def get_pattern(X, y, clf, time_point=None):
"""Get pattern from classifier on X and y at peak time.
Re-fit the classifier without cross-validation and get the patterns/
coefficients.
Parameters:
-----------
X : array
The data to fit (features).
y : vector
The response vector.
clf : sklearn classifier object
The classifier to re-fit.
time-point : float | None
If data is more than two dimensional: which time-point to fit.
Returns:
--------
pattern : array
The sensor or source pattern of coefficients.
"""
if time_point is not None:
X = X[:, :, time_point]
clf.fit(X, y)
if clf.steps[-1][0] == 'logisticregression':
pattern = get_coef(clf, 'coefs_', inverse_transform=True)
else:
pattern = get_coef(clf, 'patterns_', inverse_transform=True)
return pattern
def logreg_timedecoding(epochs, numcv=4, jobs=1):
"""
Logistic regression over sensors. Returns Evoked array containing coefficients and ROC.
Code snippets stolen from:
https://martinos.org/mne/stable/auto_tutorials/plot_sensors_decoding.html
"""
X = epochs.get_data() # MEG signals: n_epochs, n_channels, n_times
X = X.astype(float)
y = epochs.events[:, 2] # targets
# setup and run the decoder
clf = make_pipeline(StandardScaler(), LinearModel(LogisticRegression()))
time_decod = SlidingEstimator(clf, scoring='roc_auc',
n_jobs=jobs) #scoring='roc_auc',
scores = cross_val_multiscore(time_decod, X, y, cv=numcv, n_jobs=jobs)
# Mean scores across cross-validation splits
scores = np.mean(scores, axis=0)
#
time_decod = SlidingEstimator(clf, scoring='roc_auc', n_jobs=jobs)
time_decod.fit(X, y)
coef = get_coef(time_decod, 'patterns_', inverse_transform=True)
evoked = mne.EvokedArray(coef, epochs.info, tmin=epochs.times[0])
evoked.roc_auc = scores
return evoked
# 4.2) vectorise channel data for linear regression
# data to be analysed
dat = cues[subj].get_data()
dat = dat[:, :, times_to_use]
Y = Vectorizer().fit_transform(dat)
# 4.3) fit linear model with sklearn's LinearRegression
weights = compute_sample_weight(class_weight='balanced',
y=metadata.cue.to_numpy())
linear_model = LinearRegression(n_jobs=n_jobs, fit_intercept=True)
linear_model.fit(design, Y, sample_weight=weights)
# 4.4) extract the resulting coefficients (i.e., betas)
# extract betas
coefs = get_coef(linear_model, 'coef_')
inter = linear_model.intercept_
# 4.5) extract model r_squared
r2 = r2_score(Y, linear_model.predict(design), multioutput='raw_values')
# save model R-squared
r_squared[subj_ind, :] = r2
# save results
for pred_i, predictor in enumerate(design.columns):
print(pred_i, predictor)
if 'cue' in predictor:
# extract cue beats
betas[subj_ind, :] = coefs[:, pred_i]
# elif 'Intercept' in predictor:
# continue
ax.axhline(.5, color='k', linestyle='--', label='chance')
ax.set_xlabel('Times')
ax.set_ylabel('AUC') # Area Under the Curve
ax.legend()
ax.axvline(.0, color='k', linestyle='-')
ax.set_title('Sensor space decoding')
plt.show()
###############################################################################
# You can retrieve the spatial filters and spatial patterns if you explicitly
# use a LinearModel
clf = make_pipeline(StandardScaler(), LinearModel(LogisticRegression()))
time_decod = SlidingEstimator(clf, n_jobs=1, scoring='roc_auc', verbose=True)
time_decod.fit(X, y)
coef = get_coef(time_decod, 'patterns_', inverse_transform=True)
evoked = mne.EvokedArray(coef, epochs.info, tmin=epochs.times[0])
joint_kwargs = dict(ts_args=dict(time_unit='s'),
topomap_args=dict(time_unit='s'))
