How mirkozaff/aa-sgan works
Project type and purpose
• This is a research‐oriented deep learning project implementing AA-SGAN, an adversarial augmentation framework for pedestrian trajectory prediction.
• It builds on the Social GAN architecture and integrates synthetic data augmentation at training time.
File structure
• Top level:
• src/ (all code, scripts, datasets, checkpoints)
• Under src/:
• datasets/ – raw, preprocessed, train/val/test split for ETH, HOTEL, UNIV, ZARA1/2, plus JTA synthetic data
• checkpoint/ – model snapshots for four training modes (baseline, hybrid, jta, traj-gen) across five scenes
• sgan/ – core library: data loader, trajectory utilities, model definitions (G, D, A modules), loss functions, training steps, utils
• scripts:
– extract_trajectory.py – preprocess raw trajectory files into dataset splits
– evaluate_model.py – load checkpoints & compute metrics on held-out sets
– train_SGAN.py / train_SGAN_GEN.py – orchestrate the adversarial training loops
– aasgan_extract.sh / aasgan_train.sh – wrapper bash scripts to launch extraction and training
• requirements.txt – Python dependencies (e.g., PyTorch, NumPy)
README for additional insights
• The system has three logical model components—Augmenter (A), Generator (G), Discriminator (D)—arranged as an adversarial pipeline.
• Synthetic trajectories from the JTA dataset are fed into A to produce “synth-augmented” samples, which are then used alongside real trajectories for G/D training.
Final advice
• Break the system into the smallest functional units: data ingestion, augmentation, generation, discrimination, checkpointing, evaluation.
• Orchestration (bash scripts) at the periphery initiating the Extract → Train → Evaluate pipeline.