def trainer(features, cfg, psd_params, epochs, cls_params):
""" Train classifiers using computed features """
psd_subjects = {}
if 'fmin' in psd_params and 'fmax' in psd_params:
for subject in features:
psd_subjects[subject] = psd_params
else:
for subject in features:
psd_subjects[subject] = psd_params[subject]
cls1 = {}
cls2 = {}
for subject in features:
print('Training %s' % subject)
X1 = features[subject]['X1']
X2 = features[subject]['X2']
Y = features[subject]['Y']
gbp = cls_params[subject]
cls1[subject] = GradientBoostingClassifier(n_estimators=gbp['trees'], learning_rate=gbp['learning_rate'], max_depth=gbp['max_depth'], max_features=gbp['max_features'], subsample=gbp['subsample'], random_state=gbp['random_state'])
cls1[subject].fit(X1, Y)
cls1[subject].n_jobs = 1 # set to 1 for testing
cls2[subject] = GradientBoostingClassifier(n_estimators=gbp['trees'], learning_rate=gbp['learning_rate'], max_depth=gbp['max_depth'], max_features=gbp['max_features'], subsample=gbp['subsample'], random_state=gbp['random_state'])
cls2[subject].fit(X2, Y)
cls2[subject].n_jobs = 1 # set to 1 for testing
clsfile = '%s/classifiers.pkl' % MY_PATH
qc.save_obj(clsfile, dict(cls1=cls1, cls2=cls2, cfg=cfg, psd_params=psd_subjects, epochs=epochs, cls_params=cls_params))
print('Classifiers exported to %s' % clsfile)
def train_decoder(cfg, featdata, feat_file=None):
"""
Train the final decoder using all data
"""
# Init a classifier
selected_classifier = cfg.CLASSIFIER['selected']
if selected_classifier == 'GB':
cls = GradientBoostingClassifier(
loss='deviance',
learning_rate=cfg.CLASSIFIER[selected_classifier]['learning_rate'],
n_estimators=cfg.CLASSIFIER[selected_classifier]['trees'],
subsample=1.0,
max_depth=cfg.CLASSIFIER[selected_classifier]['depth'],
random_state=cfg.CLASSIFIER[selected_classifier]['seed'],
max_features='sqrt',
verbose=0,
warm_start=False,
presort='auto')
elif selected_classifier == 'XGB':
cls = XGBClassifier(
loss='deviance',
learning_rate=cfg.CLASSIFIER[selected_classifier]['learning_rate'],
n_estimators=cfg.CLASSIFIER[selected_classifier]['trees'],
subsample=1.0,
max_depth=cfg.CLASSIFIER[selected_classifier]['depth'],
random_state=cfg.GB['seed'],
max_features='sqrt',
verbose=0,
warm_start=False,
presort='auto')
elif selected_classifier == 'RF':
cls = RandomForestClassifier(
n_estimators=cfg.CLASSIFIER[selected_classifier]['trees'],
max_features='auto',
max_depth=cfg.CLASSIFIER[selected_classifier]['depth'],
n_jobs=cfg.N_JOBS,
random_state=cfg.CLASSIFIER[selected_classifier]['seed'],
oob_score=False,
class_weight='balanced_subsample')
elif selected_classifier == 'LDA':
cls = LDA()
elif selected_classifier == 'rLDA':
cls = rLDA(cfg.CLASSIFIER[selected_classifier][r_coeff])
else:
logger.error('Unknown classifier %s' % selected_classifier)
raise ValueError
# Setup features
X_data = featdata['X_data']
Y_data = featdata['Y_data']
wlen = featdata['wlen']
if cfg.FEATURES['PSD']['wlen'] is None:
cfg.FEATURES['PSD']['wlen'] = wlen
w_frames = featdata['w_frames']
ch_names = featdata['ch_names']
X_data_merged = np.concatenate(X_data)
Y_data_merged = np.concatenate(Y_data)
if cfg.CV['BALANCE_SAMPLES']:
X_data_merged, Y_data_merged = balance_samples(
X_data_merged,
