Code accompanying Induction-Gram experiments on an fMRI dataset found at openneuro.
To clone and use this dataset:
$ git clone https://github.com/SriyaM/deep_fMRI.git
then to intialize:
$ cd deep-fMRI
$ pip install .
First datalad needs to be installed which can be done with:
sudo apt get install datalad
Then, to automatically download the preprocessed data
$ cd encoding
$ python load_dataset.py -download_preprocess
This function will create a data dir if it does not exist and will use datalad to download the preprocessed data as well as feature spaces needed for fitting semantic encoding models. It will download ~20gb of data. Alternately, you can supply a different download location using the --location DATA_DIR flag. If you choose to change the default location of the data, make sure to update the config.py file with the new location.
To download the raw data you can use:
$ datalad clone https://github.com/OpenNeuroDatasets/ds003020.git
$ datalad get ds003020
The basic functionality for fitting encoding models can be found the script encoding.py, which takes a series of arguments such as subject id, feature space to use, list of training stimuli, etc.
It will automatically use the preprocessed data from the location that get_data saves the data to.
To run any of the Induction-Gram experiments on a given subject you must first run the encoding/encoding_save_presp.py script to save the top 100 principal components of the responses of a specific subject from the loaded dataset. For example, to save the PCA components for a subject (UTS03) you can run:
$ python encoding/encoding_save_presp.py --subject UTS03
Then, to fit any encoding model (incontext_infinigram) for one subject (UTS03) and test it on held-out data by using the top 900 principle components of the response vectors and the top 80% of matches for induction, you can run:
$ python encoding/encoding.py --subject UTS03 --feature incontext_infinigram --x 900 --k 0.8
For the eng1000 feature, the x and k parameters are needed. For the llama feature, k specifies the context window size. For all infinigram-based induction features, x represents the number of principal components to compress the infinigram vectors, and k represents the top k proportion of words to consider for induction matches. Refer to feature_spaces.py for more details.
The other optional parameters that encoding.py takes such as sessions, ndelays, single_alpha allow the user to change the amount of data and regularization aspects of the linear regression used.
This function will then save model performance metrics and model weights as numpy arrays.
For more information about fitting voxelwise encoding models: