README.md

---
title: SamGIS - LISA on CUDA
emoji: 🗺️
colorFrom: red
colorTo: blue
sdk: docker
app_file: app.py
pinned: true
license: mit
---

# [LISA](https://github.com/dvlab-research/LISA) + [SamGIS](https://github.com/trincadev/samgis-be) on a dedicated CUDA GPU

This project aims to permit use of [LISA](https://github.com/dvlab-research/LISA) (Reasoning Segmentation via Large Language Model) applied to geospatial data thanks to [SamGIS](https://github.com/trincadev/samgis-be). In this space I adapted LISA to HuggingFace [lisa-on-cuda](https://huggingface.co/spaces/aletrn/lisa-on-cuda) ZeroGPU space.

This [home page project](https://huggingface.co/spaces/aletrn/samgis-lisa-on-zero) is a plane Gradio interface that take a json in input to translate it to a geojson. More information about these API implementation [here](
https://aletrn-samgis-lisa-on-zero.hf.space/docs). On this [blog page](https://trinca.tornidor.com/projects/lisa-adapted-for-samgis) you can find more details, including some request and response examples with the geojson map representations.

You can also find the alternative map interface [here](https://aletrn-samgis-lisa-on-zero.hf.space/lisa/) useful to create on the fly the payload requests and to represent the geojson response.

## Custom environment variables for HuggingFace CUDA Space

Fundamental environment variables you need are:

```bash
XDG_CACHE_HOME="/data/.cache"
PROJECT_ROOT_FOLDER="/home/user/app"
WORKDIR="/home/user/app"
```

Derived ones:

```bash
MPLCONFIGDIR="/data/.cache/matplotlib"
TRANSFORMERS_CACHE="/data/.cache/transformers"
PYTORCH_KERNEL_CACHE_PATH="/data/.cache/torch/kernels"
FASTAPI_STATIC="/home/user/app/static"
VIS_OUTPUT="/home/user/app/vis_output"
MODEL_FOLDER="/home/user/app/machine_learning_models"
FOLDERS_MAP='{"WORKDIR":"/home/user/app","XDG_CACHE_HOME":"/data/.cache","PROJECT_ROOT_FOLDER":"/home/user/app","MPLCONFIGDIR":"/data/.cache/matplotlib","TRANSFORMERS_CACHE":"/data/.cache/transformers","PYTORCH_KERNEL_CACHE_PATH":"/data/.cache/torch/kernels","FASTAPI_STATIC":"/home/user/app/static","VIS_OUTPUT":"/home/user/app/vis_output"}'
```

The function `build_frontend()` from lisa_on_cuda package create all the folders required for this project using the environment variable `FOLDERS_MAP`. That's useful for cache folders (XDG_CACHE_HOME, MPLCONFIGDIR, TRANSFORMERS_CACHE, PYTORCH_KERNEL_CACHE_PATH) because missing these can slow down the inference process. Also you could keep these folders in a permanent storage disk mounted on a custom path.

To change the base relative url for custom frontend add the VITE_PREFIX environment variable, e.g.:

```bash
VITE_INDEX_URL="/custom-url"
```

## Local environment

To execute the you can simply run from the root project folder:

- `python app.py`
- `python -m uvicorn app:app --host 0.0.0.0 --port 7860`

It's also possible to make and run a docker build:

- `docker build . -t registry.gitlab.com/aletrn/samgis-lisa-on-cuda --progress=plain`
- `docker run -d --name samgis-lisa -p 7860:7860 --gpus all --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 registry.gitlab.com/aletrn/samgis-lisa; docker logs -f samgis-lisa-on-cuda`

## About HuggingFace space dependencies

For this demo simply installing `samgis-lisa` already brings all the needed dependencies. 
Now `lisa.lisa_predict()` has the optional argument `inference_decorator` useful in case of use on ZeroGPU hardware or similar.