How garvita-tiwari/sizer works
Project’s Main Components
• Data & Configuration:
• The “data” directory includes data pre-processing modules (e.g., geometry utilities) that handle 3D garment information.
• “configs” holds configuration files (YAML files for ParserNet and SizerNet) that drive experiments.
• Global environment variables and dataset directory settings are defined in “utils/global_var.py”.
• Model & Training:
• The “models” directory is at the core: it includes modules for loading data, setting up the loss functions, defining the neural network architecture (under models/network), and separate scripts for training ParserNet and SizerNet.
• “trainer.py” is the central entry point that, based on a given configuration, dispatches model training.
• “generate.py” handles evaluation or generation tasks.
• SMPL Utilities:
• The “smpl_lib” folder contains SMPL-related routines (e.g., pose mapping, linear blend skinning, geometric computations) which are essential for handling 3D garment deformations.
• Utilities & Visualisation:
• “utils” contains supplementary functions (e.g., geometry utilities, SMPL paths).
• “visualisation” is likely for rendering or displaying model outputs and intermediate results.
• Supporting Files:
• “requirements.txt” defines dependencies.
• “configs” and other scripts allow switching between training and evaluation (ParserNet vs. SizerNet) via command-line arguments.
Architectural Patterns / Design Principles
• Modular Design: The project is structured into clearly separated modules (data, models, configuration, SMPL library, utilities, and visualisation) allowing clear boundaries between data processing, model definition, and evaluation.
• Configuration Driven: The use of YAML config files enables different experiment setups that can directly modify the behavior of training/evaluation scripts.
• Domain-Specific Abstraction: The SMPL library represents an abstraction for handling 3D garment synthesis tasks, which should be clearly highlighted as a core utility shared among processing tasks.