def train_and_evaluate_MLP(training_data,
test_data,
params,
num_training_runs=100):
X_tr, y_tr = training_data
X_te, y_te = test_data
AUC_trs = []
AUC_tes = []
for i in range(num_training_runs):
e_stop = EarlyStopping(monitor='val_loss',
min_delta=0.01,
patience=5,
mode='min')
callbacks = [e_stop]
optimizer = keras.optimizers.Adam(lr=params['learning_rate'])
model = Sequential()
model.add(
Dense(params['num_neurons'],
input_dim=7,
kernel_initializer='glorot_uniform',
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio'])))
model.add(Dropout(params['dropout_rate']))
model.add(
Dense(2,
kernel_initializer='glorot_uniform',
activation='softmax',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio'])))
model.compile(loss='binary_crossentropy', optimizer=optimizer)
history = model.fit(X_tr,
y_tr,
callbacks=callbacks,
validation_data=(X_te, y_te),
epochs=100,
batch_size=params['batch_size'],
verbose=0)
probs_tr = model.predict(X_tr, batch_size=8)
probs_te = model.predict(X_te, batch_size=8)
score_tr = roc_auc_score(y_tr, probs_tr)
score_te = roc_auc_score(y_te, probs_te)
AUC_trs.append(score_tr)
AUC_tes.append(score_te)
keras.backend.clear_session()
return AUC_trs, AUC_tes
x = concatenate([dense1, dense2])
x = Dense(tune['num_neurons_final'], kernel_initializer = def_params['weight_init'], activation = def_params['hidden_activation'],
kernel_regularizer= keras.regularizers.l2(tune['l2_ratio']))(x)
x= Dropout(tune['dropout_rate'])(x)
if def_params['out_activation'] == 'softmax':
output = Dense(2,kernel_initializer = def_params['weight_init'],activation= def_params['out_activation'],
kernel_regularizer= keras.regularizers.l2(tune['l2_ratio']))(x)
else:
output = Dense(1,kernel_initializer = def_params['weight_init'],activation= def_params['out_activation'],
kernel_regularizer= keras.regularizers.l2(tune['l2_ratio']))(x)
optimizer = keras.optimizers.Adam(lr = tune['learning_rate'])
model = Model(inputs=[img_input, clin_input], outputs=[output])
model.compile(loss=def_params['loss_func'], optimizer = optimizer)
e_stop = EarlyStopping(monitor = 'val_loss', min_delta = def_params['min_delta'], patience = def_params['iter_patience'], mode='auto')
callbacks = [e_stop]
history = model.fit({'image_input' : img_X_tr,'clinical_input' : clin_X_tr}, y_tr, callbacks = callbacks,validation_data= ([img_X_val, clin_X_val],y_val),
epochs=def_params['epochs'], batch_size= tune['batch_size'], verbose=0)
probs_val = model.predict([img_X_val,clin_X_val],batch_size = 8)
score_val = roc_auc_score(y_val, probs_val)
i +=1
if i%10 == 0:
print(i)
if score_val > best_AUC:
best_AUC = score_val
def train_and_evaluate_CNN(training_data,
test_data,
params,
num_training_runs=100):
X_tr, y_tr = training_data
X_te, y_te = test_data
AUC_trs = []
AUC_tes = []
for i in range(num_training_runs):
model = Sequential()
model.add(
Conv3D(params['num_filters'][0],
params['arc_params']['filter_size'],
strides=params['arc_params']['filter_stride'],
padding="same",
kernel_regularizer=keras.regularizers.l2(params['l2_reg']),
input_shape=(156, 192, 64, 1)))
model.add(Activation('relu'))
model.add(MaxPooling3D(pool_size=params['arc_params']['pool_size']))
model.add(
Conv3D(params['num_filters'][1],
params['arc_params']['filter_size'],
strides=params['arc_params']['filter_stride'],
padding="same",
kernel_regularizer=keras.regularizers.l2(params['l2_reg'])))
model.add(Activation('relu'))
model.add(MaxPooling3D(pool_size=params['arc_params']['pool_size']))
model.add(
Conv3D(params['num_filters'][2],
params['arc_params']['filter_size'],
strides=params['arc_params']['filter_stride'],
padding="same",
kernel_regularizer=keras.regularizers.l2(params['l2_reg'])))
model.add(Activation('relu'))
model.add(MaxPooling3D(pool_size=params['arc_params']['pool_size']))
model.add(Flatten())
model.add(
Dense(params['num_neurons_in_powers'] * params['num_filters'][2],
activation='relu',
kernel_regularizer=keras.regularizers.l2(params['l2_reg'])))
model.add(Dropout(params['dropout']))
model.add(
Dense(2,
activation='softmax',
kernel_regularizer=keras.regularizers.l2(params['l2_reg'])))
optimizer = keras.optimizers.Adam(lr=params['learning_rate'])
model.compile(loss='binary_crossentropy', optimizer=optimizer)
parallel_model = multi_gpu_model(model, 2)
parallel_model.compile(loss='binary_crossentropy', optimizer=optimizer)
e_stop = EarlyStopping(monitor='val_loss',
min_delta=0.02,
patience=2,
mode='auto')
callbacks = [e_stop]
start = time.time()
history = parallel_model.fit(X_tr,
y_tr,
callbacks=callbacks,
validation_data=(X_te, y_te),
batch_size=params['batch_size'],
epochs=20,
verbose=0)
end = time.time()
model.set_weights(parallel_model.get_weights())
probs_tr = model.predict(X_tr, batch_size=8)
