How bg-bibek/master-thesis-volcano-detection works
This repository is a research toolkit for training and evaluating temporal CNN classifiers on Thalia volcanic-deformation WebDataset shards. The main workflow loads and normalizes shard samples, trains a classifier, and evaluates predictions; separate tools support unseen-data evaluation, threshold selection, permutation test-time augmentation, Grad-CAM localization, and temporal-pattern analysis. The graph also includes shard inspection and results analysis. Relationships are limited to calls or dependencies shown in excerpts and SOURCE INDEX references; some standalone scripts have no evidenced links.
Main entry points: train.py Training runner · evaluate_unseen.py Unseen-data evaluation · data_loader_fixed.py Shard data loader · analyze_temporal_patterns.py Temporal pattern analysis
Components 15
Data handling
- Shard data loader
data_loader_fixed.py - Shard inspector
inspect_shards.py
Model training
- Training runner
train.py - Focal loss
focal_loss.py
Evaluation
- Unseen-data evaluation
evaluate_unseen.py - Threshold sweep
threshold_sweep.py - Permutation TTA
permutation_tta.py
Research analysis
- Grad-CAM localization
gradcam.py - Grad-CAM comparison
compare_gradcam.py - Temporal pattern analysis
analyze_temporal_patterns.py - Results table builder
make_results_table.py - Log parser
parse_logs.py
Other components
- Researcher
- WebDataset shards
- Evaluation outputs
Connections 24
- Researcher → Training runner: starts training
- WebDataset shards → Shard data loader: supplies samples
data_loader_fixed.py - Training runner → Shard data loader: loads batches
train.py - Training runner → Focal loss: uses loss
train.py - Researcher → Shard inspector: inspects shards
inspect_shards.py - Researcher → Unseen-data evaluation: evaluates checkpoint
evaluate_unseen.py - Unseen-data evaluation → Shard data loader: loads samples
evaluate_unseen.py - Unseen-data evaluation → Training runner: reuses model code
evaluate_unseen.py - Researcher → Threshold sweep: sweeps threshold
threshold_sweep.py - Threshold sweep → Training runner: reuses model code
threshold_sweep.py - Threshold sweep → Shard data loader: creates loaders
threshold_sweep.py - Researcher → Permutation TTA: runs TTA
permutation_tta.py - Permutation TTA → Unseen-data evaluation: uses evaluation utilities
permutation_tta.py - Researcher → Grad-CAM localization: generates localization
gradcam.py - Grad-CAM localization → Shard data loader: loads samples
gradcam.py - Grad-CAM localization → Training runner: builds classifier
gradcam.py - Researcher → Temporal pattern analysis: analyzes patterns
analyze_temporal_patterns.py - Temporal pattern analysis → Shard data loader: loads samples
analyze_temporal_patterns.py - Temporal pattern analysis → Training runner: loads model state
analyze_temporal_patterns.py - Temporal pattern analysis → Unseen-data evaluation: uses evaluation utilities
analyze_temporal_patterns.py - Researcher → Grad-CAM comparison: compares localization
compare_gradcam.py - Researcher → Results table builder: builds results table
make_results_table.py - Researcher → Log parser: parses logs
parse_logs.py - Unseen-data evaluation → Evaluation outputs: reports metrics