def create_new_model(input_shape):
# facenet_model = load_model(modelFile)
inputs = Input(shape=input_shape)
#i = facenet_model(inputs)
#print(facenet_model.get_input_at(0).shape)
model1 = Model(input=facenet_model.get_input_at(0),
output=facenet_model.get_layer('add_15').output)
for layer in model1.layers:
layer.trainable = False
inter_shape = model1.output_shape[1:]
print(inter_shape)
#model2 = Model(input=model1.input1, outputs = inter_layer);
#inter_shape =
print(inter_shape)
dense = DenseNet(input_shape=inter_shape,
nb_filter=600,
include_top=False,
depth=1,
nb_dense_block=2,
growth_rate=32)(model1.output)
#d.layers.pop()
#d.layers.pop()
#d.layers.pop()
#d = d.layers.pop()
#dense.summary()
#model = d
#model.summary()
x = AveragePooling2D(pool_size=(4, 4), name='cutome_avg_pool_1')(dense)
x = Flatten(name='cutome_flatten_2')(x)
x = Dense(512, activation='relu', name='custome_dense_dense_3')(x)
x = BatchNormalization(name='cutome_batch_normalization_4')(x)
x = Dense(20, activation='relu', name='custome_dense_5')(x)
#x = Dense(1, activation='relu', name = 'custome_dense_output_1')(x)
x = Lambda(lambda x: K.l2_normalize(x, axis=1),
name='custome_l2_normalize_6')(x)
#fin = d(inter_layer.output)
model = Model(inputs=[model1.input], outputs=x)
input_1 = Input(shape=input_shape)
input_2 = Input(shape=input_shape)
input_3 = Input(shape=input_shape)
l1 = model(input_1)
l2 = model(input_2)
l3 = model(input_3)
x = concatenate([l1, l2, l3])
print(x.shape)
final_model = Model(inputs=[input_1, input_2, input_3], outputs=x)
final_model.compile(optimizer=optimizer.Adam(lr=0.0005),
loss=loss_fun,
metrics=[accuracy_c])
return final_model
def _build_model(self, input_shape):
#blks = [self.config['block_0'], self.config['block_1'], self.config['block_2'], self.config['block_3']]
#print('--------BLOCK:',blks)
#model = densenet.DenseNet(blocks=[6,6,6,6], input_shape=(32,32,3), include_top=True, weights=None, input_tensor=None, pooling=None, classes=10, backend=keras.backend, layers=keras.layers, models=keras.models)#, keras_utils= keras.utils )
depth = 40
nb_dense_block = 3
growth_rate = self.config['growth_rate']
nb_filter = 16 #TODO
dropout_rate = self.config['dropout']
weight_decay = self.config['weight_decay']
model = DenseNet(depth=depth,
nb_dense_block=nb_dense_block,
growth_rate=growth_rate,
nb_filter=nb_filter,
dropout_rate=dropout_rate,
input_shape=(32, 32, 3),
weights=None)
#model = densenet.DenseNet(blocks=[12,24,12,24], input_shape=(32,32,3), include_top=True, weights=None, input_tensor=None, pooling=None, classes=10, backend=keras.backend, layers=keras.layers, models=keras.models)#, keras_utils= keras.utils )
model.summary()
return model
def create_model(X_train, y_train, X_val, y_val, X_test, y_test):
epochs = {{choice([25, 30, 35])}}
es_patience = 5
lr_patience = 3
dropout = {{uniform(0.2, 0.8)}}
depth = 25
nb_dense_block = 3
nb_filter = 16
growth_rate = 18
bn = True
reduction_ = 0.5
bs = 32
weight_file = 'hyperas_dn_lr_optimizer_wt_5Oct_2210.h5'
nb_classes = 1
img_dim = (2, 96, 96)
n_channels = 2
model = DenseNet(depth=depth,
nb_dense_block=nb_dense_block,
growth_rate=growth_rate,
nb_filter=nb_filter,
dropout_rate=dropout,
