First of all, make sure the project structure matches the one below:

├── fswp_train.yml
├── fswp_run.yml
├── _data
│   └── # pickle files for model training
├── checkpoints
│   └── # .pth models 
├── models
│   ├── unet #unet trainer
│   └── convlstm #convlstm trainer
├── utils
└── docker 

Download convlstm training data, and data unet training. Put the received datasets in the data folder.

Download the unet model and put it in a folder checkpoints/unet

Download the convlstm model and put it in a folder checkpoints/convlstm

pip install -r requirements.txt

We use crafting to automate our experiments. You can find an example of running such a pipeline in run.yaml file. You need to have installed Docker, Nvidia drivers, and crafting package.

The crafting package is available in PyPI:

pip install crafting

To build the image run the command below in docker folder:

sh build.sh

To run an experiment specify target command in command field in run.yaml file and call crafting:

crafting configs/docker_run.yaml

train.sh -fswp_train.yml

python3 utils/predict_on_tdataset.py ./configs/run_solar_fswp.yaml

python3 utils/run_on_env.py configs/run_star_fswp.yaml False checkpoints/convlstm/CONV_LSTM_run_EricWright_94ed9438-8fae-4fc8-8afa-85409d0c6f46.pth