How peizhuoli/ganimator works
Project type and purpose
• This is a PyTorch-based research library and demo application for neural motion synthesis (GANimator). It reads BVH motion files, trains a 1D GAN generator (and discriminator), and provides downstream tasks such as style transfer, motion mixing, keyframe editing, conditional generation, evaluation, and interactive demos.
File structure
• Top-level:
• bvh/ → BVH I/O and skeleton utilities (Quaternions, parser, writer, skeleton database)
• data/ → raw BVH sequences organized by character (Crab, Joe, Elephant, etc.)
• dataset/ → dataset/motion.py builds PyTorch Dataset from BVH files
• models/ → model definitions and helpers:
– architecture.py (generator, discriminator)
– gan1d.py (1D GAN modules)
– kinematics.py (forward/inverse FK)
– contact.py (foot contact modules)
– skeleton.py (skeleton graph)
– transforms.py (spatial/temporal data transforms)
– utils.py (masking, conditioning)
• train.py → training entry point (config via option.py)
• evaluate.py → quantitative evaluation module
• demo.py / demo.sh→ demo runner for generation, style transfer, mixing, key-frame editing, conditional gen
• fix_contact.py → post-processing to correct foot sliding
• interactive_utils.py → streaming input for interactive mode
• loss_recorder.py → logging losses to TensorBoard
• option.py → CLI options parser
• utils.py → shared utilities (file paths, logging)
README for additional insights
• Stack: Python ≥3.8, PyTorch ≥1.10, TensorBoard, SciPy
• Workflow described: environment setup, quick start with pretrained models, training instructions, conditional training, applications (mixing, style transfer, keyframe editing), evaluation prerequisites