def dotheclassification(ds,
classifier,
bilateral
):
""" Dotheclassification does the classification.
Input: the dataset on which to perform a leave-one-out crossvalidation with a classifier
of choice.
Specify: the classifier to be used (gnb (linear gnb), l-sgd (linear sgd), sgd)
whether the sensitivities should be computed and stored for later use
whether the dataset has ROIs combined across hemisphere (bilateral)
"""
if classifier == 'gnb':
# set up classifier
prior = 'ratio'
if bilateral:
targets = 'bilat_ROIs'
else:
targets = 'all_ROIs'
clf = mv.GNB(common_variance=True,
prior=prior,
space=targets)
elif classifier == 'l-sgd':
# set up the dataset: If I understand the sourcecode correctly, the
# Stochastic Gradient Descent wants to have unique labels in a sample attribute
# called 'targets' and is quite stubborn with this name - I could not convince
# it to look for targets somewhere else, so now I catering to his demands
if bilateral:
ds.sa['targets'] = ds.sa.bilat_ROIs
else:
ds.sa['targets'] = ds.sa.all_ROIs
# necessary I believe regardless of the SKLLearnerAdapter
from sklearn.linear_model import SGDClassifier
# get a stochastic gradient descent into pymvpa by using the SKLLearnerAdapter.
# Get it to perform 1 vs 1 decisions (instead of one vs all) with the MulticlassClassifier
clf = mv.MulticlassClassifier(mv.SKLLearnerAdapter(SGDClassifier(loss='hinge',
penalty='l2',
class_weight='balanced'
)))
cv = mv.CrossValidation(clf, mv.NFoldPartitioner(attr='participant'),
errorfx=mv.mean_match_accuracy,
enable_ca=['stats'])
results = cv(ds)
return cv
def dotheclassification(ds_movie,
ds_loc,
classifier,
bilateral):
""" Dotheclassification does the classification.
Input: the dataset on which to perform a leave-one-out crossvalidation with a classifier
of choice.
Specify: the classifier to be used (gnb (linear gnb), l-sgd (linear sgd), sgd)
whether the sensitivities should be computed and stored for later use
whether the dataset has ROIs combined across hemisphere (bilateral)
"""
dfs = []
for idx, ds in enumerate([ds_movie, ds_loc]):
if bilateral:
ds.sa['targets'] = ds.sa.bilat_ROIs
else:
ds.sa['targets'] = ds.sa.all_ROIs
if classifier == 'gnb':
# set up classifier
prior = 'ratio'
clf = mv.GNB(common_variance=True,
prior=prior)
elif classifier == 'sgd':
# necessary I believe regardless of the SKLLearnerAdapter
from sklearn.linear_model import SGDClassifier
clf = mv.SKLLearnerAdapter(SGDClassifier(loss='hinge',
penalty='l2',
class_weight='balanced'))
elif classifier == 'l-sgd':
# necessary I believe regardless of the SKLLearnerAdapter
from sklearn.linear_model import SGDClassifier
# get a stochastic gradient descent into pymvpa by using the SKLLearnerAdapter.
# Get it to perform 1 vs 1 decisions (instead of one vs all) with the MulticlassClassifier
clf = mv.MulticlassClassifier(mv.SKLLearnerAdapter(SGDClassifier(loss='hinge',
penalty='l2',
class_weight='balanced'
)))
# prepare for callback of sensitivity extraction within CrossValidation
classifications = []
def store_class(data, node, result):
# import pdb; pdb.set_trace()
class_ds = mv.Dataset(samples=data.sa.voxel_indices)
class_ds.sa['targets'] = data.sa.targets
class_ds.sa['partitions'] = data.sa.partitions
class_ds.sa['predictions'] = clf.predict(data)
class_ds.sa['participant'] = data.sa.participant
classifications.append(class_ds)
# do a crossvalidation classification and store the classification results
cv = mv.CrossValidation(clf, mv.NFoldPartitioner(attr='participant'),
errorfx=mv.mean_match_accuracy,
enable_ca=['stats'],
callback=store_class)
# import pdb; pdb.set_trace()
results = cv(ds)
# import pdb; pdb.set_trace()
# save classification results as a Dataset
ds_type = ['movie', 'loc']
mv.h5save(results_dir + 'cv_classification_results_{}.hdf5'.format(ds_type[idx]), classifications)
print('Saved the classification results obtained during crossvalidation.')
