This repository provides the code necessary to recapitulate the analyses and figures presented in the article entitled "Sensorimotor strategies and neuronal representations for shape discrimination". https://doi.org/10.1016/j.neuron.2021.05.019
A somewhat out-of-date preprint version of the article is available at bioRxiv https://www.biorxiv.org/content/10.1101/2020.06.16.126631v1.
Update September 2022: A new version of this dataset is available at https://dandiarchive.org/dandiset/000231, with a related preprint at https://www.biorxiv.org/content/10.1101/2022.05.10.491259v1 containing extensive documentation. The advantage of the new dataset is that it includes more raw data (the entire electrophysiological recording, all the videos), and it is in the Neurodata Without Borders (NWB) format. However, this Rodgers2021 dataset remains useful for replicating the analyses presented in the original Neuron paper.
The data will be available at the following DOI: 10.5281/zenodo.4743837. https://zenodo.org/record/4743837#.YgQ7xYzMIuU
The dataset on Zenodo consists of an archive file called ShapeDataset.zip.
Download that file and unzip it. Make note of the location. The rest of
this document assumes that the unzipped data resides at ~/data/ShapeDataset/.
For a full description of the contents of this data archive, see the file
DATA_README.md in this repository.
While the data is downloading, you can start setting up an environment for analyzing it. Linux is the only OS on which this code has been tested.
If you have any problems, please create a Github Issue on this repository. Please note that I cannot promise to provide support, though I can try to answer questions as they come up. In principle, this code should run in a properly configured conda environment on any operating system; but in practice, I don't know anything about Windows or OSX, so if you have problems on those operating systems, it is very unlikely that I will be able to help.
Most of the data is stored as pickled pandas.DataFrame files. It is
very unlikely you will be able to open these files unless you use exactly
the same version of pandas as I did to create them. The following steps
show you how to set up an environment with the correct version of all
requirements.
First, clone this repository into your home directory. (You can easily modify these commands if you wish to clone it somewhere else.)
cd ~
git clone --recursive https://github.com/cxrodgers/Rodgers2021
The flag "--recursive" here trigger git to also fetch the required
submodules, which are just some of my other github repositories that this
repository depends on. Those will go into the subdirectory called submodules.
Set up a conda environment called Rodgers2021. This will install all
the necessary packages (listed in requirements.txt and
pip_requirements.txt) into that environment.
cd ~/Rodgers2021
conda create --name Rodgers2021 --file requirements.txt
conda activate Rodgers2021
pip install -r pip_requirements.txt
You must also set some environment variables. First, set $MPLCONFIGDIR
so that the "agg" background is always used. (Otherwise, these files hang
until the plot is closed.)
conda deactivate
conda activate Rodgers2021
conda env config vars set MPLCONFIGDIR=$HOME/Rodgers2021/mpl_config_dir
Second, set the $PYTHONPATH to include the path to the submodules.
conda deactivate
conda activate Rodgers2021
conda env config vars set PYTHONPATH=$HOME/Rodgers2021/submodules
Reactivate the environment once again to trigger these changes.
conda deactivate
conda activate Rodgers2021
All of the scripts use a JSON file called ~/Rodgers2021/parameters to find the
data. Open up that file with a text editor.
Near the bottom of the file will be a bunch of lines that contain the string "ShapeDataset". These paths must be fixed to go to the location of the downloaded data.
"pipeline_input_dir": "/home/jack/data/ShapeDataset",
"root_session_dir": "/home/jack/data/ShapeDataset/sessions",
"unit_db_dir": "/home/jack/data/ShapeDataset/neural_metadata",
"trainset_dir": "/home/jack/data/ShapeDataset/trainset",
"glm_fits_dir": "/home/jack/data/ShapeDataset/glm_fits",
"glm_example_dir": "/home/jack/data/ShapeDataset/glm_example",
"example_frames_dir": "/home/jack/data/ShapeDataset/example_frames"
In this case I have them pointing to /home/jack/data/ShapeDataset. Change
that to the actual location of the unzipped data on your computer.
Running this code creates many large intermediate data files. Choose a
location for those intermediate files. Let's say it's
~/data/ShapeDataset_output. These lines at the top of the file must
be set to that path.
"pipeline_output_dir": "/home/jack/data/ShapeDataset_output",
"patterns_dir": "/home/jack/data/ShapeDataset_output/patterns",
"logreg_dir": "/home/jack/data/ShapeDataset_output/logreg",
"neural_dir": "/home/jack/data/ShapeDataset_output/neural",
"glm_dir": "/home/jack/data/ShapeDataset_output/glm",
"scaling_dir": "/home/jack/data/ShapeDataset_output/scaling",
Unless your username is also "jack", the default paths will not be correct on your computer, so you must make these changes.
Once you have completed the steps above, you can run the analyses. All of the analyses exist in numbered directories:
01_patterns/
03_logreg/
04_logreg_vis/
05_behavior_vis/
06_neural/
07a_glm/
08_neural_vis/
09_neural_decoding/
Within each of these numbered directory is a set of Python scripts.
01_patterns/main1b.py
01_patterns/main1c.py
01_patterns/main2a.py
...
Each script must be run in alphabetical order, because the later ones might depend on the earlier ones.
If you're feeling lucky, you can use the script runall.py to run all
of the scripts in sequence. If anything goes wrong (most likely a missing
import or an incorrect path to the data), it will crash in place.
conda activate Rodgers2021
cd ~/Rodgers2021
python runall.py
The results will be stored as figures in the numbered subdirectories.
This CSV file lists the subdirectory and script name that generates each panel in the manuscript, as well as the image filename of the figure.