activation='sigmoid',
input_shape=img_dim,
include_top=True,
bottleneck=bn,
reduction=reduction_,
classes=nb_classes,
pooling='avg',
weights=None)
model.summary()
model.compile(loss=binary_crossentropy,
optimizer=Nadam(),
metrics=['accuracy'])
es = EarlyStopping(monitor='val_loss', patience=es_patience, verbose=1)
checkpointer = ModelCheckpoint(filepath=weight_file,
verbose=1,
save_best_only=True)
lr_reducer = ReduceLROnPlateau(monitor='val_loss',
factor=np.sqrt(0.1),
cooldown=0,
patience=lr_patience,
min_lr=0.5e-6,
verbose=1)
model.fit(X_train,
y_train,
batch_size=bs,
epochs=epochs,
callbacks=[checkpointer, es],
validation_data=(X_val, y_val),
verbose=2)
score, acc = model.evaluate(X_val, y_val)
print("current val accuracy:%0.3f" % acc)
pred = model.predict(X_val)
auc_score = roc_auc_score(y_val, pred)
print("current auc_score ------------------> %0.3f" % auc_score)
model = load_model(weight_file) #This is the best model
score, acc = model.evaluate(X_val, y_val)
print("Best saved model val accuracy:%0.3f" % acc)
pred = model.predict(X_val)
auc_score = roc_auc_score(y_val, pred)
print("best saved model auc_score ------------------> %0.3f" % auc_score)
return {'loss': -auc_score, 'status': STATUS_OK, 'model': model}
# Parameters for the DenseNet model builder
img_dim = (img_channels, img_rows,
img_cols) if K.image_data_format() == 'channels_first' else (
img_rows, img_cols, img_channels)
depth = 40
nb_dense_block = 3
growth_rate = 12
nb_filter = 16
dropout_rate = 0.0 # 0.0 for data augmentation
# Create the model (without loading weights)
model = DenseNet(depth,
nb_dense_block,
growth_rate,
nb_filter,
dropout_rate=dropout_rate,
input_shape=img_dim,
weights=None)
print('Model created')
model.summary()
optimizer = Adam(lr=1e-3) # Using Adam instead of SGD to speed up training
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['acc'])
print('Finished compiling')
(trainX, trainY), (testX, testY) = cifar10.load_data()
def fit_model(X_train, y_train, X_val, y_val):
epochs = 30
#input_shape = (1,96,96)
es_patience = 5
lr_patience = 5
#dense_filter = 512
#dropout = 0.76
dropout1 = None
depth = 13 #40
nb_dense_block = 3
nb_filter = 18
growth_rate = 12
weight_decay = 1E-4
lr = 3E-4
weight_file = 'keras_densenet_simple_wt_28Sept.h5'
nb_classes = 1
img_dim = (2, 96, 96)
n_channels = 2
model = DenseNet(depth=depth,
nb_dense_block=nb_dense_block,
growth_rate=growth_rate,
nb_filter=nb_filter,
dropout_rate=dropout1,
activation='sigmoid',
input_shape=img_dim,
include_top=True,
bottleneck=True,
reduction=0.5,
classes=nb_classes,
pooling='avg',
weights=None)
model.summary()
opt = Adam(lr=lr)
model.compile(loss=binary_crossentropy,
optimizer=opt,
metrics=['accuracy'])
es = EarlyStopping(monitor='val_loss', patience=es_patience, verbose=1)
#es = EarlyStopping(monitor='val_acc', patience=es_patience,verbose=1,restore_best_weights=True)
checkpointer = ModelCheckpoint(filepath=weight_file,
verbose=1,
save_best_only=True)
lr_reducer = ReduceLROnPlateau(monitor='val_loss',
factor=np.sqrt(0.1),
cooldown=0,
patience=lr_patience,
min_lr=0.5e-6,
verbose=1)
model.fit(X_train,
y_train,
batch_size=64,
epochs=epochs,
callbacks=[es, lr_reducer, checkpointer],