evoked.plot_joint(times=np.arange(0., .500, .100),
title='patterns',
**joint_kwargs)
###############################################################################
# Temporal Generalization
# -----------------------
#
# This runs the analysis used in [1]_ and further detailed in [2]_
#
# The idea is to fit the models on each time instant and see how it
# generalizes to any other time point.
tmin = i * 1.0
tmax = (i + 1) * 1.0
#Create :class:'Epochs <mne.Epochs>' object
epochs = mne.Epochs(raw,
events=events,
event_id=event_id,
tmin=tmin,
tmax=tmax,
baseline=None,
verbose=True,
preload=True)
for i in range(0, len(epochs.events)):
if i % 2 == 0:
epochs.events[i, 2] = 3
#epochs.plot(scalings = 'auto',block = True,n_epochs=10)
X = epochs.pick_types(meg=False, eeg=True)
y = epochs.events[:, -1]
# Define a unique pipeline to sequentially:
clf = make_pipeline(
Vectorizer(), # 1) vectorize across time and channels
StandardScaler(), # 2) normalize features across trials
LinearModel(LogisticRegression())) # 3) fits a logistic regression
clf.fit(X, y)
coef = get_coef(clf, 'patterns_', inverse_transform=True)
evoked = EvokedArray(coef, epochs.info, tmin=epochs.tmin)
fig = evoked.plot_topomap(title='EEG Patterns', size=3, show=False)
fig.savefig(title + "_ti_" + str(tmin) + "_tf_" + str(tmax) + '.png')
# The data was vectorized to fit a single model across all time points and
# all channels. We thus reshape it:
coef = coef.reshape(len(meg_epochs.ch_names), -1)
# Plot
evoked = EvokedArray(coef, meg_epochs.info, tmin=epochs.tmin)
evoked.plot_topomap(title='MEG %s' % name, time_unit='s')
###############################################################################
# Let's do the same on EEG data using a scikit-learn pipeline
X = epochs.pick_types(meg=False, eeg=True)
y = epochs.events[:, 2]
# Define a unique pipeline to sequentially:
clf = make_pipeline(
Vectorizer(), # 1) vectorize across time and channels
StandardScaler(), # 2) normalize features across trials
LinearModel(
LogisticRegression(solver='lbfgs'))) # 3) fits a logistic regression
clf.fit(X, y)
# Extract and plot patterns and filters
for name in ('patterns_', 'filters_'):
# The `inverse_transform` parameter will call this method on any estimator
# contained in the pipeline, in reverse order.
coef = get_coef(clf, name, inverse_transform=True)
evoked = EvokedArray(coef, epochs.info, tmin=epochs.tmin)
evoked.plot_topomap(title='EEG %s' % name[:-1], time_unit='s')
random = np.random.RandomState(random_state)
# number of random samples
boot = 2000
# create empty array for saving the bootstrap samples
boot_betas = np.zeros((boot, Y.shape[1], len(predictors)))
# run bootstrap for regression coefficients
for i in range(boot):
# extract random epochs from data
resamples = random.choice(range(n_epochs), n_epochs, replace=True)
# set up model and fit model
model = LinearRegression(fit_intercept=False)
model.fit(X=design.iloc[resamples], y=Y[resamples, :])
# extract regression coefficients
boot_betas[i, :, :] = get_coef(model, 'coef_')
# delete the previously fitted model
del model
###############################################################################
# compute lower and upper boundaries of confidence interval based on
# distribution of bootstrap betas.
lower, upper = np.quantile(boot_betas, [.025, .975], axis=0)
###############################################################################
# fit linear regression model to original data and store the results in
# MNE's evoked format for convenience
# set up linear model
linear_model = LinearRegression(fit_intercept=False)
# fit model
sleep_epochs = myload(sleep_path_data, typ='epoch_preprocessed', sbj=sbj, preload=True) # SLEEP!
sleep_epochs.event_id = sleep_event_id # event_id remapping. For wake this step works during preprocessing # SLEEP !
