rectipy 0.4.0

Recurrent neural network training in Python

RectiPy

Recurrent neural network training in Python (RectiPy) is a software package developed by Richard Gast that allows for lightweight implementations of recurrent neural networks (RNNs) based on ordinary or delayed differential equations. RectiPy provides an intuitive YAML interface for model definition, and leverages PyRates to translate these model definitions into PyTorch functions. This way, users can easily define their own neuron models, spike-based or rate-based, and use them to create a RNN model. All model training, testing, as well as numerical integration of the differential equations is also performed in PyTorch. Thus, RectiPy comes with all the gradient-based optimization and parallelization features that PyTorch provides.

Basic Features

1. Model definition

• RNN layers are defined via ordinary or delayed differential equations that govern the neuron dynamics
• neurons can either be rate neurons or spiking neurons
• RNN layers can either be defined via YAML templates (see documentation of PyRates for a detailed documentation of the YAML-based model definition) or via custom PyTorch modules.
• linear input and output layers can be added, thus connecting the RNN into a layered neural network

2. Model training and testing

• input and output weights, as well as any parameters of the RNN layers can be trained
• autograd functions by PyTorch are used for the parameter optimization
• most loss functions and optimization algorithms implemented in PyTorch are available

3. Model outputs

• record any RNN state variable, loss, or model outputs via the Observer class
• choose at which rate to sample your recordings
• visualize for recordings via lightweight plotting functions
• connect the RectiPy network to larger deep learning architectures

Installation

Stable release (PyPi)

You can install the most recent stable version of RectiPy via the pip command. To this end, execute the following command via the terminal within the Python environment you would like to install RectiPy in:

pip install rectipy

This will also install the dependencies of the software listed below.

Development version (github)

To install the most recent development version of RectiPy as available on the master branch, clone this repository and run the following line from the directory in which the repository was cloned:

python setup.py install

Again, this will also install the dependencies of the software listed below.

Dependencies

• torch
• pyrates
• numpy
• matplotlib

Contact

If you have any questions, want to contribute to the software, or just get in touch, feel free to post an issue or contact Richard Gast.