codejoetheduke/27th-Solution-Lacuna-Solar-Challenge

The objective of this challenge is to develop a machine learning model capable of accurately detecting and counting solar panels and solar boilers in satellite and drone imagery of Madagascar. # **Lacuna Solar Survey Challenge – EfficientNetB5 Baseline** This project provides a full PyTorch-based training pipeline for the **Lacuna Solar Survey Challenge** on Kaggle. It predicts **boil number** and **pan number** from solar panel images using a **deep learning regression model** with **metadata fusion**. --- ## **Overview** The pipeline combines **image features** extracted by an EfficientNetV2 backbone with **metadata embeddings** (panel placement and image origin). It implements a **5-Fold Stratified Cross-Validation** setup for robust training and evaluation, and uses **Test-Time Augmentation (TTA)** for improved predictions. --- ## **Main Features** ✅ Full end-to-end deep learning pipeline (train → validate → predict) ✅ EfficientNetV2 backbone from `timm` with pretrained weights ✅ Metadata embeddings for `img_origin` and `placement` ✅ Albumentations-based augmentation for better generalization ✅ Automatic fold assignment using Stratified K-Fold ✅ Mixed-precision training with gradient scaling (`torch.cuda.amp`) ✅ Test-Time Augmentation (horizontal flip averaging) --- ## **Project Structure** ``` 📁 /kaggle/input/lacuna-solar-survey-challenge/ ├── Train.csv ├── images/ │ ├── 0001.jpg │ ├── 0002.jpg │ └── ... ``` - **Train.csv** — Contains metadata and target variables (`boil_nbr`, `pan_nbr`). - **images/** — Contains corresponding `.jpg` files named by `ID`. --- ## **Model Architecture** - **Backbone:** `tf_efficientnet_b5` (from `timm`) - **Metadata Embeddings:** - `img_origin` → Embedding(2, 8) - `placement` → Embedding(2, 8) - **Fusion Layer:** Concatenation of image + metadata features - **Head:** Two fully connected layers for regression output (2 targets) --- ## **Installation** Run in **Kaggle** or **Google Colab** environment with GPU support. ```bash !pip install timm albumentations ``` --- ## **Usage** 1. **Check GPU** ```python !nvidia-smi …