deep_neuro

deep_neuro presents multiple ways of exploring the Gray dataset. At its core, it features a convolutional neural network trying to predict either stimulus class or response type from different brain regions and at different points in time (CNN module). It also transforms the raw data from MATLAB to NumPy-ready and applies the usual pre-processing steps (matnpy module) as well as dealing with the analysis and visualization of the results (monkeyplot module).

Installation

To get started, create a project directory using mkdir my_project/ and change into it (cd my_project/). Then clone this repository using git clone https://github.com/rpaul23/deep_neuro.git. Once cloned, change into the directory (cd deep_neuro/) and source the environment generator file:

. generate_environment.sh

This will set up your folder structure (see below) and install all required packages. Move raw data into data/raw/ (and/or already pre-processed data into data/pre-processed/) and you should be ready to go.

my_project
|___data
|   |___raw
|   |___pre-processed
|
|___scripts
|   |____params
|   |___deep_neuro (git repo)
|       |___lib
|           ...
|
|___results
    |___training
        |___pvals
        |___summary
        |___plots

Pre-process raw data

The pre-prossesing parameters are set in lib/matnpy/matnpy.py. To pre-process raw data, just cd into the scripts/deep_neuro/ directory and source the prep_data.sh file using

. prep_data.sh -u <user_name> -s <session> -t <trial_length>.

The trial_length parameter is optional and set to 500 ms by default. How pre-processing using the matnpy module works is that it cuts every trial of a given session into five intervals:

• pre-sample (500 ms before stimulus onset)
• sample (500 ms after stimulus onset)
• delay (500 ms after stimulus offset)
• pre-match (500 ms before match onset)
• match (500 ms after match onset)

Example call: . prep_data.sh -u jannesschaefer -s 141023 -t 500

Train classifier

To train a classifier, cd into your scripts/deep_neuro/ directory and source the submit file using

. submit_training.sh -a <user_name> -s <session> -n <node>

where user_name refers to your cluster account name (e. g. jannesschaefer) and session to the session number that you want to train the classifier on (e. g. 141023). If you want to use a specific node on the cluster, you can use the optional -n flag (e. g. -n n04). The job will be submitted to the cluster and processed once the resources are available.

Example call: . submit_training.sh -u jannesschaefer -s 140123 -n n04

Get results

To get results, just cd into the scripts/deep_neuro/ directory and source the get_results.sh file. This will generate a summary file in results/training/summary/ and a file of pvalues in results/training/pvals/. Plots will be stored in results/training/plots/.

Example call: . get_results.sh -u jannesschaefer -s 141023