How ism-ail11/snap-uq works
SNAP-UQ is a research toolkit for training compact image/audio classifiers, fitting self-supervised next-activation predictors, and evaluating uncertainty offline or on streams. The typical workflow configures data and a backbone, trains the backbone, trains SNAP-UQ heads, optionally calibrates scores, then evaluates detection or selective risk. Baselines and experiment runners support the paper’s comparisons and ablations. Source-index references and README statements support the unsampled branches; those branches are represented at their exact enclosing directories rather than implying inspected implementation details.
Main entry points: preprocess Data preprocessing · vision.py Vision datasets · experiments Experiment configs · snapuq.py SNAP-UQ predictors
Components 20
Data
- Data preprocessing
scripts/preprocess - Vision datasets
snapuq/data/vision.py - Audio datasets
snapuq/data/audio.py - Corruption suites
snapuq/data/corruptions
Models and training
- Backbone models
snapuq/models - Model factory
snapuq/models/factory.py - Classifier training
snapuq/train.py - Predictor training
snapuq/train.py - Runtime utilities
snapuq/utils
Uncertainty and evaluation
- SNAP-UQ predictors
snapuq/snapuq.py - Uncertainty baselines
snapuq/baselines.py - Score calibration
snapuq/calibration.py - Offline evaluation
snapuq/eval.py - Detection metrics
snapuq/metrics.py - Streaming evaluation
snapuq/streaming.py
Research workflows
- Experiment configs
configs/experiments - Ablation runner
scripts/ablations/run_ablations.py - Appendix runner
scripts/appendix/run_appendix.py - Results runner
scripts/results/run_results.py
Other components
- Researcher
Connections 23
- Researcher → Experiment configs: configures
README.md - Researcher → Data preprocessing: prepares data
README.md - Experiment configs → Model factory: selects model
- Model factory → Backbone models: constructs
snapuq/models/factory.py - Data preprocessing → Vision datasets: prepares inputs
README.md - Data preprocessing → Audio datasets: prepares inputs
README.md - Vision datasets → Corruption suites: supports corruption data
README.md - Audio datasets → Runtime utilities: imports safe helpers
snapuq/data/audio.py - Vision datasets → Runtime utilities: imports safe helpers
snapuq/data/vision.py - Classifier training → Backbone models: trains classifier
snapuq/train.py - Classifier training → Runtime utilities: uses training helpers
snapuq/train.py - Experiment configs → SNAP-UQ predictors: configures through builder
- Predictor training → SNAP-UQ predictors: optimizes predictors
snapuq/train.py - Researcher → Predictor training: trains heads
README.md - SNAP-UQ predictors → Score calibration: maps scores
- Offline evaluation → Detection metrics: computes measures
snapuq/eval.py - Streaming evaluation → Detection metrics: reports event measures
README.md - Researcher → Offline evaluation: evaluates offline
README.md - Researcher → Streaming evaluation: evaluates streams
README.md - Researcher → Uncertainty baselines: compares methods
README.md - Researcher → Ablation runner: runs ablations
README.md - Researcher → Appendix runner: runs experiments
README.md - Researcher → Results runner: runs results