def train_predict_test(subject,clf,X,X_test,enhance_size = 0):
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
data_grouped_by_hour = load_grouped_train_data('preprocessed/cnn/', subject, filenames_grouped_by_hour)
X, y = generate_overlapped_data(data_grouped_by_hour, overlap_size=10,
window_size=X.shape[-1],
overlap_interictal=True,
overlap_preictal=True)
X, scalers = scale_across_time(X, x_test=None)
X_test, _ = scale_across_time(X_test, x_test=None, scalers=scalers)
print X.shape
X = X.reshape(X.shape[0],X.shape[1]*X.shape[2]*X.shape[3])
X_test = X_test.reshape(X_test.shape[0],X_test.shape[1]*X_test.shape[2]*X_test.shape[3])
X,xt,y,yt = train_test_split(X,y,test_size = .25)
print "train size", X.shape
print "test_size", xt.shape
#print "done loading"
clf.fit(X)
preds_proba = clf.predict(X_test)
#print preds_proba.shape
validation_preds = clf.predict(xt)
return preds_proba,list(validation_preds),list(yt)
def train_predict_test_cnn(subject,clf,X,X_test,enhance_size = 0):
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
data_grouped_by_hour = load_grouped_train_data('preprocessed/cnn/', subject, filenames_grouped_by_hour)
X, y = generate_overlapped_data(data_grouped_by_hour, overlap_size=10,
window_size=X.shape[-1],
overlap_interictal=True,
overlap_preictal=True)
X, scalers = scale_across_time(X, x_test=None)
X_test, _ = scale_across_time(X_test, x_test=None, scalers=scalers)
# X,xt,y,yt = split_evenly(X,y,test_size = .5)
# if enhance_size > 0:
# X,y = enhance_data(X,y,enhance_size,cnn=True)
# xt,yt = enhance_data(xt,yt,enhance_size/2,cnn=True)
X, xt, y, yt = train_test_split(X, y, test_size=0.25, random_state=42)
print "train size", X.shape
print "test_size", xt.shape
preds_proba = np.zeros(( X.shape[1], X_test.shape[0] ))
val_proba = np.zeros(( xt.shape[1], xt.shape[0] ))
weighting = np.zeros((X.shape[1],))
for i in range(0, X.shape[1]):
print "Progress: " + str(100*i/X.shape[1]) + '%'
X_train = X[:,i,:,:]
xt_train = xt[:,i,:,:]
weighting[i], val_proba[i,] = clf.fit(X_train,y,xt_train,yt)
train_loss = np.array([])
valid_loss = np.array([])
X_test_subset = X_test[:,i,:,:]
preds_proba[i,] = clf.predict_proba(X_test_subset)
#idx = np.argmax(weighting)
sc = np.amax(weighting)
print "Best score:" + str(sc)
weighting -= weighting.min()
weighting /= weighting.sum()
#preds_proba = preds_proba[idx,]
preds_proba = np.average(preds_proba, axis=0, weights=weighting)
preds_scaled = preds_proba
#preds_scaled = min_max_scale(preds_proba)
#validation_preds = val_proba[idx,]
validation_preds = np.average(val_proba, axis=0, weights=weighting)
return preds_scaled,preds_proba,list(validation_preds),list(yt),train_loss,valid_loss
def plot_features(subject, data_path, model_path, test_labels, dataset='test'):
with open(model_path + '/' + subject + '.pickle', 'rb') as f:
state_dict = cPickle.load(f)
cnn = ConvNet(state_dict['params'])
cnn.set_weights(state_dict['weights'])
scalers = state_dict['scalers']
if dataset == 'test':
d = load_test_data(data_path, subject)
x = d['x']
y = test_labels['preictal']
elif dataset == 'train':
d = load_train_data(data_path, subject)
x, y = d['x'], d['y']
else:
raise ValueError('dataset')
x, _ = scale_across_time(x, x_test=None, scalers=scalers) if state_dict['params']['scale_time'] \
else scale_across_features(x, x_test=None, scalers=scalers)
