map_generation

Project init

1. Install PDM (only if not already installed)

   pip install pdm

2. Install package locally (will download all dev packages and create a local venv)

   # Optional if you want to create venv in a specific version. More info: https://pdm.fming.dev/latest/usage/venv/#create-a-virtualenv-yourself
   pdm venv create 3.10 # tested version
   
   pdm install

3. There is some problem with kaleido, you might have to install it with pip if you need it.

Notebooks

To see workflow process use notebooks from workflow_raster. It also contains less memory consuming training procedure.

Run code in container

• specify configs in config.py (dirs inside container, base model, batch size, epoch, n_columns used to generate prompt, dataset)
• dataset can be located in data/tiles (specify in config data/tiles) or huggingface dataset (like mprzymus/osm_tiles_small which contains tiles from Wrocław)
• tiles can be downloaded from https://drive.google.com/drive/u/0/folders/1rIsECcHQNBj60905VhGzPUkeQuaBKgmI
• you can specify more arguments in run.sh according to https://huggingface.co/docs/diffusers/training/text2image
• docker build --tag diffusion .
• docker run -it --gpus all -v ./logs_cont:/app/result_path diffusion
• to monitor logs run tensorboard --logdir logs_cont/logs
• after training model is located inside log_cont directory

Demo app

Demo app is available! To run it with existing model use:

pdm sync -G prod 
streamlit run map_generation/app/app.py 

You can also use existing docker image:

docker pull mprzymus/diffusion_demo
docker run --gpus 1 -p 6661:5000 mprzymus/diffusion_demo