validation_data=(X_val, y_val),
verbose=2)
score, acc = model.evaluate(X_val, y_val)
print('current Test accuracy:', acc)
pred = model.predict(X_val)
auc_score = roc_auc_score(y_val, pred)
print("current auc_score ------------------> ", auc_score)
model = load_model(weight_file) #This is the best model
score, acc = model.evaluate(X_val, y_val)
print('Best saved model Test accuracy:', acc)
pred = model.predict(X_val)
auc_score = roc_auc_score(y_val, pred)
print("best saved model auc_score ------------------> ", auc_score)
threshold = 0.6
pred_scores2 = (pred > threshold).astype(int)
test_acc2 = accuracy_score(y_val, pred_scores2)
print('Test accuracy 0.6:', test_acc2)
return auc_score, model
def create_model():
epochs = 4
es_patience = 7
lr_patience = 5
dropout = None
#depth = {{choice([7,13,19,25,31])}}
#nb_dense_block = {{choice([2,3])}}
depth = 7
nb_dense_block = 2
nb_filter = 16
#growth_rate = {{choice([6,10,14,18])}}
growth_rate = 10
bn = True
reduction_ = 0.5
bs = 32
lr = 1E-3 #########################################################CHange file name##########################################
nb_classes = 1
img_dim = (2, 96, 96)
n_channels = 2
model = DenseNet(depth=depth,
nb_dense_block=nb_dense_block,
growth_rate=growth_rate,
nb_filter=nb_filter,
dropout_rate=dropout,
activation='sigmoid',
input_shape=img_dim,
include_top=True,
bottleneck=bn,
reduction=reduction_,
classes=nb_classes,
pooling='avg',
weights=None)
model.summary()
opt = Adam(lr=lr)
model.compile(loss=binary_crossentropy,
optimizer=opt,
metrics=['accuracy'])
"""es = EarlyStopping(monitor='val_loss', patience=es_patience,verbose=1)
#es = EarlyStopping(monitor='val_acc', patience=es_patience,verbose=1,restore_best_weights=True)
checkpointer = ModelCheckpoint(filepath=weight_file,verbose=1, save_best_only=True)
lr_reducer = ReduceLROnPlateau(monitor='val_loss', factor=np.sqrt(0.1), cooldown=0, patience=lr_patience, min_lr=0.5e-6,verbose=1)
model.fit(X_train,y_train,
batch_size=bs,
epochs=epochs,
callbacks=[es,lr_reducer,checkpointer],
validation_data=(X_val,y_val),
verbose=2)
score, acc = model.evaluate(X_test, y_test)
print('current Test accuracy:', acc)
pred = model.predict(X_test)
auc_score = roc_auc_score(y_test,pred)
print("current auc_score ------------------> ",auc_score)"""
"""model = load_model(weight_file) #This is the best model
score, acc = model.evaluate(X_test, y_test)
print('Best saved model Test accuracy:', acc)
pred = model.predict(X_test)
auc_score = roc_auc_score(y_test,pred)
print("best saved model auc_score ------------------> ",auc_score)"""
return model
def create_model(X_train, y_train, X_val, y_val, X_test, y_test):
epochs = 1
es_patience = 5
lr_patience = 3
dropout = None
depth = {{choice([7, 16, 25, 34])}}
nb_dense_block = {{choice([2, 3, 4])}}
nb_filter = 16
growth_rate = {{choice([6, 14, 22, 30])}}
bn = True
reduction_ = 0.5
bs = 32
lr = {{choice([1E-2, 5E-2, 1E-3, 5E-3, 1E-4, 5E-4, 1E-5, 5E-5])}}
weight_file = 'keras_densenet_simple_wt_02Oct_1300.h5'
nb_classes = 1
img_dim = (2, 96, 96)
n_channels = 2
print("Depth: ", depth, " Growth_rate: ", growth_rate, " Nb_dense_block: ",
nb_dense_block, " lr: ", lr)
model = DenseNet(depth=depth,