cnn.batch_size.set_value(x.shape[0])
get_features = theano.function([cnn.x, Param(cnn.training_mode, default=0)], cnn.feature_extractor.output,
allow_input_downcast=True)
logits_test = get_features(x)
model = TSNE(n_components=2, random_state=0)
z = model.fit_transform(np.float64(logits_test))
plt.scatter(z[:, 0], z[:, 1], s=60, c=y)
plt.show()
def plot_train_probs(subject, data_path, model_path):
with open(model_path + '/' + subject + '.pickle', 'rb') as f:
state_dict = pickle.load(f)
cnn = ConvNet(state_dict['params'])
cnn.set_weights(state_dict['weights'])
scalers = state_dict['scalers']
d = load_train_data(data_path, subject)
x, y = d['x'], d['y']
x, _ = scale_across_time(x, x_test=None, scalers=scalers) if state_dict['params']['scale_time'] \
else scale_across_features(x, x_test=None, scalers=scalers)
cnn.batch_size.set_value(x.shape[0])
probs = cnn.get_test_proba(x)
fpr, tpr, threshold = roc_curve(y, probs)
c = np.sqrt((1 - tpr)**2 + fpr**2)
opt_threshold = threshold[np.where(c == np.min(c))[0]]
print(opt_threshold)
x_coords = np.zeros(len(y), dtype='float64')
rng = np.random.RandomState(42)
x_coords += rng.normal(0.0, 0.08, size=len(x_coords))
plt.scatter(x_coords, probs, c=y, s=60)
plt.title(subject)
plt.show()
def predict(subject, data_path, model_path, submission_path):
patient_filenames = [
filename for filename in os.listdir(model_path)
if subject in filename and filename.endswith('.pickle')
]
for filename in patient_filenames:
print(filename)
d = load_test_data(data_path, subject)
x, id = d['x'], d['id']
with open(model_path + '/' + filename, 'rb') as f:
state_dict = pickle.load(f)
scalers = state_dict['scalers']
x, _ = scale_across_time(x, x_test=None, scalers=scalers) if state_dict['params']['scale_time'] \
else scale_across_features(x, x_test=None, scalers=scalers)
cnn = ConvNet(state_dict['params'])
cnn.set_weights(state_dict['weights'])
test_proba = cnn.get_test_proba(x)
ans = list(zip(id, test_proba))
df = DataFrame(data=ans, columns=['clip', 'preictal'])
csv_name = '.'.join(
filename.split('.')[:-1]) if '.' in filename else filename
df.to_csv(submission_path + '/' + csv_name + '.csv',
index=False,
header=True)
def predict(subject, data_path, model_path, submission_path):
patient_filenames = [filename for filename in os.listdir(model_path) if
subject in filename and filename.endswith('.pickle')]
for filename in patient_filenames:
print filename
d = load_test_data(data_path, subject)
x, id = d['x'], d['id']
with open(model_path + '/' + filename, 'rb') as f:
state_dict = cPickle.load(f)
scalers = state_dict['scalers']
x, _ = scale_across_time(x, x_test=None, scalers=scalers) if state_dict['params']['scale_time'] \
else scale_across_features(x, x_test=None, scalers=scalers)
cnn = ConvNet(state_dict['params'])
cnn.set_weights(state_dict['weights'])
test_proba = cnn.get_test_proba(x)
ans = zip(id, test_proba)
df = DataFrame(data=ans, columns=['clip', 'preictal'])
csv_name = '.'.join(filename.split('.')[:-1]) if '.' in filename else filename
df.to_csv(submission_path + '/' + csv_name + '.csv', index=False, header=True)
def plot_train_probs(subject, data_path, model_path):
with open(model_path + "/" + subject + ".pickle", "rb") as f:
state_dict = cPickle.load(f)
cnn = ConvNet(state_dict["params"])
