How gjeusel/deeproof works
DeepRoof is a PyTorch roof-orientation classifier. The main workflow configures a model and image transforms, builds training and validation loaders from labeled data, trains and scores the model, saves the best checkpoint, then loads it to predict the submission examples and write a CSV. A separate downloader retrieves satellite imagery for those examples. The excerpts show the handler owning training, validation, persistence, and prediction calls; unsampled modules and runtime wiring are not inferred.
Main entry points: model_handler.py Workflow handler · main.py Run configuration · dataset.py Image dataset · common.py Paths and labels
Components 16
Data preparation
- Image downloader
data/download_images.py - Labeled data
data/train.csv - Submission examples
data/sample_submission.csv - Roof images
- Image dataset
src/deeproof/dataset.py
Model learning
- Workflow handler
src/deeproof/model_handler.py - ResNet and ShortNet
src/deeproof/neuro/handcraft.py - Dual Path Network
src/deeproof/neuro/dpn.py - Classification metrics
src/deeproof/metrics.py - Training loop
src/deeproof/train.py - Validation scoring
src/deeproof/validation.py
Persistence and output
- Model registry
src/deeproof/database_models.py - Prediction
src/deeproof/prediction.py - Paths and labels
src/deeproof/common.py
Other components
- Run configuration
src/deeproof/main.py - Satellite imagery
Connections 19
- Satellite imagery → Image downloader: serves tiles
data/download_images.py - Image downloader → Roof images: writes images
data/download_images.py - Run configuration → Workflow handler: configures
src/deeproof/main.py - Run configuration → ResNet and ShortNet: selects model
src/deeproof/main.py - Run configuration → Dual Path Network: imports option
src/deeproof/main.py - Run configuration → Classification metrics: imports loss
src/deeproof/main.py - Workflow handler → Image dataset: builds loaders
src/deeproof/model_handler.py - Image dataset → Labeled data: reads labels
src/deeproof/dataset.py - Image dataset → Roof images: loads images
src/deeproof/dataset.py - Workflow handler → Training loop: calls training
src/deeproof/model_handler.py - Workflow handler → Validation scoring: calls validation
src/deeproof/model_handler.py - Workflow handler → Model registry: checks and saves
src/deeproof/model_handler.py - Workflow handler → Submission examples: loads examples
src/deeproof/model_handler.py - Workflow handler → Prediction: calls prediction
src/deeproof/model_handler.py - Workflow handler → Prediction: writes submission
src/deeproof/model_handler.py - Validation scoring → Classification metrics: imports scoring
src/deeproof/validation.py - Model registry → Paths and labels: uses snapshot path
src/deeproof/database_models.py - Workflow handler → Paths and labels: uses data paths
src/deeproof/model_handler.py - Run configuration → Paths and labels: sets up logging
src/deeproof/main.py