def predict_with_existing(self,
X_train, y_train_regression, y_train_classification,
X_val, y_val_regression, y_val_classification,
X_test, y_test_regression, y_test_classification,
PMTNN_weight_file):
model = self.setup_model_ensemble()
model.load_weights(PMTNN_weight_file)
y_pred_on_train = model.predict(X_train)
y_pred_on_val = model.predict(X_val)
y_pred_on_test = model.predict(X_test)
print
print('train precision: {}'.format(precision_auc_single(y_train_classification, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train_classification, y_pred_on_train)))
print('train bedroc: {}'.format(bedroc_auc_single(y_train_classification, y_pred_on_train)))
print
print('validation precision: {}'.format(precision_auc_single(y_val_classification, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val_classification, y_pred_on_val)))
print('validation bedroc: {}'.format(bedroc_auc_single(y_val_classification, y_pred_on_val)))
print
print('test precision: {}'.format(precision_auc_single(y_test_classification, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test_classification, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test_classification, y_pred_on_test)))
print
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(y_test_classification, y_pred_on_test, EF_ratio)
print('ratio: {}, EF: {},\tactive: {}'.format(EF_ratio, ef, n_actives))
return
def predict_with_existing(self, X_train, y_train, X_val, y_val, X_test,
y_test, PMTNN_weight_file):
model = self.setup_model()
model.load_weights(PMTNN_weight_file)
y_pred_on_train = model.predict(X_train)[:, -1]
y_train = y_train[:, -1]
y_pred_on_val = model.predict(X_val)[:, -1]
y_val = y_val[:, -1]
if X_test is not None:
y_pred_on_test = model.predict(X_test)[:, -1]
y_test = y_test[:, -1]
print y_train.shape, '\t', y_pred_on_test.shape
print
print('train precision: {}'.format(
precision_auc_single(y_train, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train, y_pred_on_train)))
print('train bedroc: {}'.format(
bedroc_auc_single(y_train, y_pred_on_train)))
print
print('validation precision: {}'.format(
precision_auc_single(y_val, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val,
y_pred_on_val)))
print('validation bedroc: {}'.format(
bedroc_auc_single(y_val, y_pred_on_val)))
print
if X_test is not None:
print('test precision: {}'.format(
precision_auc_single(y_test, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test,
y_pred_on_test)))
print('test bedroc: {}'.format(
bedroc_auc_single(y_test, y_pred_on_test)))
print
if X_test is not None:
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(
y_test, y_pred_on_test, EF_ratio)
nef = ef / ef_max
print('ratio: {}, EF: {},\tactive: {}'.format(
EF_ratio, ef, n_actives))
print('ratio: {}, NEF: {}'.format(EF_ratio, nef))
return
def train_and_predict(self,
X_train, y_train_regression, y_train_classification,
X_val, y_val_regression, y_val_classification,
X_test, y_test_regression, y_test_classification,
PMTNN_weight_file):
model = self.setup_model()
sw = get_sample_weight(self, y_train_regression)
print 'Sample Weight\t', sw
model.compile(loss=self.compile_loss, optimizer=self.compile_optimizer)
model.fit(x=X_train, y=y_train_regression,
nb_epoch=self.fit_nb_epoch,
batch_size=self.fit_batch_size,
verbose=self.fit_verbose,
sample_weight=sw,
validation_data=[X_val, y_val_regression],
shuffle=True)
model.save_weights(PMTNN_weight_file)
y_pred_on_train = model.predict(X_train)
y_pred_on_val = model.predict(X_val)
if X_test is not None:
y_pred_on_test = model.predict(X_test)
print
print('train precision: {}'.format(precision_auc_single(y_train_classification, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train_classification, y_pred_on_train)))
print('train bedroc: {}'.format(bedroc_auc_single(y_train_classification, y_pred_on_train)))
print
print('validation precision: {}'.format(precision_auc_single(y_val_classification, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val_classification, y_pred_on_val)))
print('validation bedroc: {}'.format(bedroc_auc_single(y_val_classification, y_pred_on_val)))
print
if X_test is not None:
print('test precision: {}'.format(precision_auc_single(y_test_classification, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test_classification, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test_classification, y_pred_on_test)))
print
if X_test is not None:
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(y_test_classification, y_pred_on_test, EF_ratio)
nef = ef / ef_max
print('ratio: {}, EF: {},\tactive: {}'.format(EF_ratio, ef, n_actives))
print('ratio: {}, NEF: {}'.format(EF_ratio, nef))
return
def get_rf(self, X_train, y_train, X_val, y_val, X_test, y_test):
max_features = 'log2'
n_estimators = 4000
min_samples_leaf = 1
class_weight = 'balanced'
rnd_state = 1337
np.random.seed(seed=rnd_state)
rf = RandomForestClassifier(n_estimators=n_estimators,
max_features=max_features,
min_samples_leaf=min_samples_leaf,
n_jobs=3,
class_weight=class_weight,
random_state=rnd_state,
oob_score=False,
verbose=1)
rf.fit(X_train, y_train)
y_pred_on_train = reshape_data_into_2_dim(rf.predict(X_train))
y_pred_on_val = reshape_data_into_2_dim(rf.predict(X_val))
y_pred_on_test = reshape_data_into_2_dim(rf.predict(X_test))