Y_data_merged,
cfg.CV['BALANCE_SAMPLES'],
verbose=True)
# Start training the decoder
logger.info_green('Training the decoder')
timer = qc.Timer()
cls.n_jobs = cfg.N_JOBS
cls.fit(X_data_merged, Y_data_merged)
logger.info('Trained %d samples x %d dimension in %.1f sec' %\
(X_data_merged.shape[0], X_data_merged.shape[1], timer.sec()))
cls.n_jobs = 1 # always set n_jobs=1 for testing
# Export the decoder
classes = {c: cfg.tdef.by_value[c] for c in np.unique(Y_data)}
if cfg.FEATURES['selected'] == 'PSD':
data = dict(cls=cls,
ch_names=ch_names,
psde=featdata['psde'],
sfreq=featdata['sfreq'],
picks=featdata['picks'],
classes=classes,
epochs=cfg.EPOCH,
w_frames=w_frames,
w_seconds=cfg.FEATURES['PSD']['wlen'],
wstep=cfg.FEATURES['PSD']['wstep'],
spatial=cfg.SP_FILTER,
spatial_ch=featdata['picks'],
spectral=cfg.TP_FILTER[cfg.TP_FILTER['selected']],
spectral_ch=featdata['picks'],
notch=cfg.NOTCH_FILTER[cfg.NOTCH_FILTER['selected']],
notch_ch=featdata['picks'],
multiplier=cfg.MULTIPLIER,
ref_ch=cfg.REREFERENCE[cfg.REREFERENCE['selected']],
decim=cfg.FEATURES['PSD']['decim'])
clsfile = '%s/classifier/classifier-%s.pkl' % (cfg.DATA_PATH,
platform.architecture()[0])
qc.make_dirs('%s/classifier' % cfg.DATA_PATH)
qc.save_obj(clsfile, data)
logger.info('Decoder saved to %s' % clsfile)
# Reverse-lookup frequency from FFT
fq = 0
if type(cfg.FEATURES['PSD']['wlen']) == list:
fq_res = 1.0 / cfg.FEATURES['PSD']['wlen'][0]
else:
fq_res = 1.0 / cfg.FEATURES['PSD']['wlen']
fqlist = []
while fq <= cfg.FEATURES['PSD']['fmax']:
if fq >= cfg.FEATURES['PSD']['fmin']:
fqlist.append(fq)
fq += fq_res
# Show top distinctive features
if cfg.FEATURES['selected'] == 'PSD':
logger.info_green('Good features ordered by importance')
if selected_classifier in ['RF', 'GB', 'XGB']:
keys, values = qc.sort_by_value(list(cls.feature_importances_),
rev=True)
elif selected_classifier in ['LDA', 'rLDA']:
keys, values = qc.sort_by_value(cls.w, rev=True)
keys = np.array(keys)
values = np.array(values)
if cfg.EXPORT_GOOD_FEATURES:
if feat_file is None:
gfout = open('%s/classifier/good_features.txt' % cfg.DATA_PATH,
'w')
else:
gfout = open(feat_file, 'w')
if type(wlen) is not list:
ch_names = [ch_names[c] for c in featdata['picks']]
else:
ch_names = []
for w in range(len(wlen)):
for c in featdata['picks']:
ch_names.append('w%d-%s' % (w, ch_names[c]))
chlist, hzlist = features.feature2chz(keys, fqlist, ch_names=ch_names)
valnorm = values[:cfg.FEAT_TOPN].copy()
valsum = np.sum(valnorm)
if valsum == 0:
valsum = 1
valnorm = valnorm / valsum * 100.0
# show top-N features
for i, (ch, hz) in enumerate(zip(chlist, hzlist)):
if i >= cfg.FEAT_TOPN:
break
txt = '%-3s %5.1f Hz normalized importance %-6s raw importance %-6s feature %-5d' %\
(ch, hz, '%.2f%%' % valnorm[i], '%.2f%%' % (values[i] * 100.0), keys[i])
logger.info(txt)
if cfg.EXPORT_GOOD_FEATURES:
gfout.write('Importance(%) Channel Frequency Index\n')
for i, (ch, hz) in enumerate(zip(chlist, hzlist)):
gfout.write('%.3f\t%s\t%s\t%d\n' %
(values[i] * 100.0, ch, hz, keys[i]))
gfout.close()
def balance_tpr(cfg, featdata):
"""
Find the threshold of class index 0 that yields equal number of true positive samples of each class.
Currently only available for binary classes.
