How chetrajpandey/fulldiskattention works
This repository builds full-disk solar magnetogram classifiers to predict whether an M1.0-or-greater flare will occur in the next 24 hours. The documented workflow obtains and converts HMI images, labels timestamped images from GOES flare records, trains attention or standard-CNN models using four-fold datasets, and analyzes predictions and attention maps. The graph preserves the documented analysis branches; source excerpts establish training, labeling, model, and download behavior, while notebook wiring is based on README descriptions.
Main entry points: train.py Training runner · trained_models Trained weights · download_jp2.py Download and convert · attention_model.py CNN and attention
Components 19
Image acquisition
- JP2 magnetograms
- Download and convert
download_mag/download_jp2.py - JPG magnetograms
Flare labeling
- GOES flare records
data_labeling/data_source/goes_flares_integrated.csv - Label and split
data_labeling/labeling.py - Four-fold labels
data_labeling/data_labels/simplified_data_labels
Model training
- Training runner
modeling/train.py - Image datasets
modeling/dataloader.py - CNN and attention
modeling/attention_model.py - Attention blocks
modeling/blocks.py - Weight initialization
modeling/initialize.py - Skill metrics
modeling/evaluation.py - Trained weights
modeling/trained_models - Fold results
modeling/trained_models
Result analysis
- Attention predictions
result_analysis/Attention_Predictions.ipynb - CNN predictions
result_analysis/No_Attention_Predictions.ipynb - Attention maps
result_analysis/Attention_Maps_Visualize.ipynb - Plot locations
result_analysis/plots.csv
Other components
- Helioviewer
Connections 20
- Helioviewer → Download and convert: provides images
- Download and convert → JP2 magnetograms: downloads JP2
- Download and convert → JPG magnetograms: converts images
- GOES flare records → Label and split: supplies flare data
- Label and split → Four-fold labels: writes labels
- Four-fold labels → Training runner: provides fold CSVs
- JPG magnetograms → Image datasets: provides images
- Training runner → Image datasets: constructs datasets
- Training runner → CNN and attention: runs classifier
- CNN and attention → Attention blocks: uses attention blocks
- CNN and attention → Weight initialization: initializes weights
- Training runner → Skill metrics: computes skill scores
- Training runner → Trained weights: saves model weights
- Training runner → Fold results: writes fold results
- Trained weights → Attention predictions: evaluates attention models
- Fold results → Attention predictions: supplies results
- Trained weights → CNN predictions: evaluates CNN models
- Fold results → CNN predictions: supplies results
- Trained weights → Attention maps: visualizes attention
- Plot locations → Attention maps: supplies instances