cnn.set_weights(state_dict["weights"])
scalers = state_dict["scalers"]
d = load_train_data(data_path, subject)
x, y = d["x"], d["y"]
x, _ = (
scale_across_time(x, x_test=None, scalers=scalers)
if state_dict["params"]["scale_time"]
else scale_across_features(x, x_test=None, scalers=scalers)
)
cnn.batch_size.set_value(x.shape[0])
probs = cnn.get_test_proba(x)
fpr, tpr, threshold = roc_curve(y, probs)
c = np.sqrt((1 - tpr) ** 2 + fpr ** 2)
opt_threshold = threshold[np.where(c == np.min(c))[0]]
print opt_threshold
x_coords = np.zeros(len(y), dtype="float64")
rng = np.random.RandomState(42)
x_coords += rng.normal(0.0, 0.08, size=len(x_coords))
plt.scatter(x_coords, probs, c=y, s=60)
plt.title(subject)
plt.show()
def train(subject, data_path, model_path, model_params, validation_params):
d = load_train_data(data_path, subject)
x, y, filename_to_idx = d['x'], d['y'], d['filename_to_idx']
x_test = load_test_data(data_path,
subject)['x'] if model_params['use_test'] else None
# --------- add params
model_params['n_channels'] = x.shape[1]
model_params['n_fbins'] = x.shape[2]
model_params['n_timesteps'] = x.shape[3]
print '============ parameters'
for key, value in model_params.items():
print key, ':', value
print '========================'
x_train, y_train = None, None
x_valid, y_valid = None, None
if model_params['overlap']:
# no validation if overlap
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
data_grouped_by_hour = load_grouped_train_data(
data_path, subject, filenames_grouped_by_hour)
x, y = generate_overlapped_data(data_grouped_by_hour,
overlap_size=model_params['overlap'],
window_size=x.shape[-1],
overlap_interictal=True,
overlap_preictal=True)
print x.shape
x, scalers = scale_across_time(x, x_test=None) if model_params['scale_time'] \
else scale_across_features(x, x_test=None)
cnn = ConvNet(model_params)
cnn.train(train_set=(x, y), max_iter=175000)
state_dict = cnn.get_state()
state_dict['scalers'] = scalers
with open(model_path + '/' + subject + '.pickle', 'wb') as f:
cPickle.dump(state_dict, f, protocol=cPickle.HIGHEST_PROTOCOL)
return
else:
if validation_params['random_split']:
skf = StratifiedShuffleSplit(y,
n_iter=1,
test_size=0.25,
random_state=0)
for train_idx, valid_idx in skf:
x_train, y_train = x[train_idx], y[train_idx]
x_valid, y_valid = x[valid_idx], y[valid_idx]
else:
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
d = split_train_valid_filenames(subject, filenames_grouped_by_hour)
train_filenames, valid_filenames = d['train_filenames'], d[
'valid_filnames']
train_idx = [filename_to_idx[i] for i in train_filenames]
valid_idx = [filename_to_idx[i] for i in valid_filenames]
x_train, y_train = x[train_idx], y[train_idx]
x_valid, y_valid = x[valid_idx], y[valid_idx]
if model_params['scale_time']:
x_train, scalers_train = scale_across_time(x=x_train, x_test=x_test)
x_valid, _ = scale_across_time(x=x_valid,
x_test=x_test,
scalers=scalers_train)
else:
x_train, scalers_train = scale_across_features(x=x_train,
x_test=x_test)
x_valid, _ = scale_across_features(x=x_valid,
x_test=x_test,
scalers=scalers_train)
del x, x_test
print '============ dataset'
print 'train:', x_train.shape
print 'n_pos:', np.sum(y_train), 'n_neg:', len(y_train) - np.sum(y_train)
print 'valid:', x_valid.shape
print 'n_pos:', np.sum(y_valid), 'n_neg:', len(y_valid) - np.sum(y_valid)
