!!! Note: GIN does not support bulk downloads. You need to download the files individually !!!
Sensorium (Mouse 1):
static26872-17-20-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
Sensorium+ (Mouse 2):
static27204-5-13-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
Pre-training scans (Mice 3-7):
static21067-10-18-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
static22846-10-16-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
static23343-5-17-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
static23656-14-22-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
static23964-4-22-GrayImageNet-94c6ff995dac583098847cfecd43e7b6
Below we provide a brief explanation of the dataset structure and how to access all the information contained in them.
Have a look at our white paper for in depth description of the data. White paper on arXiv
We provide the datasets in the .zip format. Unzipping them will create two folders data and meta.
X.npy
contains the image that was shown to the mouse in trial X
.X.npy
contains the deconvolved calcium traces (i.e. responses) recorded from the mouse in trial X
in response to the particular presented image.1 x 3
) where each single X.npy
contains the behavioral variables (in the same order that was mentioned earlier) for trial X
.1 x 2
) for horizontal and vertical eye positions.meta: includes meta data of the experiment
area.npy
: contains the area of each neuroncell_motor_coordinates.npy
: contains the position (x, y, z) of each neuron in the cortex, given in microns. Note: Thelayer.npy
: contains the cortex layer to which neuron belongs tounit_ids.npy
: contains a unique id for each neuronstatistics: This directory contains statistics (i.e. mean, median, etc.) of the experimental variables (i.e. behavior, images, pupil_center, and responses).
However, for the evaluation of submissions in the competition, we require the responses to be standardized (i.e. r = r/(std_r)
).
trials: This directory contains trial-specific meta data. They contain single 1-d NumPy arrays for each trial variable.
How to relate these meta data to the neuronal data (images, responses, ...)?
The indices of these arrays correspond to the .npy
files in data. For example:
# get meta data array
image_ids = np.load('./meta/trials/frame_image_id.npy')
# relate meta data with neuronal data
trial_image_id = image_ids[0]
corresponding_image = np.load('./data/images/0.npy')
corresponding_neuronal_response = np.load('./data/responses/0.npy')
Below are a list of important variables in this directory.
frame_image_id.npy
: contains unique image id. If the image is presented multiple times (which is the case in the test set) this image ID will be present multiple times.tiers.npy
: contains labels that are used to split the data into train, validation, and test set
trial_idx.npy
: contains a unique index for each trial. While the true trial index is available for the “pre-training” datasets, it is hidden (i.e. hashed) in the competition datasets.
The datasets 26872-17-20
(Sensorium) 27204-5-13
(Sensorium+) are different from the 5 other full datasets in these ways:
For Sensorium, the behavioral variables and eye position are withheld (arrays are present, but zeroed out)
This data is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. This license requires that you contact us before you use the data in your own research. In particular, this means that you have to ask for permission if you intend to publish a new analysis performed with this data (no derivative works-clause).