How markos-diomataris/wandr works
Project type and purpose
• This is a research‐oriented Python machine‐learning pipeline (WANDR) for intention‐guided human motion generation.
• It provides data ingestion (SMPL-X body model), configuration via Hydra, model definition (CVAE, motion prior), training/evaluation loops, and rendering of generated motion.
File structure
• conf/ (Hydra config hierarchy): dataset, dataloader, model, optimizer, scheduler, logger, rendering, evaluator, paths, train/evaluate settings, custom callbacks.
• dataset/, dataloader, data/: raw data definitions, parsing, feature extraction.
• models/: abstract base modules and concrete implementations (motion_model, CVAE).
• motion_filter/: filters applied to raw motion sequences.
• schedulers/, utils/: supporting code (custom LR schedulers, math transforms, SMPL utilities).
• evaluators/: motion quality metrics.
• rendering/: utilities for visualizing and exporting motion.
• scripts/: environment setup (SMPL-X extraction, venv).
• Top‐level entry points: train.py (training), evaluate.py (quantitative evaluation), demo.py (inference + rendering), calculate_stats.py.
Extracted insights from the README
• External dependency: SMPL-X body model must be downloaded separately and put under data/body_models.
• setup.sh extracts the model and creates a virtual environment.
• demo.py produces an output.mp4 of generated motion.
• The project uses Hydra for config management, PyTorch (implied by modules), and optionally Weights & Biases for logging.
Diagramming recommendations
– Blue: Data (SMPL-X, deps, Dataset)
– Green: Core compute (Model, Filter, Training Loop)
– Orange: Configuration & Logging (Hydra, WandB)
– Purple: Output & Visualization (Evaluation, Rendering)