Ryan-J-Gilbert/Lacuna-Solar-Survey-Challenge

Some of my work for the Lacuna Solar Survey Challenge hosted by Zindi.africa # Solar Panel and Boiler Counting Model This repository contains a deep learning pipeline for counting solar panels and boilers in aerial images. The model integrates image features with metadata and employs advanced techniques such as cross-attention, test-time augmentation (TTA), and calibration for optimal performance. ## Pipeline Overview 1. **Data Preprocessing**: - Images are resized and normalized. - Metadata (e.g., image origin, placement) is encoded using one-hot encoding. 2. **Model Architecture**: - **Backbone**: EfficientNetV2 is used for feature extraction from images. - **Metadata Processor**: A fully connected network processes metadata. - **Cross-Attention**: Combines image and metadata features using multi-head attention. - **Counting Head**: A custom regressor predicts the number of solar panels and boilers. 3. **Training**: - Uses a combination of Huber Loss and L1 Loss for robust training. - Employs advanced augmentation techniques to improve generalization. - Implements early stopping and a OneCycleLR scheduler for efficient training. 4. **Validation**: - Performs cross-validation with K-Fold splitting. - Optimizes rounding thresholds for better integer predictions. 5. **Inference**: - Supports Test-Time Augmentation (TTA) to improve robustness. - Applies calibration thresholds to refine predictions. 6. **Submission**: - Generates three types of submissions: - Raw predictions. - Integer predictions (standard rounding). - Calibrated predictions (optimized thresholds). ## Features - **Metadata Integration**: Combines image and metadata features for improved accuracy. - **Cross-Attention**: Enhances feature interaction between image and metadata. - **Advanced Augmentation**: Includes geometric and color transformations. - **Test-Time Augmentation (TTA)**: Averages predictions over multiple augmented versions of test images. - **Calibration**: Optimizes rounding thresholds for better integer predictions. ## Requirements - Python 3.8+ - P …