How jannerm/diffuser works
Project type and purpose
• This is an open-source research tool for diffusion-based planning in offline reinforcement learning (RL).
• It provides functionality for:
• Training diffusion models on D4RL locomotion datasets
• Training value functions
• Generating action plans via guided sampling
• Visualizing results and integrating with cloud and HPC back-ends
File structure
Top-level directories:
• diffuser/ ← core Python package
• datasets/ ← data buffers, normalization, sequence handling
• environments/ ← MuJoCo/PyBullet environment wrappers and XML assets
• models/ ← diffusion and temporal model definitions
• sampling/ ← sampling functions, guides, policies
• utils/ ← I/O, logging, rendering, training loops, serialization, cloud helpers
• config/ ← default hyperparameter definitions for diffusion, values, plans
• scripts/ ← CLI entrypoints (train.py, train_values.py, plan_guided.py, plotting notebooks, download helpers)
• plotting/ ← score plotting and table-generation scripts
• azure/, gcp/, slurm/ ← cloud/HPC orchestration, launch & sync scripts for Azure, GCP, and Slurm
• Dockerfile, Singularity.def ← container definitions
• environment.yml, setup.py ← environment and packaging
• logs/ (created at runtime) ← storage of model checkpoints, configs, sample images, plan outputs
README for additional insights
• Core workflows:
a. Download or train diffusion model (scripts/train.py)
b. Train value function (scripts/train_values.py)
c. Plan via guided sampling (scripts/plan_guided.py)
• Integration: Docker, Singularity, Azure (using azure/launch.py), GCP, and Slurm scripts.
• Dependencies: D4RL, mujoco/pybullet, PyTorch, Hugging Face Diffusers integration.
Detail
• Submodules for clarity (e.g., within Core: datasets, models, sampling, utils)
• Both local and remote execution paths (Docker/Singularity vs. Azure/GCP/Slurm)
• Storage of checkpoints, logs, and upload/download via azure/utils or gcp scripts