This repo contains materials from CosmiQ Works' open source and/or public workshops on using solaris for geospatial ML applications. Each sub-directory corresponds to a specific workshop or tutorial. Some of the workshop materials will require additional data download and/or preparation beyond the contents of the repository; see the README files within each workshop for additional instructions.
Note that many of the workshops may require access to (at times substantial) GPU resources for model training. We've tried to indicate that both in the outline below as well as within the workshop materials themselves.
Familiarity with Python, Jupyter Notebooks, and basic concepts around geospatial data (what satellite imagery is, what vector-formatted labels are). Solaris must already be installed on your computer. One of the notebooks requires GPU compute (to complete in any reasonable amount of time).
- An intro to neural nets and machine learning for geospatial applications (slides)
- Running inference with a pre-trained model from a SpaceNet Challenge (Jupyter Notebooks)
- Evaluating model performance (slides and Jupyter Notebook)
- Fine-tuning a pre-trained model on new imagery (Jupyter Notebook) requires GPU resources
- Multi-channel training targets and combination loss functions (slides and Jupyter Notebook)