How anuragmundhada/pose-hgreg-gan works
Project type and purpose
• This is a research‐oriented machine‑learning pipeline for 3D human‑pose estimation using a GAN (generator + discriminator) approach in PyTorch, providing weak depth supervision when full 3D ground truth is unavailable.
File structure
• Top‑level “data/” holds raw datasets (Human3.6M, MPII) with annotations.
• “exp/” is the experiment output root (checkpoints, logs).
• “src/” contains:
• datasets/ – data loaders and wrappers (h36m.py, mpii.py, fusion.py)
• models/ – network definitions (hg_3d.py, hg_3d_gan.py, discriminators.py) and custom layers (Residual, FusionCriterion, adversarial losses)
• train.py & demo.py – entry points for training and inference
• main.py – orchestrates parsing options (opts.py), kicking off train/ref processes
• utils/ – logging, image/pose utilities, evaluation metrics
• tools/ – dataset download/preparation scripts
• Config and dependency hints: readme.md (PyTorch, h5py, OpenCV, tensorboard), .gitignore, LICENSE.
README
• Confirms GAN approach for weak supervision, based on Zhou et al.’s PyTorch pose code.
• Lists deep‑learning stack and optional tensorboard.
1. Data Sources
• raw H36M & MPII folders
• annotation files (.h5, .mat, .txt)
2. Data Loader & Preprocessing
• src/datasets modules
• normalization, augmentation (fusion.py)
3. Model
• Generator (hg_3d.py)
• Discriminator (discriminators.py)
• Custom Layers (Residual, FusionCriterionAdversarial)
4. Training Engine
• train.py orchestrating forward/backward passes
• Loss computation (L2, adversarial)
• Optimizer step (SGD/Adam)
5. Evaluation & Metrics
• src/utils/eval.py
• tensorboard summaries / progress bar
6. Experiment Management
• src/utils/logger.py
• exp/ checkpoint storage
7. Demo / Inference
• demo.py for visualization of predicted 3D pose
8. Supporting Tools
• tools/GetH36M.py
• Data Sources → Data Loader
• Data Loader → Model (Generator + Discriminator)
• Model outputs & ground‑truth → Loss Module
• Loss Module → Training Engine → Model weights
• Training Engine → Experiment Manager (logging, checkpoints)
• Model weights → Demo / Inference
c. Architectural patterns & principles:
• Layered modular structure (separate folders for datasets, models, utils)
• Clear separation of concerns: data ingestion, model definition, training loop, evaluation
• Reusable components: custom layers and loss criteria factored into their own modules
• Script‑based orchestration (main.py, train.py, demo.py)
d. Key technologies & frameworks:
• PyTorch (core deep‑learning)
• Python h5py, OpenCV (data handling)
• Optional TensorBoard for visualization
• Standard Python packaging and script invocation