Generalized linear models for neural spike train modeling, in Python! With GPU-accelerated fully-Bayesian inference, MAP inference, and network priors.
First, install Theano version 0.6.0 from here: http://deeplearning.net/software/theano/install.html#install
Make sure you have g++ installed, otherwise performance will be terrible! Theano will warn you if it cannot find g++.
Then, you can quickly demo the code by running:
-
cd pyglm
-
mkdir data
-
mkdir results
-
python -m test.generate_synth_data -N 4 -T 60 -m standard_glm -r data(outputs something like 'Creating results directory: data/2014_02_18-21_40')
-
python -m test.synth_map -d data/2014_02_18-21_40/data.pkl -r results(outputs something like 'Plots can be found in directory: results/2014_02_18-21_43')
-
open the PDFs in the results directory. Hopefully they look good!
GLMs are one of the most popular models for neural spike trains, yet there is no industry standard implementation. This package is an attempt to provide a simple, extensible framework for building GLMs, simulating spike trains, and fitting them using off-the-shelf optimization tools.
We build upon Theano, a package that offloads much of the work of deriving and verifying gradients and Hessians using automatic differentiation, and then compile the gradients into optimized C or CUDA code for fast execution. This makes model development quite easy - we work entirely with symbolic expressions. We pay a bit of a penalty by rederiving gradients, but in many cases this penalty is minor relative to the benefits of rapid development.
One of the primary motivations for using Python is the potential for easily parallelizing the code and offloading to Amazon EC2. This is still a work in progress, but should be coming soon!
The code is object oriented to mimic our modeling abstractions. We have a single GLM class which contains subclasses for the impulse responses, the stimulus response (via the background model), the bias, and the nonlinearity.
GLMs are bound together in a Population with shared variables. This allows us to model the underlying network of functional correlations and other variables that may be shared among individual neurons.
A more thorough code overview is in the works, but for now feel free to browse!
Some basic instructions for running with IPython.parallel (note, these instructions are specifically for my cluster with SGE queueing system. For your own cluster, consult the IPython.parallel tutorial on starting the controller and engines).
To start the controller (on a node visible from all engines):
ipcontroller --profile=sge --ip='*'
To start the engines (on cluster worker nodes):
ipcluster engines --profile=sge --n=8
Then, run the parallel test script:
python -m test.parallel_coord_descent -m standard_glm -d data/synth.mat -p sge
Ideally, this project will offer a substrate for sharing our expertise in GLMs. Feel free to contact me if you'd like to contribute!