print('train precision: {}'.format(precision_auc_single(y_train, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train, y_pred_on_train)))
print('train bedroc: {}'.format(bedroc_auc_single(y_train, y_pred_on_train)))
print
print('validation precision: {}'.format(precision_auc_single(y_val, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val, y_pred_on_val)))
print('validation bedroc: {}'.format(bedroc_auc_single(y_val, y_pred_on_val)))
print
print('test precision: {}'.format(precision_auc_single(y_test, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test, y_pred_on_test)))
print
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(y_test, y_pred_on_test, EF_ratio)
print('ratio: {}, EF: {},\tactive: {}'.format(EF_ratio, ef, n_actives))
return rf
def get_EF_score_with_existing_model(self,
X_test, y_test, y_test_classification,
file_path, EF_ratio):
model = self.setup_model_ensemble()
model.load_weights(file_path)
y_pred_on_test = model.predict(X_test)
print('test precision: {}'.format(precision_auc_single(y_test_classification, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test_classification, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test_classification, y_pred_on_test)))
print
n_actives, ef, ef_max = enrichment_factor_single(y_test_classification, y_pred_on_test, EF_ratio)
print('EF: {},\tactive: {}'.format(ef, n_actives))
return
def get_EF_values_single_task(task, X_test, y_test, model_weight,
EF_ratio_list):
model = task.setup_model()
model.load_weights(model_weight)
y_pred_on_test = model.predict(X_test)
print('test precision: {}'.format(
precision_auc_single(y_test, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test, y_pred_on_test)))
print
ef_values = []
ef_max_values = []
for EF_ratio in EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(
y_test, y_pred_on_test, EF_ratio)
ef_values.append(ef)
ef_max_values.append(ef_max)
return ef_values, ef_max_values
def train_and_predict(self,
X_train, y_train, y_train_classification,
X_val, y_val, y_val_classification,
X_test, y_test, y_test_classification,
mode):
model = Sequential()
conf = self.conf
batch_normalizer_epsilon = conf['batch']['epsilon']
batch_normalizer_mode = conf['batch']['mode']
batch_normalizer_axis = conf['batch']['axis']
batch_normalizer_momentum = conf['batch']['momentum']
batch_normalizer_weights = conf['batch']['weights']
batch_normalizer_beta_init = conf['batch']['beta_init']
batch_normalizer_gamma_init = conf['batch']['gamma_init']
batch_normalizer = BatchNormalization(epsilon=batch_normalizer_epsilon,
mode=batch_normalizer_mode,
axis=batch_normalizer_axis,
momentum=batch_normalizer_momentum,
weights=batch_normalizer_weights,
beta_init=batch_normalizer_beta_init,
gamma_init=batch_normalizer_gamma_init)
if mode == 'classification':
model.add(Dense(2048, input_dim=1024, init='glorot_normal', activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1024, init='glorot_normal', activation='relu'))
model.add(Dropout(0.5))
if self.batch_is_use:
model.add(batch_normalizer)
model.add(Dense(1, init='glorot_normal', activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='adam')
else:
model.add(Dense(2048, input_dim=1024, init='glorot_normal', activation='sigmoid'))
model.add(Dropout(0.5))
model.add(Dense(1024, init='glorot_normal', activation='sigmoid'))
model.add(Dropout(0.5))
if self.batch_is_use:
model.add(batch_normalizer)
model.add(Dense(1, init='glorot_normal', activation='linear'))
model.compile(loss='mse', optimizer='adam')
model.fit(X_train, y_train,
batch_size=self.fit_batch_size,
nb_epoch=self.fit_nb_epoch,
verbose=self.fit_verbose,
validation_data=(X_val, y_val))
y_pred_on_train = model.predict(X_train)
y_pred_on_val = model.predict(X_val)
y_pred_on_test = model.predict(X_test)
print
print 'this is mode ', mode
print('train precision: {}'.format(precision_auc_single(y_train_classification, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train_classification, y_pred_on_train)))
print('train bedroc: {}'.format(bedroc_auc_single(y_train_classification, y_pred_on_train)))
print
print('validation precision: {}'.format(precision_auc_single(y_val_classification, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val_classification, y_pred_on_val)))
print('validation bedroc: {}'.format(bedroc_auc_single(y_val_classification, y_pred_on_val)))
print
print('test precision: {}'.format(precision_auc_single(y_test_classification, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test_classification, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test_classification, y_pred_on_test)))
print
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(y_test_classification, y_pred_on_test, EF_ratio)
print('ratio: {}, EF: {},\tactive: {}'.format(EF_ratio, ef, n_actives))
return y_pred_on_test
def train_and_predict_ensemble(self,
X_train, y_train_regression, y_train_classification,
X_val, y_val_regression, y_val_classification,
X_test, y_test_regression, y_test_classification,
PMTNN_weight_file):
model = self.setup_model_ensemble()
# TODO: remove
print model.summary()
if self.weight_schema == 'weighted':
loss_weight = {'classification_output_layer': 1., 'regression_output_layer': 100.}
elif self.weight_schema == 'no_weight':
loss_weight = {'classification_output_layer': 1., 'regression_output_layer': 1.}
else:
raise ValueError('Wrong weight schema. Should be no_weight, or weighted.')
