Code for Training deep neural density estimators to identify mechanistic models of neural dynamics. Each application has its own subfolder.

The experiments in the paper were run using SNPE based on Theano, using the delfi toolbox as installed below. For new applications of SNPE, we recommend using the sbi toolbox, which is based on PyTorch and has extended functionality.

Setup a base environment for running these experiments as follows:

conda create -n ind python=3.7
conda activate ind
conda install numpy scipy matplotlib ipython jupyter jupyterlab pandas seaborn

pip install dill python-box svgutils cython
pip install -r https://raw.githubusercontent.com/Lasagne/Lasagne/master/requirements.txt
pip install https://github.com/Lasagne/Lasagne/archive/master.zip
pip install theano --upgrade
pip install parameters cma tqdm dill==0.2.7.1 python-box==3.1.1

git clone https://github.com/mackelab/delfi.git
cd delfi
pip install -r requirements.txt
pip install .

ipython kernel install --name "ind" --user

Use the ind kernel when running notebooks. If applications have extra dependencies or require compilation of Cython models, this is stated in the README.md of the application subfolder.

Make sure to use float32 precision with theano, e.g. by adding this to the theano config file (~/.theanorc):

[global]
floatX = float32

@article{gonccalves2020training,
  title     = {Training deep neural density estimators to identify mechanistic models of neural dynamics},
  author    = {Gon{\c{c}}alves, Pedro J and Lueckmann, Jan-Matthis and Deistler, Michael and Nonnenmacher, Marcel and {\"O}cal, Kaan and Bassetto, Giacomo and Chintaluri, Chaitanya and Podlaski, William F and Haddad, Sara A and Vogels, Tim P and Greenberg, David S. and Macke, Jakob H.},
  year      = {2020},
  doi       = {10.7554/eLife.56261},
  publisher = {eLife},
  journal   = {eLife}
}

MIT