TorchWM Documentation

TorchWM is a modular PyTorch library for world models, latent-dynamics planning, and representation learning workflows (Dreamer, JEPA, IRIS, DiT, and more).

Documentation

• Getting Started
  • Installation
  • Logging with Weights & Biases and TensorBoard
  • Quick Start: Dreamer
  • Quick Start: Modular RSSM
  • Environment Backends
  • Typical Training Flow
• Package Overview
  • Core APIs
  • Environment Integration
  • World Model Building Blocks
  • Representation Learning and Diffusion
  • Data and Memory
  • Utilities
  • Which API Should I Use?

Algorithms

• Dreamer: Model-Based RL with Latent Dynamics
  • Key Idea
  • Architecture
  • Components
  • Training
  • DreamerV2 Enhancements
  • Environment Support
  • References
• JEPA: Joint Embedding Predictive Architecture
  • Key Idea
  • Architecture
  • Components
  • Training
  • Masking Strategies
  • Uses for Learned Representations
  • Comparison to Other Methods
  • References
• IRIS: Transformers for Sample-Efficient World Models
  • Key Idea
  • Architecture
  • Key Components
  • Training
  • Benchmark Results
  • Usage
  • References
• DiT: Diffusion Transformer
  • Key Idea
  • Architecture
  • Components
  • Training
  • Sampling (Generation)
  • Comparison to CNN-Based Diffusion
  • Applications
  • References

Reference

• API Reference
  • Core Public APIs
  • Dreamer
  • JEPA and ViT
  • IRIS (Sample-Efficient World Models)
  • Diffusion
  • Datasets and Transforms
  • Memory and Controllers
  • Utilities