# get the classification list into a pandas dataframe
for i, classification in enumerate(classifications):
df = pd.DataFrame(data={'voxel_indices': list(classification.samples),
'targets': list(classification.sa.targets),
'predictions': list(classification.sa.predictions),
'partitions': list(classification.sa.partitions),
'participants': list(classification.sa.participant),
'ds_type': [ds_type[idx]] * len(classification.sa.predictions)
}
)
dfs.append(df)
# two helper functions for later use in a lamda function
def hits(row):
if row['predictions'] == row['targets']:
return 1
else:
return 0
def parts(row):
if row['partitions'] == 1:
return "train"
elif row['partitions'] == 2:
return "test"
# get all folds into one dataframe, disregard the index
all_classifications = pd.concat(dfs, ignore_index=True)
# compute hits as correspondence between target and prediction
all_classifications['hits'] = all_classifications.apply(lambda row: hits(row), axis=1)
# assign string labels to testing and training partitions (instead of 1, 2)
all_classifications['parts'] = all_classifications.apply(lambda row: parts(row), axis=1)
# transform voxel coordinates from arrays (unhashable) into tuples
all_classifications['voxel_indices'] = all_classifications['voxel_indices'].apply(tuple)
# subset the dataset to contain only the testing data
all_testing = all_classifications[all_classifications.parts == "test"]
# check that every participant is in the data
assert len(all_testing.participants.unique()) == 15
# to check for correspondence between the sum of the two experiments confusion matrices,
# do sth like this: len(all_testing[(all_testing['predictions'] == 'PPA') & (all_testing['targets'] == 'VIS')])
# this counts hits per fold across experiments (2 if both experiments classified correctly,
# 1 if 1 experiment classified correctly, 0 is none did). Also, append the targets per voxel.
# we use 'min' here because aggregate needs any function, but targets are the same between
# the experiments
compare_exp = all_testing.groupby(['voxel_indices', 'participants']).agg(
{'hits': 'sum', 'targets': 'min'}).reset_index().sort_values(['voxel_indices', 'participants'])
all_testing_movie = all_testing[all_testing.ds_type == 'movie'].sort_values(
['voxel_indices', 'participants']).reset_index()
all_testing_loc = all_testing[all_testing.ds_type == 'loc'].sort_values(
['voxel_indices', 'participants']).reset_index()
# append movie and loc predictions to the dataframe
compare_exp['pred_movie'] = all_testing_movie.predictions
compare_exp['pred_loc'] = all_testing_loc.predictions
# get the ROIS from the classification
ROIS = np.unique(ds_movie.sa.targets)
# there can't be values greater than two or lower than zero
assert compare_exp.hits.max() <= 2
assert compare_exp.hits.min() >= 0
return compare_exp, all_testing, ROIS
def dotheclassification(ds, bilateral, store_sens=True):
""" Dotheclassification does the classification. It builds a
linear gaussian naive bayes classifier, performs a leave-one-out
crossvalidation and stores the sensitivities from the GNB classifier of each
fold in a combined dataset for further use in a glm.
