This is a Python package for spike inference from calcium imaging data using deep neural networks that are trained unsupervised with variational autoencoders.

The details of the algorithm are presented in the paper Fast amortized inference of neural activity from calcium imaging data with variational autoencoders (Speiser, Yan, Archer, Buesing, Turaga and Macke)

The code requires Theano and Lasagne.

The repository includes four notebooks that show how the algorithm is used:

• EX1: Training a CNN on data simulated from a simple linear model of fluorescence dynamics.
• EX2: Training a RNN on data simulated from a nonlinear model of fluorescence dynamics.
• EX3: Training a RNN on publically available real data.
• GC6s_prep: Preprocessing of the real data.

While the CNN can be trained on a CPU in reasonable time, training the RNN requires a GPU.