Python package for easy implementation of inverted encoding modeling as described in Scotti, Chen, & Golomb (in-prep).
Contact: [email protected] (Paul Scotti)
Run the following to install:
pip install inverted-encodingfrom inverted_encoding import IEM, permutation, circ_diff
import numpy as np
predictions, confidences, aligned_at_prediction_recons, aligned_at_zero_recons = IEM(trialbyvoxel,features,stim_max=180,nfolds=num_runs,is_circular=True)
# use "help(IEM)" for more information, below is a summary:
# trialbyvoxel: your matrix of brain activations, does not necessarily have to be voxels
# features: array of your stimulus features (must be integers within range defined by stim_max)
# stim_max=180 means that your stimulus space ranges 0-179° degrees
# nfolds refers to the K to use for KFold cross-validation. We recommend setting this to the number of runs you have for 1-run-left-out CV.
# is_circular=True for a circular stimulus space, False for non-circular stimulus space
# predictions: array of predicted stimulus for each trial
# confidences: array of goodness of fit values for each trial
# aligned_at_prediction_recons: trial-by-trial reconstructions (matrix of num_trials x stim_max) such that
# when plotted ideally each reconstruction is centered at the original trial stimulus
# aligned_at_zero_recons: trial-by-trial reconstructions aligned at zero, such that when
# plotted ideally each reconstruction is centered at zero on the x axis (e.g., plt.plot(aligned_at_zero_recons[trial,:]))
## Compute mean absolute error (MAE) by doing the following, then compare to null distribution:
if is_circular: # if your stimulus space is circular, need to compute circular differences
mae = np.mean(np.abs(circ_diff(predictions,features,stim_max)))
else:
mae = np.mean(np.abs(predictions-features))
null_mae_distribution = permutation(features,stim_max=180,num_perm=1000,is_circular=True)