sleep_epochs = sleep_epochs.crop(tmin=tmin, tmax=tmax)
X1, y1 = get_Xy_balanced(sleep_epochs, contrast1)
clf = make_pipeline(Vectorizer(), StandardScaler(), LinearModel(LogisticRegression(max_iter = 4000))) #StandardScaler(),
cv = StratifiedKFold(n_splits=2, shuffle=True)
coef_folds = []
for train_idx, test_idx in cv.split(X1, y1):
clf.fit(X1[train_idx], y=y1[train_idx])
#scores1.append(clf.score(X1[test_idx], y=y1[test_idx]))
coef_folds.append(get_coef(clf, attr='patterns_', inverse_transform=True))
coef = np.asarray(coef_folds).mean(0).reshape([173, -1]) #mean folds and reshape
evoked = EvokedArray(coef, sleep_epochs.info, tmin=tmin)
evokeds.append(evoked)
ga = mne.grand_average(evokeds)
#SLEEP
f = ga.plot_topomap([0., 0.2, 0.4, 0.6, 0.8], scalings=0.1, vmin=-2, vmax=2)
#f = ga.plot_topomap([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8], scalings=0.1, vmin=-6, vmax=6)
#f = ga.plot_topomap([0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5], scalings=0.1, vmin=-6, vmax=6)
f.savefig(os.path.join(save_path, contrast1[0].split('/')[1] + 'coefs_new_detail.tiff'))
#WAKE
# f = ga.plot_topomap([0., 0.2, 0.4, 0.6, 0.8], scalings=0.1, vmin=-2, vmax=2)
ax.axhline(.5, color='k', linestyle='--', label='chance')
ax.set_xlabel('Times')
ax.set_ylabel('AUC') # Area Under the Curve
ax.legend()
ax.axvline(.0, color='k', linestyle='-')
ax.set_title('Sensor space decoding')
###############################################################################
# You can retrieve the spatial filters and spatial patterns if you explicitly
# use a LinearModel
clf = make_pipeline(StandardScaler(),
LinearModel(LogisticRegression(solver='lbfgs')))
time_decod = SlidingEstimator(clf, n_jobs=1, scoring='roc_auc', verbose=True)
time_decod.fit(X, y)
coef = get_coef(time_decod, 'patterns_', inverse_transform=True)
evoked = mne.EvokedArray(coef, epochs.info, tmin=epochs.times[0])
joint_kwargs = dict(ts_args=dict(time_unit='s'),
topomap_args=dict(time_unit='s'))
evoked.plot_joint(times=np.arange(0., .500, .100), title='patterns',
**joint_kwargs)
###############################################################################
# Temporal generalization
# ^^^^^^^^^^^^^^^^^^^^^^^
#
# Temporal generalization is an extension of the decoding over time approach.
# It consists in evaluating whether the model estimated at a particular
# time instant accurately predicts any other time instant. It is analogous to
# transferring a trained model to a distinct learning problem, where the
# problems correspond to decoding the patterns of brain activity recorded at
coef = scaler.inverse_transform([coef])[0]
# The data was vectorized to fit a single model across all time points and
# all channels. We thus reshape it:
coef = coef.reshape(len(meg_epochs.ch_names), -1)
# Plot
evoked = EvokedArray(coef, meg_epochs.info, tmin=epochs.tmin)
evoked.plot_topomap(title='MEG %s' % name)
###############################################################################
# Let's do the same on EEG data using a scikit-learn pipeline
X = epochs.pick_types(meg=False, eeg=True)
y = epochs.events[:, 2]
# Define a unique pipeline to sequentially:
clf = make_pipeline(
Vectorizer(), # 1) vectorize across time and channels
StandardScaler(), # 2) normalize features across trials
LinearModel(LogisticRegression())) # 3) fits a logistic regression
clf.fit(X, y)
# Extract and plot patterns and filters
for name in ('patterns_', 'filters_'):