# -------------- validate
cnn = ConvNet(model_params)
best_iter = cnn.validate(train_set=(x_train, y_train),
valid_set=(x_valid, y_valid),
valid_freq=validation_params['valid_freq'],
max_iter=validation_params['max_iter'],
fname_out=model_path + '/' + subject + '.txt')
# ---------------- scale
d = load_train_data(data_path, subject)
x, y, filename_to_idx = d['x'], d['y'], d['filename_to_idx']
x_test = load_test_data(data_path,
subject)['x'] if model_params['use_test'] else None
x, scalers = scale_across_time(x=x, x_test=x_test) if model_params['scale_time'] \
else scale_across_features(x=x, x_test=x_test)
del x_test
cnn = ConvNet(model_params)
cnn.train(train_set=(x, y), max_iter=best_iter)
state_dict = cnn.get_state()
state_dict['scalers'] = scalers
with open(model_path + '/' + subject + '.pickle', 'wb') as f:
cPickle.dump(state_dict, f, protocol=cPickle.HIGHEST_PROTOCOL)
def train(subject, data_path, model_path, model_params, validation_params):
d = load_train_data(data_path, subject)
x, y, filename_to_idx = d['x'], d['y'], d['filename_to_idx']
x_test = load_test_data(data_path, subject)['x'] if model_params['use_test'] else None
# --------- add params
model_params['n_channels'] = x.shape[1]
model_params['n_fbins'] = x.shape[2]
model_params['n_timesteps'] = x.shape[3]
print '============ parameters'
for key, value in model_params.items():
print key, ':', value
print '========================'
x_train, y_train = None, None
x_valid, y_valid = None, None
if model_params['overlap']:
# no validation if overlap
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
data_grouped_by_hour = load_grouped_train_data(data_path, subject, filenames_grouped_by_hour)
x, y = generate_overlapped_data(data_grouped_by_hour, overlap_size=model_params['overlap'],
window_size=x.shape[-1],
overlap_interictal=True,
overlap_preictal=True)
print x.shape
x, scalers = scale_across_time(x, x_test=None) if model_params['scale_time'] \
else scale_across_features(x, x_test=None)
cnn = ConvNet(model_params)
cnn.train(train_set=(x, y), max_iter=175000)
state_dict = cnn.get_state()
state_dict['scalers'] = scalers
with open(model_path + '/' + subject + '.pickle', 'wb') as f:
cPickle.dump(state_dict, f, protocol=cPickle.HIGHEST_PROTOCOL)
return
else:
if validation_params['random_split']:
skf = StratifiedShuffleSplit(y, n_iter=1, test_size=0.25, random_state=0)
for train_idx, valid_idx in skf:
x_train, y_train = x[train_idx], y[train_idx]
x_valid, y_valid = x[valid_idx], y[valid_idx]
else:
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
d = split_train_valid_filenames(subject, filenames_grouped_by_hour)
train_filenames, valid_filenames = d['train_filenames'], d['valid_filnames']
train_idx = [filename_to_idx[i] for i in train_filenames]
valid_idx = [filename_to_idx[i] for i in valid_filenames]
x_train, y_train = x[train_idx], y[train_idx]
x_valid, y_valid = x[valid_idx], y[valid_idx]
if model_params['scale_time']:
x_train, scalers_train = scale_across_time(x=x_train, x_test=x_test)
x_valid, _ = scale_across_time(x=x_valid, x_test=x_test, scalers=scalers_train)
else:
x_train, scalers_train = scale_across_features(x=x_train, x_test=x_test)
x_valid, _ = scale_across_features(x=x_valid, x_test=x_test, scalers=scalers_train)
del x, x_test
print '============ dataset'
