segformer_playground

Messin around with

• https://huggingface.co/docs/transformers/model_doc/segformer and
• https://huggingface.co/blog/fine-tune-segformer

This repo trains a model that predicts land cover using aerial imagery downloaded from https://earthexplorer.usgs.gov/. This is a semantic segmentation task: see prediction_rasters for example prediction rasters, which you can visualize using e.g. QGIS.

./save_tiles.sh

sudo docker build ~/segformer_playground --tag=segformer_playground
sudo docker run --gpus all -it -v ~/segformer_playground:/home/segformer_playground segformer_playground bash
python segformer/train_model.py

sudo docker run -v ~/segformer_playground:/home/segformer_playground segformer_playground pytest