model.compile(optimizer=self.compile_optimizer,
loss={'classification_output_layer': 'binary_crossentropy', 'regression_output_layer': 'mse'},
loss_weights=loss_weight)
model.fit({'input_layer': X_train},
{'classification_output_layer': y_train_classification,
'regression_output_layer': y_train_regression},
nb_epoch=self.fit_nb_epoch,
batch_size=self.fit_batch_size,
verbose=self.fit_verbose,
validation_data=({'input_layer': X_val},
{'classification_output_layer': y_val_classification,
'regression_output_layer': y_val_regression}),
shuffle=True)
model.save_weights(PMTNN_weight_file)
y_pred_on_train_ensemble = np.array(model.predict(X_train))
y_pred_on_val_ensmble = np.array(model.predict(X_val))
y_pred_on_test_ensemble = np.array(model.predict(X_test))
print
print 'TreeNet Ensemble'
mode_list = ['TreeNet classification', 'TreeNet regression']
for mode in range(2):
print
print mode_list[mode]
y_pred_on_train = y_pred_on_train_ensemble[mode]
y_pred_on_val = y_pred_on_val_ensmble[mode]
y_pred_on_test = y_pred_on_test_ensemble[mode]
print('train precision: {}'.format(precision_auc_single(y_train_classification, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train_classification, y_pred_on_train)))
print('train bedroc: {}'.format(bedroc_auc_single(y_train_classification, y_pred_on_train)))
print
print('validation precision: {}'.format(precision_auc_single(y_val_classification, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val_classification, y_pred_on_val)))
print('validation bedroc: {}'.format(bedroc_auc_single(y_val_classification, y_pred_on_val)))
print
print('test precision: {}'.format(precision_auc_single(y_test_classification, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test_classification, y_pred_on_test)))
print('test bedroc: {}'.format(bedroc_auc_single(y_test_classification, y_pred_on_test)))
print
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(y_test_classification, y_pred_on_test, EF_ratio)
print('ratio: {}, EF: {},\tactive: {}'.format(EF_ratio, ef, n_actives))
return y_pred_on_test_ensemble
def train_and_predict(self, X_train, y_train, X_val, y_val, X_test, y_test,
PMTNN_weight_file):
model = self.setup_model()
if self.early_stopping_option == 'auc':
early_stopping = KeckCallBackOnROC(
X_train,
y_train,
X_val,
y_val,
patience=self.early_stopping_patience,
file_path=PMTNN_weight_file)
callbacks = [early_stopping]
elif self.early_stopping_option == 'precision':
early_stopping = KeckCallBackOnPrecision(
X_train,
y_train,
X_val,
y_val,
patience=self.early_stopping_patience,
file_path=PMTNN_weight_file)
callbacks = [early_stopping]
else:
callbacks = []
cw = get_class_weight(self, y_train)
print 'cw ', cw
model.compile(loss=self.compile_loss, optimizer=self.compile_optimizer)
model.fit(X_train,
y_train,
nb_epoch=self.fit_nb_epoch,
batch_size=self.fit_batch_size,
verbose=self.fit_verbose,
class_weight=cw,
shuffle=True,
callbacks=callbacks)
if self.early_stopping_option == 'auc' or self.early_stopping_option == 'precision':
model = early_stopping.get_best_model()
y_pred_on_train = model.predict(X_train)
y_pred_on_val = model.predict(X_val)
if X_test is not None:
y_pred_on_test = model.predict(X_test)
print
print('train precision: {}'.format(
precision_auc_single(y_train, y_pred_on_train)))
print('train roc: {}'.format(roc_auc_single(y_train, y_pred_on_train)))
print('train bedroc: {}'.format(
bedroc_auc_single(y_train, y_pred_on_train)))
print
print('validation precision: {}'.format(
precision_auc_single(y_val, y_pred_on_val)))
print('validation roc: {}'.format(roc_auc_single(y_val,
y_pred_on_val)))
print('validation bedroc: {}'.format(
bedroc_auc_single(y_val, y_pred_on_val)))
print
if X_test is not None:
print('test precision: {}'.format(
precision_auc_single(y_test, y_pred_on_test)))
print('test roc: {}'.format(roc_auc_single(y_test,
y_pred_on_test)))
print('test bedroc: {}'.format(
bedroc_auc_single(y_test, y_pred_on_test)))
print
if X_test is not None:
for EF_ratio in self.EF_ratio_list:
n_actives, ef, ef_max = enrichment_factor_single(
y_test, y_pred_on_test, EF_ratio)
nef = ef / ef_max
print('ratio: {}, EF: {},\tactive: {}'.format(
EF_ratio, ef, n_actives))
print('ratio: {}, NEF: {}'.format(EF_ratio, nef))
return