Note: The latest PyTorch version (0.4) merged Tensors and Variables. The current convis release was adapted to this change, but it is still possible that something broke in the change, so please report any Exceptions thrown by PyTorch in the issue tracker.

The convis package

This python package provides an implementation of the Virtual Retina developed by Adrien Wohrer. It uses PyTorch to simulate spike trains of retinal ganglion cells by directing the input through a number of computation layers. Each layer might do linear or nonlinear computations, eg. convolve the inpute with a spatio-temporal kernel or apply gain control.

TravisCI on the master branch:

We are supporting Python 3.5 and 2.7 right now, but are aiming to support Python 3.6 as well.

Convis is under development and some features might not work in the current master branch or the PyPi releases. If you discover unexpected behaviour, please leave an Issue on github.

Also there are two mailing lists for people interested in Convis:

• To recieve announcements of changes, please subsribe to: [email protected]
• If you want to participate in the development, please subscribe to: [email protected]

Find the documentation at: https://jahuth.github.io/convis/

Usage Example:

import numpy as np
import matplotlib.pylab as plt
import convis

c = convis.retina.RetinaConfiguration()
ret = convis.retina.Retina(c)
inp = np.zeros((2000,50,50))
inp[:,20:30,20:30] = 255.0*(rand(*inp[:,20:30,20:30].shape)<0.2)
out = ret.run(inp)

plt.plot(np.mean(out[0],(0,1,3,4)))
plt.plot(np.mean(out[1],(0,1,3,4)))

An earlier version using theano has been put on hold, but is still available here. If you are interested in continued development of the theano version, please let me know! An even older version was published as the retina package

When using convis for your scientific publications, please cite:

• Huth J, Masquelier T and Arleo A (2018) Convis: A Toolbox to Fit and Simulate Filter-Based Models of Early Visual Processing. Front. Neuroinform. 12:9. doi: 10.3389/fninf.2018.00009

Installation

Installing convis and PyTorch itself is not complicated.

Requirements for the base installation are: Python 2.7 or Python 3.5, Numpy, SciPy.

The current PyTorch version built for CUDA 8 can be installed via conda or pip:

conda install pytorch -c pytorch 
# or
pip install torch torchvision

For other CUDA versions, installing via pip on Windows etc., see the PyTorch website.

Once the requirements are installed, convis can be installed via pip for the latest release:

pip install convis

or from the most up-to-date version from github:

pip install git+https://github.com/jahuth/convis.git

Alternatively, you can clone the repository and install it locally:

git clone https://github.com/jahuth/convis.git
cd convis
# change something in the source code
pip install -e convis

I recommend installing opencv, and jupyter notebook, if you do not already have it installed:

pip install convis notebook
# eg. for ubuntu:
sudo apt-get install python-opencv

Introduction

convis provides a retina model which is identical to VirtualRetina and tools to create either simpler or more complex models.

The paper about convis is available in frontiers in Neuroinformatics: "Convis: A Toolbox To Fit and Simulate Filter-based Models of Early Visual Processing".

The Retina model

A description of all parameters for the retina model can be obtained directly from an instantiated model. This description contains doc strings for each parameter.

import convis
retina = convis.retina.Retina()
print(retina)

Here is a graph of the model:

To use the model, supply a numpy array as an argument to the Retina (for short input) or to the run function with a dt keyword to split the input in smaller chunks (and automatically reassemble the output):

import numpy as np
import matplotlib.pylab as plt
inp = np.ones((100,20,20))
output = retina(inp)
    
inp = np.ones((2000,20,20))
output = retina.run(inp,dt=100)

It will return an Output object containing all outputs of the model (the default for retina is two outputs: spikes of On and Off cells).

convis.plot_5d_time(output[0])
title('On Cells (1 line = 1 pixel)')
figure()
convis.plot_5d_matshow(output[0][:,:,::50,:,:])
title('Every 50th frame of activity')
figure()
# dimension 2 is time, so we mean over all others
# to get the average activity
convis.plot_5d_time(output[0].mean((0,1,3,4)))
convis.plot_5d_time(output[1].mean((0,1,3,4)))
title('Mean Activitiy of On and Off Cells')

The output object holds the ouput of the computation (in most cases torch.autograd.Variables) but can be converted to a numpy array with output.array(...). Outputs can be addressed with numbers or names.

>>> type(output[0])
<class 'torch.autograd.variable.Variable'>
>>> type(output[0].data.cpu().numpy())
<class 'numpy.ndarray'>
>>> type(output.array(0))
<class 'numpy.ndarray'>
>>> output[0] is output.ganglion_spikes_ON
True
>>> output[0] is output['ganglion_spikes_ON']
True
>>> np.sum(abs(output.array('ganglion_spikes_ON') - output.array(0)))
0

If instead of spikes, only the firing rate should be returned, the retina can be initialized without a spiking mechanism:

retina_without_spikes = convis.retina.Retina(spikes = False)

Found a bug or want to contribute?

Bug reports and feature requests are always welcome! The best place to put them is the github issue tracker. If you have questions about usage of functions and classes and can not find an answer in the documentation and docstrings, this is considered a bug and I appreciate it if you open an issue for that!

If you want, you can flag your issue already with one of the labels:

• bug
• enhancement
• missing documentation
• question

If you have fixed a bug or added a feature and you would like to see the change included in the main repository, the preferred method is for you to commit the change to a fork on your own github account and submit a pull request.

General discussion is encouraged on the two mailing lists:

• [email protected] for announcements and user questions
• [email protected] for discussions about the development