How pranoy-panda/causal-feature-subset-selection works
Project type and purpose
• It runs as a command-line pipeline (via main.py), processes vision datasets (MNIST, CIFAR, FMNIST), loads a pretrained model, computes causal feature subsets, evaluates them, and exports visualization results.
File structure
• Top-level orchestrator: main.py
• Configuration manager: config.py (datasets, hyperparameters, paths)
• Dataset loader and preprocessor: dataset_file.py
• Core algorithms and utilities: utils.py (causal metric computation, mask generation, plotting, evaluation)
• Results output directory: results/ (images saved from experiments)
• Dependency manifest: requirements.txt
README for additional insights
• External dependencies: PyTorch, NumPy, Matplotlib, skimage, captum.
a. Main components to represent
– Configuration Manager (config.py)
– Data Ingestion & Preprocessing (dataset_file.py)
– Causal Feature Selector (core routines in utils.py)
– Evaluation Module (post-hoc accuracy, Average Causal Effect in utils.py)
– Visualization & Results Generator (plotting routines in utils.py → results/)
– External Libraries & Frameworks (PyTorch, NumPy, skimage, captum, Matplotlib)
b. Relationships and data flow
– Configuration Manager feeds hyperparameters and dataset/model choices into Data Ingestion and Model Loader.
– Causal Selector computes feature subsets (via calls to captum or custom utility functions).
– Evaluation Module measures metrics on the selected subset vs. original output.
– Visualization Module generates images of selected superpixels and writes to results/.
c. Architectural patterns and design principles
– Separation of concerns: each module handles a distinct responsibility.
– Modular and extensible: swapping models, datasets, or causal metrics by editing config.py.
d. Key technologies and libraries
– Core language: Python 3.5+
– Deep learning: PyTorch
– Array & math: NumPy, scikit-image
– Interpretation: captum
– Plotting: Matplotlib