print 'train:', x_train.shape
print 'n_pos:', np.sum(y_train), 'n_neg:', len(y_train) - np.sum(y_train)
print 'valid:', x_valid.shape
print 'n_pos:', np.sum(y_valid), 'n_neg:', len(y_valid) - np.sum(y_valid)
# -------------- validate
cnn = ConvNet(model_params)
best_iter = cnn.validate(train_set=(x_train, y_train),
valid_set=(x_valid, y_valid),
valid_freq=validation_params['valid_freq'],
max_iter=validation_params['max_iter'],
fname_out=model_path + '/' + subject + '.txt')
# ---------------- scale
d = load_train_data(data_path, subject)
x, y, filename_to_idx = d['x'], d['y'], d['filename_to_idx']
x_test = load_test_data(data_path, subject)['x'] if model_params['use_test'] else None
x, scalers = scale_across_time(x=x, x_test=x_test) if model_params['scale_time'] \
else scale_across_features(x=x, x_test=x_test)
del x_test
cnn = ConvNet(model_params)
cnn.train(train_set=(x, y), max_iter=best_iter)
state_dict = cnn.get_state()
state_dict['scalers'] = scalers
with open(model_path + '/' + subject + '.pickle', 'wb') as f:
cPickle.dump(state_dict, f, protocol=cPickle.HIGHEST_PROTOCOL)
def train_predict_test_cnn(subject,clf,X,X_test,enhance_size = 0):
filenames_grouped_by_hour = cPickle.load(open('filenames.pickle'))
data_grouped_by_hour = load_grouped_train_data('preprocessed/cnn/', subject, filenames_grouped_by_hour)
X, y = generate_overlapped_data(data_grouped_by_hour, overlap_size=10,
window_size=X.shape[-1],
overlap_interictal=True,
overlap_preictal=True)
X, scalers = scale_across_time(X, x_test=None)
X_test, _ = scale_across_time(X_test, x_test=None, scalers=scalers)
X,xt,y,yt = split_evenly(X,y,test_size = .25)
#X,xt,y,yt = train_test_split(X,y,test_size = .25)
if enhance_size > 0:
X,y = enhance_data(X,y,enhance_size,cnn=True,even=True)
xt,yt = enhance_data(xt,yt,enhance_size,cnn=True,even=True)
print "train size", X.shape
print "test_size", xt.shape
#print "done loading"
clf.fit(X,y,xt,yt)
#train_loss = np.array([])
#valid_loss = np.array([])
#print "train,valid size",train_loss.shape,valid_loss.shape
#print "done fitting"
preds_proba = clf.predict_proba(X_test)[:,1]
# unsup_size = int(X_test.shape[0]/5)
# top_ind = np.argpartition(preds_proba,-unsup_size)[-unsup_size:]
# bot_ind = preds_proba.argsort()[:unsup_size]
# x_new_p = X_test[top_ind]
# x_new_i = X_test[bot_ind]
# y_p = np.ones(x_new_p.shape[0])
# y_i = np.zeros(x_new_i.shape[0])
#print y_p.shape,y_i.shape
#print x_new_p.shape, x_new_i.shape
# x_new = np.vstack((x_new_p,x_new_i))
# y_new = np.append(y_p,y_i)
# #print x_new.shape,y_new.shape
# #X,xt,y,yt = split_evenly(x_new,y_new,test_size = .25)
# if enhance_size > 0:
# x_new,y_new = enhance_data(x_new,y_new,enhance_size,cnn=True)
# print "train size", X.shape
# print "test_size", xt.shape
# #print "done loading"
# clf2 = CNN(subject)
# clf2.fit(x_new,y_new,xt,yt)
# preds_proba = clf2.predict_proba(X_test)[:,1]
train_loss = np.array([i["train_loss"] for i in clf.convnet.train_history_])
valid_loss = np.array([i["valid_loss"] for i in clf.convnet.train_history_])
#preds_proba = set_median_to_half(preds_proba)[:,1]
preds_scaled = min_max_scale(preds_proba)
#print preds_proba.shape
validation_preds = min_max_scale(clf.predict_proba(xt)[:,1])
return preds_scaled,preds_proba,list(validation_preds),list(yt),train_loss,valid_loss