Params
======
cfg: config module
feetdata: feature data computed using compute_features()
"""
n_jobs = cfg.N_JOBS
if n_jobs is None:
n_jobs = mp.cpu_count()
if n_jobs > 1:
logger.info('balance_tpr(): Using %d cores' % n_jobs)
pool = mp.Pool(n_jobs)
results = []
# Init a classifier
selected_classifier = cfg.CLASSIFIER[cfg.CLASSIFIER['selected']]
if selected_classifier == 'GB':
cls = GradientBoostingClassifier(
loss='deviance',
learning_rate=cfg.CLASSIFIER['GB']['learning_rate'],
n_estimators=cfg.CLASSIFIER['GB']['trees'],
subsample=1.0,
max_depth=cfg.CLASSIFIER['GB']['depth'],
random_state=cfg.CLASSIFIER[selected_classifier]['seed'],
max_features='sqrt',
verbose=0,
warm_start=False,
presort='auto')
elif selected_classifier == 'XGB':
cls = XGBClassifier(
loss='deviance',
learning_rate=cfg.CLASSIFIER['XGB']['learning_rate'],
n_estimators=cfg.CLASSIFIER['XGB']['trees'],
subsample=1.0,
max_depth=cfg.CLASSIFIER['XGB']['depth'],
random_state=cfg.CLASSIFIER['XGB']['seed'],
max_features='sqrt',
verbose=0,
warm_start=False,
presort='auto')
elif selected_classifier == 'RF':
cls = RandomForestClassifier(
n_estimators=cfg.CLASSIFIER['RF']['trees'],
max_features='auto',
max_depth=cfg.CLASSIFIER['RF']['depth'],
n_jobs=cfg.N_JOBS,
random_state=cfg.CLASSIFIER['RF']['seed'],
oob_score=False,
class_weight='balanced_subsample')
elif selected_classifier == 'LDA':
cls = LDA()
elif selected_classifier == 'rLDA':
cls = rLDA(cfg.CLASSIFIER['rLDA'])
else:
logger.error('Unknown classifier type %s' % selected_classifier)
raise ValueError
# Setup features
X_data = featdata['X_data']
Y_data = featdata['Y_data']
wlen = featdata['wlen']
if cfg.CLASSIFIER['PSD']['wlen'] is None:
cfg.CLASSIFIER['PSD']['wlen'] = wlen
# Choose CV type
ntrials, nsamples, fsize = X_data.shape
selected_CV = cfg.CV_PERFORM[cfg.CV_PERFORM['selected']]
if cselected_CV == 'LeaveOneOut':
logger.info_green('\n%d-fold leave-one-out cross-validation' % ntrials)
if SKLEARN_OLD:
cv = LeaveOneOut(len(Y_data))
else:
cv = LeaveOneOut()
elif selected_CV == 'StratifiedShuffleSplit':
logger.info_green(
'\n%d-fold stratified cross-validation with test set ratio %.2f' %
(cfg.CV_PERFORM[selected_CV]['folds'],
cfg.CV_PERFORM[selected_CV]['test_ratio']))
if SKLEARN_OLD:
cv = StratifiedShuffleSplit(
Y_data[:, 0],
cfg.CV_PERFORM[selected_CV]['folds'],
test_size=cfg.CV_PERFORM[selected_CV]['test_ratio'],
random_state=cfg.CV_PERFORM[selected_CV]['random_seed'])
else:
cv = StratifiedShuffleSplit(
n_splits=cfg.CV_PERFORM[selected_CV]['folds'],
test_size=cfg.CV_PERFORM[selected_CV]['test_ratio'],
random_state=cfg.CV_PERFORM[selected_CV]['random_seed'])
else:
logger.error('%s is not supported yet. Sorry.' % selected_CV)
raise NotImplementedError
logger.info('%d trials, %d samples per trial, %d feature dimension' %
(ntrials, nsamples, fsize))
# For classifier itself, single core is usually faster
cls.n_jobs = 1
Y_preds = []
if SKLEARN_OLD:
splits = cv
else:
splits = cv.split(X_data, Y_data[:, 0])
for cnum, (train, test) in enumerate(splits):
X_train = np.concatenate(X_data[train])
X_test = np.concatenate(X_data[test])
Y_train = np.concatenate(Y_data[train])
Y_test = np.concatenate(Y_data[test])
if n_jobs > 1:
results.append(
pool.apply_async(
get_predict_proba,
[cls, X_train, Y_train, X_test, Y_test, cnum + 1]))
else:
Y_preds.append(
get_predict_proba(cls, X_train, Y_train, X_test, Y_test,
cnum + 1))
cnum += 1
# Aggregate predictions
if n_jobs > 1:
pool.close()
pool.join()
for r in results:
Y_preds.append(r.get())
Y_preds = np.concatenate(Y_preds, axis=0)
# Find threshold for class index 0
Y_preds = sorted(Y_preds)
mid_idx = int(len(Y_preds) / 2)
if len(Y_preds) == 1:
return 0.5 # should not reach here in normal conditions
elif len(Y_preds) % 2 == 0:
thres = Y_preds[mid_idx -
1] + (Y_preds[mid_idx] - Y_preds[mid_idx - 1]) / 2
else:
thres = Y_preds[mid_idx]
return thres
