How dhiraj0-del/galaxai works
GALAXAI is a research system for multi-attribute galaxy morphology classification. Its user-facing workflow accepts an uploaded HST image, runs the trained model, and presents morphology predictions; its research workflow prepares multi-task data, trains a CNN, and evaluates results. The graph also retains the documented architecture-comparison objective and legacy pipeline. Source excerpts establish the UI-to-inference and current model/training relationships; unsampled data-preparation and legacy internals are represented conservatively.
Main entry points: train_multitask.py Training orchestration · models.py Multi-task CNN · inference.py Image inference · prepare_multitask_dataset.py Dataset preparation
Components 13
Image analysis
- Streamlit interface
app.py
Model runtime
- Image inference
src/inference.py - Multi-task CNN
src/models.py - Trained weights
checkpoints/galaxai_model1_v2_best.pth
Research workflow
- Dataset preparation
src/prepare_multitask_dataset.py - Training orchestration
src/train_multitask.py - Dataset and loss
src/multitask_training.py - Model evaluation
src/evaluate_model1.py - Class weights
src/calculate_class_weights.py - Target analysis
src/analyze_10_targets.py - Split validation
src/check_multitask_splits.py
Legacy pipeline
- Earlier CNN pipeline
archive/old_pipeline
Other components
- Image user
Connections 13
- Image user → Streamlit interface: uploads image
README.md - Streamlit interface → Image inference: requests predictions
app.py - Image inference → Multi-task CNN: runs model
src/inference.py - Image inference → Trained weights: loads weights
src/inference.py - Training orchestration → Multi-task CNN: trains model
src/train_multitask.py - Training orchestration → Dataset and loss: uses training utilities
src/train_multitask.py - Model evaluation → Multi-task CNN: evaluates model
src/evaluate_model1.py - Model evaluation → Dataset and loss: uses dataset utilities
src/evaluate_model1.py - Dataset preparation → Training orchestration: supplies dataset
- Class weights → Training orchestration: supplies weights
- Target analysis → Dataset preparation: analyzes targets
- Split validation → Dataset preparation: checks splits
- Earlier CNN pipeline → Multi-task CNN: uses earlier models
archive/old_pipeline/train.py