How runa91/bite_release works
File structure
• Top‐level “scripts” directory holds entry-point scripts for training (train.py, train_withref.py), inference (full_inference_including_ttopt.py), and demo (gradio_demo.py).
• “src” contains modular packages:
• bps_2d (body part segmentation snippets)
• combined_model (image → 3D prediction and refinement losses)
• graph_networks (graph CNN architectures and mesh utilities)
• smal_pytorch (differentiable renderer and SMAL animal model)
• stacked_hourglass (2D keypoint detection model, dataset loaders, loss, train/predict code)
• test_time_optimization (bite inference model plus TT-opt utilities)
• configs (YAML/JSON files for paths, hyperparameters, loss weights)
• metrics (evaluation metrics)
• Environment and dependency files: requirements.txt, packages.txt, plus standard Git config.
README for insights
• Confirms a two-stage training pipeline: first BARC+ (2D→3D breed‐aware regressor), then a refinement network (full BITE).
• Data preparation instructions show where “checkpoint,” “data,” and “datasets” live at runtime.
• Usage sections outline separate flows for training, inference, and a Gradio demo.