🚧 Under Development 🚧
This repository contains code for training and validating segmentation models to perform building delineation on different types of satellite imagery: Low-Resolution (LR), High-Resolution (HR), and Super-Resolution (SR). The goal is to compare the performance of segmentation models across these varying resolutions.
The project leverages PyTorch Lightning for model training and Weights & Biases (W&B) for experiment tracking. It includes scripts to train segmentation models and validate them by calculating relevant metrics.
- train.py: Script to train the segmentation models using configurations specified in YAML files.
- validate.py: Script to validate the trained models and calculate segmentation metrics.
- configs/: Directory containing YAML configuration files for different training setups.
- model_files/: Contains model definitions and utilities.
The following segmentation models are implemented and can be selected through the configuration files in the configs/ directory:
| Model Name | Number of Bands | Pretrained Status |
|---|---|---|
| UNet | 4 | Scratch |
| UNet++ | 4 | Scratch |
| DeepLabV3 | 4 | Scratch |
| DeepLabV3+ | 4 | Scratch |
| TorchGeo ResNet18 | 3 | Backbone pretrained on S2 RGB |
| TorchGeo FarSeg | 3 | Backbone pretrained on S2 RGB |
These models are customizable via YAML configurations and are compatible with LR, HR, and SR imagery workflows. Important settings when changing models:
- Set number of bands in both model and data section
- Set appropriate loss, define wether sigmoid needs to be applied
To train a segmentation model:
- Update Configuration: Modify the configuration files in the configs/ directory to set your training parameters. Things to consider:
- Model Selection: Currently implemented are DeepLabV3, UNet and UNet++
- Training parameters: Optimizers, Schedulers, LRs etc
- Set the LR-SR-HR paramter
- if using dataloaders from this project, make sure to change the data information like path and interpolation setttings
- Run Training: Run train.py to start training, adjust which config to use. Either pass config file as argument from CL, or hardcode.
python train.py configs/config_hr.yaml- Validate: Run validate.py
- Give models and loaded weights + dataloaders to opensr-usecases package to get validation metrics, hardcoded in script.
- output in package format, as both text and mAP curve image.
- optionally, send output to txt file
python validate.py >> logs/output.txtCheck the training status at WandB
The validation is based on an external package. It outputs a numerical caluclation of the improvement of SR basic imagery over LR, as well as mAP curves for all data types
