How edward1997104/wavelet-generation works
File structure
• Top-level:
– configs/ (config.py, config_highs.py)
– data/ (flip_axis.py, to_manifold.py, sample_sdf.py, convert_wavelet.py)
– external/ (Manifold, mesh_to_sdf)
– wavelet_data/ (storage for generated .npy files)
– trainer/ (trainer.py)
– models/
· module/ (diffusion_network.py, gaussian_diffusion.py, dwt.py, etc.)
· network.py (training entry model)
· network_gen.py (inference entry)
– pretrain/ (pretrained checkpoints)
– utils/ (debugger.py, meter.py, other_utils.py)
– debug/ (output_folder)
README for additional insights
• Describes environment setup, dataset preparation steps (ShapeNet → data scripts → wavelet_data), training (trainer.py + config), and inference (network_gen.py + pretrain).
• External dependencies: PyTorch Wavelets, Manifold, mesh_to_sdf.
Main components
A. Raw Dataset (ShapeNet)
B. Data-Processing Pipeline (scripts in data/ plus external libs)
• flip_axis.py → orientation fix
• to_manifold.py → watertight mesh (uses external/Manifold)
• sample_sdf.py → SDF sampling (uses external/mesh_to_sdf)
• convert_wavelet.py → wavelet coefficients → output to wavelet_data/
C. Data Store: wavelet_data/ (coarse & detail .npy)
D. Configuration Layer: configs/config.py & config_highs.py
E. Training Engine: trainer/trainer.py
• Loads .npy via config
• Uses models/network.py & models/module/* to train diffusion & detail predictor
• Outputs checkpoints to pretrain/ or user‐defined folder
F. Model Modules:
• Core building blocks (diffusion, resample, DWT, FP16 utilities)
• network.py (assembles modules)
G. Pretrained Model Store: pretrain/
H. Inference Engine: models/network_gen.py
• Loads checkpoints & config params
• Generates wavelet coefficients → reconstructs mesh → writes to debug/output_folder or user folder
I. Utilities: utils/ (logging, debugging, metrics)
J. External Dependencies:
• pytorch_wavelets, external/Manifold, external/mesh_to_sdf
Relationships & interactions
• Utilities are tapped into by both Trainer and Inference for logging/metrics.
• External libs are referenced only by specific data scripts.
Architectural patterns & principles
• Modular pipeline separation (Data-Prep ↔ Training ↔ Inference)
• Config-driven workflows decoupling code from hyperparameters
• Plugin-style model modules (easy to swap or extend building blocks)
• CLI orchestration (Python entry scripts in each stage)