2v2

2v2.py

import numpy as np
from sklearn.metrics import pairwise_distances
from numpy.random import permutation
from time import time


def cc_2v2(true, pred, metric="cosine"):
    assert len(true) == len(pred)
    ns = len(true)
    first = permutation(ns)  # first group of TR
    second = permutation(ns)  # second group of TR
    while (first == second).any():  # check that distinct TRs in pairs
        first[first == second] = np.random.choice((first == second).sum())

    correct = 0.
    for i, j in zip(first, second):
        # compute the 4 distances
        r = pairwise_distances(true[[i, j]], pred[[i, j]], metric)
        diag = np.diag(r).sum()  # distances of corresponding TR
        cross = r.sum() - diag  # distance of cross TR
        correct += 1. * (diag < cross)  # comparison
    acc = correct / ns
    return acc


def jr_2v2(true, pred, metric="cosine"):
    """Tsonova et al 2019 https://arxiv.org/pdf/2009.08424.pdf"""
    assert len(true) == len(pred)
    ns = len(true)
    first = permutation(ns)  # first group of TR
    second = permutation(ns)  # second group of TR
    while (first == second).any():  # check that distinct TRs in pairs
        first[first == second] = np.random.choice((first == second).sum())

    r = pairwise_distances(true, pred)
    s1 = r[first, first] + r[second, second]
    s2 = r[first, second] + r[second, first]
    
    acc = np.mean(1.*(s1<s2))
    return acc


n_samples, n_voxels = 1000, 20
true = np.random.randn(n_samples, n_voxels)
snr = .1
pred = snr*true + np.random.randn(n_samples, n_voxels)

start = time()
cc = cc_2v2(true, pred)
print('cc acc=', cc, 'time', time()-start)

start = time()
jr = jr_2v2(true, pred)
print('jr acc=', jr, 'time', time()-start)

assert jr==cc