If sens == False, the sensitivities are not stored, and only a
classification is performed"""
import matplotlib.pyplot as plt
# set up classifier
prior = 'ratio'
if bilateral:
targets = 'bilat_ROIs'
else:
targets = 'all_ROIs'
gnb = mv.GNB(common_variance=True, prior=prior, space=targets)
# prepare for callback of sensitivity extraction within CrossValidation
sensitivities = []
if store_sens:
def store_sens(data, node, result):
sens = node.measure.get_sensitivity_analyzer(
force_train=False)(data)
# we also need to manually append the time attributes to the sens ds
sens.fa['time_coords'] = data.fa['time_coords']
sens.fa['chunks'] = data.fa['chunks']
sensitivities.append(sens)
# do a crossvalidation classification
cv = mv.CrossValidation(gnb,
mv.NFoldPartitioner(attr='participant'),
errorfx=mv.mean_match_accuracy,
enable_ca=['stats'],
callback=store_sens)
else:
cv = mv.CrossValidation(gnb,
mv.NFoldPartitioner(attr='participant'),
errorfx=mv.mean_match_accuracy,
enable_ca=['stats'])
results = cv(ds)
# save classification results
with open(results_dir + 'avmovie_clf.txt', 'a') as f:
f.write(cv.ca.stats.as_string(description=True))
# printing of the confusion matrix
if bilateral:
desired_order = ['VIS', 'LOC', 'OFA', 'FFA', 'EBA', 'PPA']
else:
desired_order = [
'brain', 'VIS', 'left LOC', 'right LOC', 'left OFA', 'right OFA',
'left FFA', 'right FFA', 'left EBA', 'right EBA', 'left PPA',
'right PPA'
]
labels = get_known_labels(desired_order, cv.ca.stats.labels)
# plot the confusion matrix with pymvpas build-in plot function currently fails
# cv.ca.stats.plot(labels=labels,
# numbers=True,
# cmap='gist_heat_r')
# plt.savefig(results_dir + 'confusion_matrix.png')
if niceplot:
ACC = cv.ca.stats.stats['mean(ACC)']
plot_confusion(cv,
labels,
fn=results_dir + 'confusion_matrix_avmovie.svg',
figsize=(9, 9),
vmax=100,
cmap='Blues',
ACC='%.2f' % ACC)
mv.h5save(results_dir + 'gnb_cv_classification_results.hdf5', results)
print('Saved the crossvalidation results.')
if store_sens:
mv.h5save(results_dir + 'sensitivities_nfold.hdf5', sensitivities)
print('Saved the sensitivities.')
# results now has the overall accuracy. results.samples gives the
# accuracy per participant.
# sensitivities contains a dataset for each participant with the
# sensitivities as samples and class-pairings as attributes
return sensitivities, cv
# print len(dataset.targets) # print dataset.chunks # print len(dataset.chunks) # REDUCE TO CLASS LABELS, AND ONLY KEEP CONDITIONS OF INTEREST (JAPANESE VS ENGLISH) dataset.targets = [t[0:2] for t in dataset.targets] dataset = dataset[N.array([l in ['jj', 'je', 'ej', 'ee'] for l in dataset.sa.targets], dtype='bool')] print '... and only',dataset.shape[0],'cases of interest (Language Switch between Japanese vs English)' dataset=M.datasets.miscfx.remove_invariant_features(dataset) print 'saving as compressed file',trimmedCache pickleFile = gzip.open(trimmedCache, 'wb', 5); pickle.dump(dataset, pickleFile); anovaSelectedSMLR = M.FeatureSelectionClassifier( M.GNB(common_variance=True), M.SensitivityBasedFeatureSelection( M.OneWayAnova(), M.FixedNElementTailSelector(500, mode='select', tail='upper') ), ) foldwiseCvedAnovaSelectedSMLR = M.CrossValidation( anovaSelectedSMLR, M.NFoldPartitioner(), enable_ca=['samples_error','stats', 'calling_time','confusion'] ) # run classifier print 'learning on detrended, normalised, averaged, Language Switch ...',datetime.datetime.now() results = foldwiseCvedAnovaSelectedSMLR(dataset) print '... done',datetime.datetime.now() print 'accuracy',N.round(100-N.mean(results)*100,1),'%',datetime.datetime.now()