# The `inverse_transform` parameter will call this method on any estimator
# contained in the pipeline, in reverse order.
coef = get_coef(clf, name, inverse_transform=True)
evoked = EvokedArray(coef, epochs.info, tmin=epochs.tmin)
evoked.plot_topomap(title='EEG %s' % name[:-1])
# *** 2.1) create bootstrap sample ***
# extract random subjects from overall sample
resampled_subjects = random.choice(range(betas_cue[0, ...].shape[0]),
betas_cue[0, ...].shape[0],
replace=True)
# resampled betas
resampled_betas = betas_cue[0, ...][resampled_subjects, :]
# *** 2.2) estimate effect of moderator (i.e., PBI) on group-level ***
# set up and fit moderator (i.e., PBI) model using bootstrap sample
model_boot = LinearRegression(fit_intercept=False)
model_boot.fit(X=group.iloc[resampled_subjects], y=resampled_betas)
# extract regression coefficients
group_coefs = get_coef(model_boot, 'coef_')
# save bootstrap betas
for pred_i, predictor in enumerate(group.columns):
if 'pbi_rt' in predictor:
# store regression coefficient for moderator (i.e., PBI)
group_pp_betas[i, :] = group_coefs[:, pred_i]
# remove prev object
del resampled_betas
# *** 2.3) compute test statistic for bootstrap sample ***
# compute standard error
resampled_betas = betas_null[resampled_subjects, :]
se = resampled_betas.std(axis=0) / np.sqrt(resampled_betas.shape[0])
else:
sel = np.where(~np.isnan(y))[0]
td = SlidingEstimator(clf,
scoring=make_scorer(scorer),
n_jobs=24,
**kwargs)
# run decoding
cv = StratifiedKFold(8)
scores = list()
patterns = list()
filters = list()
for train, test in cv.split(X[sel], y[sel]):
td.fit(X[sel][train], y[sel][train])
score = td.score(X[sel][test], y[sel][test])
scores.append(score)
patterns.append(get_coef(td, 'patterns_', inverse_transform=True))
filters.append(get_coef(td, 'filters_', inverse_transform=True))
scores = np.mean(scores, axis=0)
patterns = np.mean(patterns, axis=0)
filters = np.mean(filters, axis=0)
if 'angle' in analysis:
patterns = np.mean(np.abs(patterns), axis=1)
filters = np.mean(np.abs(filters), axis=1)
scores = np.reshape(scores, (n_freqs, n_times))
patterns = np.reshape(patterns, (n_channels, n_freqs, n_times))
filters = np.reshape(filters, (n_channels, n_freqs, n_times))
# save cross-validated scores
fname = results_folder +\
'%s_tf_scores_%s_%s.npy' % (subject, epoch_type, analysis)
np.save(fname, np.array(scores))
fname = results_folder +\
# customized the temporal decoding process
# define 10-fold cross validation
cv = StratifiedShuffleSplit(n_splits=10, random_state=12345)
coefs = []
scores = []
for time_ in tqdm(range(data.shape[-1]), desc='temporal decoding'):
coef = []
scores_ = []
# at each time point, we use the frequency information in each channel as the features
for train, test in cv.split(data, labels):
data_ = data[train, :, :, time_]
clf = make_clf(True, True)
clf.fit(data_, labels[train])
# print(time_,metrics.classification_report(clf.predict(data[test,:,:,time_]),labels[test]))
# get the patterns decoded by the classifier
coef_ = get_coef(clf, 'patterns_', True)
coef.append(coef_)
temp = metrics.roc_auc_score(
labels[test],
clf.predict_proba(data[test, :, :, time_])[:, -1])
scores_.append(temp)
print('\n', '%d' % time_, 'auc = ', np.mean(scores_), '\n')
coefs.append(np.array(coef))
scores.append(scores_)
coefs = np.array(coefs)
scores = np.array(scores)
# get info object to plot the "pattern"
temp_epochs = mne.read_epochs(working_dir + 'sub5_d2-eventsRelated-epo.fif')
info = temp_epochs.info
import pickle
pickle.dump([scores, info, coefs], open(saving_dir + 'score_info_coefs.p',
# --- 2) vectorize (eeg-channel) data for linear regression analysis ---
# data to be analysed
data = subject.get_data()
# vectorize data across channels
Y = Vectorizer().fit_transform(data)
# --- 3) fit linear model with sklearn's LinearRegression ---
# we already have an intercept column in the design matrix,
# thus we'll call LinearRegression with fit_intercept=False
linear_model = LinearRegression(fit_intercept=False)
linear_model.fit(design, Y)
# --- 4) extract the resulting coefficients (i.e., betas) ---
# extract betas
coefs = get_coef(linear_model, 'coef_')
# only keep relevant predictor
betas[iteration, :] = coefs[:, pred_col]
# calculate coefficient of determination (r-squared)
r_squared[iteration, :] = r2_score(Y,
linear_model.predict(design),
multioutput='raw_values')
# clean up
del linear_model
###############################################################################
# create design matrix from group-level regression
# z-core age predictor
def run_gat(subj, decoder="ridge", n_jobs=2):
"""
Function to run Generalization Across Time (GAT).