probs_te = model.predict(X_te, batch_size=8)
score_tr = roc_auc_score(y_tr, probs_tr)
score_te = roc_auc_score(y_te, probs_te)
AUC_trs.append(score_tr)
AUC_tes.append(score_te)
print('Training time for run %i was around %i minutes' %
(i, np.floor((end - start) / 60)))
keras.backend.clear_session()
return AUC_trs, AUC_tes
def train_and_evaluate_feat_extract(img_X_tr,
clin_X_tr,
y_tr,
img_X_te,
clin_X_te,
y_te,
params,
num_training_runs=100):
AUC_trs = []
AUC_tes = []
for i in range(num_training_runs):
img_input = Input(shape=(img_X_tr.shape[1], ), name='image_input')
clin_input = Input(shape=(clin_X_tr.shape[1], ), name='clinical_input')
dense1 = Dense(params['num_neurons_embedding'][0],
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(clin_input)
dense2 = Dense(params['num_neurons_embedding'][1],
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(img_input)
x = concatenate([dense1, dense2])
x = Dense(params['num_neurons_final'],
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x)
x = Dropout(params['dropout_rate'])(x)
output = Dense(2,
activation='softmax',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x)
optimizer = keras.optimizers.Adam(lr=params['learning_rate'])
model = Model(inputs=[img_input, clin_input], outputs=[output])
model.compile(loss='binary_crossentropy', optimizer=optimizer)
e_stop = EarlyStopping(monitor='val_loss',
min_delta=0.01,
patience=5,
mode='auto')
callbacks = [e_stop]
history = model.fit(
{
'image_input': img_X_tr,
'clinical_input': clin_X_tr
},
y_tr,
callbacks=callbacks,
validation_data=([img_X_te, clin_X_te], y_te),
epochs=100,
batch_size=params['batch_size'],
verbose=0)
probs_tr = model.predict([img_X_tr, clin_X_tr], batch_size=8)
probs_te = model.predict([img_X_te, clin_X_te], batch_size=8)
score_tr = roc_auc_score(y_tr, probs_tr)
score_te = roc_auc_score(y_te, probs_te)
AUC_trs.append(score_tr)
AUC_tes.append(score_te)
keras.backend.clear_session()
return AUC_trs, AUC_tes
def train_and_evaluate_end_to_end(img_X_tr,
clin_X_tr,
y_tr,
img_X_te,
clin_X_te,
y_te,
params,
num_training_runs=100):
AUC_trs = []
AUC_tes = []
for i in range(num_training_runs):
img_input = Input(shape=(156, 192, 64, 1), name='image_input')
clin_input = Input(shape=(clin_X_tr.shape[1], ), name='clinical_input')
x1 = Conv3D(params['num_filters'][0], (3, 3, 3),
strides=(1, 1, 1),
padding="same",
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(img_input)
x1 = Activation('relu')(x1)
x1 = MaxPooling3D(pool_size=(3, 3, 3))(x1)
x1 = Conv3D(params['num_filters'][1], (3, 3, 3),
strides=(1, 1, 1),
padding="same",
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x1)
x1 = Activation('relu')(x1)
x1 = MaxPooling3D(pool_size=(3, 3, 3))(x1)
x1 = Conv3D(params['num_filters'][2], (3, 3, 3),
strides=(1, 1, 1),
padding="same",
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x1)
x1 = Activation('relu')(x1)
x1 = MaxPooling3D(pool_size=(3, 3, 3))(x1)
x1 = Flatten()(x1)
x1 = Dense(params['num_filters'][2] * 2,
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x1)
x1 = Dropout(params['dropout_rate'])(x1)
x1 = Dense(params['num_neurons_embedding'][1],
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x1)
x2 = Dense(params['num_neurons_MLP'],
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(clin_input)
x2 = Dropout(params['dropout_rate'])(x2)
x2 = Dense(params['num_neurons_embedding'][0],
activation='relu',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x2)
x = concatenate([x1, x2])
x = Dense(params['num_neurons_final'],
activation='relu',
kernel_regularizer=keras.regularizers.l1(
params['l2_ratio']))(x)
x = Dropout(params['dropout_rate'])(x)
output = Dense(2,
activation='softmax',
kernel_regularizer=keras.regularizers.l2(
params['l2_ratio']))(x)
model = Model(inputs=[img_input, clin_input], outputs=[output])
optimizer = keras.optimizers.Adam(lr=params['learning_rate'])
model.compile(loss='binary_crossentropy', optimizer=optimizer)
e_stop = EarlyStopping(monitor='val_loss',
min_delta=0.02,
patience=2,
mode='auto')
callbacks = [e_stop]
start = time.time()
history = model.fit(
{
'image_input': img_X_tr,
'clinical_input': clin_X_tr
}, #inputs
y_tr, #output
callbacks=callbacks,
validation_data=([img_X_te, clin_X_te], y_te),
epochs=20,
batch_size=params['batch_size'],
verbose=0)
end = time.time()
probs_tr = model.predict([img_X_tr, clin_X_tr], batch_size=8)
probs_te = model.predict([img_X_te, clin_X_te], batch_size=8)
score_tr = roc_auc_score(y_tr, probs_tr)
score_te = roc_auc_score(y_te, probs_te)
AUC_trs.append(score_tr)
AUC_tes.append(score_te)
print('Training time for run %i was around %i minutes' %
(i, np.floor((end - start) / 60)))
keras.backend.clear_session()
return AUC_trs, AUC_tes