How chanyoung-ken/ai_proj works
Project Type and Purpose
• This is an ML research/pipeline project focused on training and evaluating robust image-classification models (ResNet variants) on CIFAR-10 and CIFAR-100. It includes modules for data generation, model training (including adversarially trained variants), hyperparameter tuning, ensemble construction, robustness evaluation, and visualization (Grad-CAM, plotting).
File-Structure Analysis
• Top-level scripts and orchestration:
• run_pipeline.sh, run_at_alp.sh, run_ensemble.sh, run_epsilon_sweep.sh – bash drivers that chain together data preparation, training, ensembling, and evaluation.
• benchmark.py, hyperparameter_tuning.py, evaluate_models.py, multi_ensem_logs/… – standalone Python scripts for specific tasks.
• Key directories:
• data_generator/ – contains data_generator.py (creates synthetic/adversarial data), train_resnet.py (model training loop), pipeline.ipynb for exploration, checkpoints and models subfolders.
• ensemble_models/ – stores final ensemble weights.
• robustness_results/, ensemble_robustness_results/ – CSVs and model files capturing evaluation metrics under different attacks.
• gradcam_visualizations/, graphs/ – visual output of model explanations and performance plots.
• Logging and artifacts:
• logs/, multi_ensem_logs/, nohup.out, sweep_output.log – track runtime logs by component.
README Insights
• Minimal README; primary clues come from file names and directory layout. The goal is end-to-end robustness research: data → single-model training → ensemble assembly → evaluation → visualization.
a. Main Components
1. Data Ingestion & Preprocessing (data_generator/data_generator.py)
2. Model Training Service
– Standard training (train_resnet.py)
– Adversarial training (attentionlpt.py, ensemble_multi_attak.py)
– Hyperparameter tuner (hyperparameter_tuning.py)
3. Ensemble Orchestrator (realensemble.py, ensemble_models/)
4. Evaluation & Robustness Analysis (evaluate_models.py, robustness_results/, ensemble_robustness_results/)
5. Visualization & Reporting (gradcam.py → gradcam_visualizations/, plotting.py → graphs/)
6. Orchestration Scripts (run_*.sh)
7. Artifact Storage (checkpoints/, models/, logs/, CSV result folders)
b. Relationships & Data Flows
– Raw CIFAR data → Data Ingestion → Preprocessed data/augmented adversarial examples
– Preprocessed data → Model Training → Checkpointed weights
– Checkpoints → Ensemble Orchestrator → ensemble_models/*.pth
– Single and ensemble models → Evaluation module → robustness_results/*.csv
– Raw evaluation results → Plotting/Grad-CAM modules → graphs/ and gradcam_visualizations/
– Orchestration scripts tie above modules in pipelines and log to logs/ and multi_ensem_logs/
c. Architectural Patterns & Principles
– Modular pipeline: each step isolated in its own script or directory for reusability.
– Artifact versioning: checkpoints and logs organized by timestamped folders.
– Separation of concerns: training, ensembling, evaluation, and visualization are distinct components.
d. Technologies & Frameworks
– Python 3.x, PyTorch for model definition and training
– NumPy/Pandas for data handling
– Matplotlib/Seaborn (in plotting.py) for visualizations
– Bash scripts for orchestration
– Jupyter notebooks for exploratory analysis (pipeline.ipynb)
Level-of-Detail
• Every major directory as a component.
• Interactions between training and evaluation through shared model artifacts.
• External dependencies (PyTorch library, adversarial-attack utilities).