Setting up Python

The examples are tested with Python 3.12 and the package versions in requirements.txt. We use conda for the environment and pip for the packages. From the repository root, run

conda create -n estimating-hank-nn python=3.12
conda activate estimating-hank-nn
pip install -r requirements.txt
pip install -e .

The last line installs the shared estimating_hank_nn package in editable mode, so the examples can import it and changes to the source take effect without reinstalling. To check the installation and see which accelerators PyTorch can use, run

python -c "import torch, estimating_hank_nn; print(torch.__version__, torch.cuda.is_available(), torch.backends.mps.is_available())"

Then open examples/rank/rank_baseline.py in VS Code, select the estimating-hank-nn environment as the kernel, and run the cells in order.

The New Keynesian examples run on a CPU, except Extension 10, which defaults to CUDA. An NVIDIA GPU ("cuda") or an Apple GPU ("mps") speeds up training, and Extension 10 shows how to select one. Training the heterogeneous agent models needs an NVIDIA GPU, but their saved networks let you reproduce the results on a laptop.