}$ & $48.8$ & $74.3$ \scriptsize{\raisebox{1pt}{$\pm 1.4$}} \\ 
Medium & Walker2d & $75.3$ & $72.5$ & $\textbf{\color{thighlight}78.3}$ & $74.0$ & $\textbf{\color{thighlight}79.0}$ & $17.8$ & $\textbf{\color{thighlight}77.8}$ & $41.0$ & $\textbf{\color{thighlight}79.6}$ \scriptsize{\raisebox{1pt}{$\pm 0.55$}} \\ 
\midrule
Medium-Replay & HalfCheetah & $36.6$ & $45.5$ & $44.2$ & $36.6$ & $41.9$ & $\textbf{\color{thighlight}53.1}$ & $40.2$ & $42.3$ & $37.7$ \scriptsize{\raisebox{1pt}{$\pm 0.5$}} \\ 
Medium-Replay & Hopper & $18.1$ & $\textbf{\color{thighlight}95.0}$ & $\textbf{\color{thighlight}94.7}$ & $82.7$ & $\textbf{\color{thighlight}91.5}$ & $67.5$ & $\textbf{\color{thighlight}93.6}$ & $12.4$ & $\textbf{\color{thighlight}93.6}$ \scriptsize{\raisebox{1pt}{$\pm 0.4$}} \\ 
Medium-Replay & Walker2d & $26.0$ & $77.2$ & $73.9$ & $66.6$ & $\textbf{\color{thighlight}82.6}$ & $39.0$ & $49.8$ & $9.7$ & $70.6$ \scriptsize{\raisebox{1pt}{$\pm 1.6$}} \\ 
\midrule
\multicolumn{2}{c}{\bf Average} & 51.9 & \textbf{\color{thighlight}77.6} & \textbf{\color{thighlight}77.0} & 74.7 & \textbf{\color{thighlight}78.9} & 42.1 & 72.9 & 47.8 & \textbf{\color{thighlight}77.5} \hspace{.6cm} \\ 
\bottomrule
\end{tabularx}
\vspace{-.0cm}
\caption{
}
\label{table:locomotion}
\end{table*}

Planning

To plan with guided sampling, run:

python scripts/plan_guided.py --dataset halfcheetah-medium-expert-v2 --logbase logs/pretrained

The --logbase flag points the experiment loaders to the folder containing the pretrained models. You can override planning hyperparameters with flags, such as --batch_size 8, but the default hyperparameters are a good starting point.

Training from scratch

1. Train a diffusion model with:

python scripts/train.py --dataset halfcheetah-medium-expert-v2

The default hyperparameters are listed in locomotion:diffusion. You can override any of them with flags, eg, --n_diffusion_steps 100.

2. Train a value function with:

python scripts/train_values.py --dataset halfcheetah-medium-expert-v2

See locomotion:values for the corresponding default hyperparameters.

3. Plan using your newly-trained models with the same command as in the pretrained planning section, simply replacing the logbase to point to your new models:

python scripts/plan_guided.py --dataset halfcheetah-medium-expert-v2 --logbase logs

See locomotion:plans for the corresponding default hyperparameters.

Deferred f-strings. Note that some planning script arguments, such as --n_diffusion_steps or --discount, do not actually change any logic during planning, but simply load a different model using a deferred f-string. For example, the following flags:

---horizon 32 --n_diffusion_steps 20 --discount 0.997
--value_loadpath 'f:values/defaults_H{horizon}_T{n_diffusion_steps}_d{discount}'

will resolve to a value checkpoint path of values/defaults_H32_T20_d0.997. It is possible to change the horizon of the diffusion model after training (see here for an example), but not for the value function.

Docker

1. Build the image:

docker build -f Dockerfile . -t diffuser

2. Test the image:

docker run -it --rm --gpus all \
    --mount type=bind,source=$PWD,target=/home/code \
    --mount type=bind,source=$HOME/.d4rl,target=/root/.d4rl \
    diffuser \
    bash -c \
    "export PYTHONPATH=$PYTHONPATH:/home/code && \
    python /home/code/scripts/train.py --dataset halfcheetah-medium-expert-v2 --logbase logs"

Singularity

1. Build the image:

singularity build --fakeroot diffuser.sif Singularity.def

2. Test the image:

singularity exec --nv --writable-tmpfs diffuser.sif \
        bash -c \
        "pip install -e . && \
        python scripts/train.py --dataset halfcheetah-medium-expert-v2 --logbase logs"

Running on Azure

Setup

1. Tag the Docker image (built in the Docker section) and push it to Docker Hub:

export DOCKER_USERNAME=$(docker info | sed '/Username:/!d;s/.* //')
docker tag diffuser ${DOCKER_USERNAME}/diffuser:latest
docker image push ${DOCKER_USERNAME}/diffuser

3. Update azure/config.py, either by modifying the file directly or setting the relevant environment variables. To set the AZURE_STORAGE_CONNECTION variable, navigate to the Access keys section of your storage account. Click Show keys and copy the Connection string.

4. Download azcopy: ./azure/download.sh

Usage

Launch training jobs with python azure/launch.py. The launch script takes no command-line arguments; instead, it launches a job for every combination of hyperparameters in params_to_sweep.

Viewing results

To rsync the results from the Azure storage container, run ./azure/sync.sh.

To mount the storage container:

1. Create a blobfuse config with ./azure/make_fuse_config.sh
2. Run ./azure/mount.sh to mount the storage container to ~/azure_mount

To unmount the container, run sudo umount -f ~/azure_mount; rm -r ~/azure_mount

Reference

@inproceedings{janner2022diffuser,
  title = {Planning with Diffusion for Flexible Behavior Synthesis},
  author = {Michael Janner and Yilun Du and Joshua B. Tenenbaum and Sergey Levine},
  booktitle = {International Conference on Machine Learning},
  year = {2022},
}

Acknowledgements

The diffusion model implementation is based on Phil Wang's denoising-diffusion-pytorch repo. The organization of this repo and remote launcher is based on the trajectory-transformer repo.