Parameters
----------
subj: int
decoder: str
Specify type of classifier -'ridge' for Ridge Regression (default),
'lin-svm' for linear SVM 'svm' for nonlinear (RBF) SVM and 'log_reg'
for Logistic Regression
n_jobs: int
The number of jobs to run in parallel.
"""
# load cue A and cue B epochs
epochs = get_epochs(subj)['Correct A', 'Correct B']
# specify whether to use a linear or nonlinear SVM if SVM is used
lin = '' # if not svm it doesn't matter, both log_reg and ridge are linear
if "svm" in decoder:
decoder, lin = decoder.split("-")
# build classifier pipeline #
# pick a machine learning algorithm to use (ridge/SVM/logistic regression)
decoder_dict = {
"ridge":
RidgeClassifier(class_weight='balanced', random_state=42,
solver="sag"),
"svm":
SVC(class_weight='balanced',
kernel=("rbf" if "non" in lin else "linear"),
random_state=42),
"log_reg":
LogisticRegression(class_weight='balanced', random_state=42)
}
# get data and targets
data = epochs.get_data()
labels = epochs.events[:, -1]
# create classifier pipeline
clf = make_pipeline(StandardScaler(), decoder_dict[decoder])
gen_clf = GeneralizingEstimator(clf, scoring="roc_auc", n_jobs=n_jobs)
# compute cross validated performance scores
scores = cross_val_multiscore(gen_clf, data, labels, cv=5,
n_jobs=n_jobs).mean(0)
# calculate prediction confidence scores
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
preds = np.empty((len(labels), data.shape[2], data.shape[2]))
for train, test in cv.split(data, labels):
gen_clf.fit(data[train], labels[train])
d = gen_clf.decision_function(data[test])
preds[test] = d
# compute topographical patterns
dat = Vectorizer().fit_transform(data)
clf.fit(dat, labels)
dat = dat - dat.mean(0, keepdims=True)
# look for the type of classifier and get the weights
if decoder == 'ridge':
filt_ = clf.named_steps.ridgeclassifier.coef_.copy()
elif decoder == 'svm':
filt_ = clf.named_steps.svc.coef_.copy()
elif decoder == 'log_reg':
filt_ = clf.named_steps.logisticregression.coef_.copy()
# Compute patterns using Haufe's trick: A = Cov_X . W . Precision_Y
# cf.Haufe, et al., 2014, NeuroImage,
# doi:10.1016/j.neuroimage.2013.10.067)
inv_y = 1.
patt_ = np.cov(dat.T).dot(filt_.T.dot(inv_y)).T
# store the patterns accordingly
if decoder == 'ridge':
clf.named_steps.ridgeclassifier.patterns_ = patt_
elif decoder == 'svm':
clf.named_steps.svc.patterns_ = patt_
elif decoder == 'log_reg':
clf.named_steps.logisticregression.patterns_ = patt_
# back transform using steps in pipeline
patterns = get_coef(clf, 'patterns_', inverse_transform=True)
# return subject scores, prediction confidence and topographical patterns
return scores, preds, patterns