def internal_cons(self, splithalfno, test_inds):
rng = np.random.RandomState(splithalfno)
split1, split2 = utils.splithalf(self.neural_feats_reps[:, test_inds],
rng=rng)
r = utils.pearsonr_matrix(split1, split2)
rc = utils.spearman_brown_correct(r, n=2)
return rc
def raw_fit(self, train_inds, test_inds):
neural_feats = np.nanmean(self.neural_feats_reps, axis=0)
self.reg.fit(self.model_feats[train_inds],
np.squeeze(neural_feats[train_inds]))
pred = np.squeeze(self.reg.predict(self.model_feats[test_inds]))
actual = neural_feats[test_inds]
rs = utils.pearsonr_matrix(actual, pred)
return np.squeeze(rs)
def splithalf_corr(data, niter=10, seed=None):
rng = np.random.RandomState(seed)
df = []
for i in range(niter):
split1, split2 = utils.splithalf(data, rng=rng)
r = utils.pearsonr_matrix(split1, split2)
rc = utils.spearman_brown_correct(r, n=2)
df.extend([(i, site, rci) for site, rci in enumerate(rc)])
df = pandas.DataFrame(df, columns=['splithalf', 'site', 'internal_cons'])
return df
def mapping_cons(self, splithalfno, train_inds, test_inds):
"""
Split data in half over reps, run PLS on each half on the train set,
get predictions for the test set, correlate the two, Spearman-Brown
"""
rng = np.random.RandomState(splithalfno)
split1, split2 = utils.splithalf(self.neural_feats_reps[:, train_inds],
rng=rng)
self.reg.fit(self.model_feats[train_inds], split1)
pred1 = self.reg.predict(self.model_feats[test_inds])
self.reg.fit(self.model_feats[train_inds], split2)
pred2 = self.reg.predict(self.model_feats[test_inds])
rs = utils.pearsonr_matrix(pred1, pred2)
return np.squeeze(utils.spearman_brown_correct(rs, n=2))
def nfit(model_feats,
neural,
labels,
n_splits=10,
n_components=200,
test_size=.25):
if model_feats.shape[1] > n_components:
model_feats = PCA(n_components=n_components).fit_transform(model_feats)
skf = StratifiedShuffleSplit(n_splits=n_splits, test_size=test_size)
df = []
for it, (train_idx, test_idx) in enumerate(skf.split(model_feats, labels)):
reg = PLSRegression(n_components=25, scale=False)
reg.fit(model_feats[train_idx], neural[train_idx])
pred = reg.predict(model_feats[test_idx])
rs = utils.pearsonr_matrix(neural[test_idx], pred)
df.extend([(it, site, r) for site, r in enumerate(rs)])
df = pandas.DataFrame(df, columns=['split', 'site', 'fit_r'])
return df