nb_dense_block=nb_dense_block,
growth_rate=growth_rate,
nb_filter=nb_filter,
dropout_rate=dropout,
activation='sigmoid',
input_shape=img_dim,
include_top=True,
bottleneck=bn,
reduction=reduction_,
classes=nb_classes,
pooling='avg',
weights=None)
model.summary()
opt = Adam(lr=lr)
model.compile(loss=binary_crossentropy,
optimizer=opt,
metrics=['accuracy'])
es = EarlyStopping(monitor='val_loss', patience=es_patience, verbose=1)
checkpointer = ModelCheckpoint(filepath=weight_file,
verbose=1,
save_best_only=True)
lr_reducer = ReduceLROnPlateau(monitor='val_loss',
factor=np.sqrt(0.1),
cooldown=0,
patience=lr_patience,
min_lr=0.5e-6,
verbose=1)
model.fit(X_train,
y_train,
batch_size=bs,
epochs=epochs,
callbacks=[lr_reducer, es],
validation_data=(X_val, y_val),
verbose=2)
score, acc = model.evaluate(X_val, y_val)
print('current val accuracy:', acc)
pred = model.predict(X_val)
auc_score = roc_auc_score(y_val, pred)
print("current auc_score ------------------> ", auc_score)
#model = load_model(weight_file) #This is the best model
#score, acc = model.evaluate(X_val,y_val)
#print('Best saved model val accuracy:', acc)
#pred = model.predict(X_val)
#auc_score = roc_auc_score(y_val,pred)
#print("best saved model auc_score ------------------> ",auc_score)
return {'loss': -auc_score, 'status': STATUS_OK, 'model': model}
def fit_model(X_train, y_train, X_val, y_val, G):
epochs = 5
es_patience = 5
lr_patience = 3
dropout1 = None
depth = 25 #40
nb_dense_block = 3
nb_filter = 18
growth_rate = 18
lr = 3E-1
weight_file = 'keras_densenet_simple_wt_30Sept.h5'
bn = True
reduction_ = 0.5
nb_classes = 1
img_dim = (2, 96, 96)
n_channels = 2
model = DenseNet(depth=depth,
nb_dense_block=nb_dense_block,
growth_rate=growth_rate,
nb_filter=nb_filter,
dropout_rate=dropout1,
activation='sigmoid',
input_shape=img_dim,
include_top=True,
bottleneck=bn,
reduction=reduction_,
classes=nb_classes,
pooling='avg',
weights=None)
model.summary()
opt = Adam(lr=lr)
parallel_model = multi_gpu_model(model, gpus=G)
parallel_model.compile(loss=binary_crossentropy,
optimizer=Adadelta(),
metrics=['accuracy'])
es = EarlyStopping(monitor='val_loss', patience=es_patience, verbose=1)
#es = EarlyStopping(monitor='val_acc', patience=es_patience,verbose=1,restore_best_weights=True)
checkpointer = ModelCheckpoint(filepath=weight_file,
verbose=1,
save_best_only=True)
lr_reducer = ReduceLROnPlateau(monitor='val_loss',
factor=np.sqrt(0.1),
cooldown=0,
patience=lr_patience,
min_lr=0.5e-6,
verbose=1)
parallel_model.fit(X_train,
y_train,
batch_size=64 * G,
epochs=epochs,
callbacks=[es, lr_reducer, checkpointer],
validation_data=(X_val, y_val),
verbose=2)
score, acc = parallel_model.evaluate(X_val, y_val)
print('current Test accuracy:', acc)
pred = parallel_model.predict(X_val)
auc_score = roc_auc_score(y_val, pred)
print("current auc_score ------------------> ", auc_score)
"""model = load_model(weight_file) #This is the best model
score, acc = model.evaluate(X_val, y_val)
print('Best saved model Test accuracy:', acc)
pred = model.predict(X_val)
auc_score = roc_auc_score(y_val,pred)
print("best saved model auc_score ------------------> ",auc_score)"""
return auc_score, parallel_model
