def build_model_combine_features(self, load_weight=False):
cnn_branch = Sequential()
cnn_branch.add(
Conv2D(filters=16, kernel_size=5, strides=1, padding='valid', activation='relu', input_shape=(11, 11, 3),
name='Conv1'))
cnn_branch.add(Conv2D(filters=24, kernel_size=3, strides=1, padding='same', activation='relu', name='Conv2'))
cnn_branch.add(Conv2D(filters=32, kernel_size=3, strides=1, padding='same', activation='relu', name='Conv3'))
cnn_branch.add(MaxPooling2D(pool_size=(3, 3), strides=2))
cnn_branch.add(Conv2D(filters=64, kernel_size=3, strides=1, padding='same', activation='relu', name='Conv4'))
cnn_branch.add(MaxPooling2D(pool_size=(3, 3), strides=2))
cnn_branch.add(Conv2D(filters=96, kernel_size=3, strides=1, padding='same', activation='relu', name='Conv5'))
cnn_branch.add(Flatten())
location_branch = Sequential()
location_branch.add(Dense(2, input_shape=(2,), activation='relu'))
model = Concatenate([location_branch, cnn_branch])
model.add(Dense(500, activation='relu'))
model.add(Dense(2, activation='softmax'))
model.compile(optimizer=Adam(lr=self.lr), loss='categorical_crossentropy', metrics=['accuracy'])
if load_weight:
print("Loading weight...")
model.load_weight(WEIGHT_DIR + "")
print("Weight loaded.")
return model
def build_model_combine_features(self, load_weight=False):
cnn_branch = Sequential()
cnn_branch.add(
Conv2D(filters=16,
kernel_size=5,
strides=1,
padding='valid',
activation='relu',
input_shape=(11, 11, 3),
name='Conv1'))
cnn_branch.add(
Conv2D(filters=24,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
name='Conv2'))
cnn_branch.add(
Conv2D(filters=32,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
name='Conv3'))
cnn_branch.add(MaxPooling2D(pool_size=(3, 3), strides=2))
cnn_branch.add(
Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
name='Conv4'))
cnn_branch.add(MaxPooling2D(pool_size=(3, 3), strides=2))
cnn_branch.add(
Conv2D(filters=96,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
name='Conv5'))
cnn_branch.add(Flatten())
location_branch = Sequential()
location_branch.add(Dense(2, input_shape=(2, ), activation='relu'))
model = Concatenate([location_branch, cnn_branch])
model.add(Dense(500, activation='relu'))
model.add(Dense(2, activation='softmax'))
model.compile(optimizer=Adam(lr=self.lr),
loss='categorical_crossentropy',
metrics=['accuracy'])
if load_weight:
print("Loading weight...")
model.load_weight(WEIGHT_DIR + "")
print("Weight loaded.")
return model
def build_base_network(self, X):
input = Input(shape=X.T.shape, name='image_input')
# VGG Layer
vgg = VGG16(weights='imagenet', include_top=False)
# can be removed
vgg.add(Dense(4096))
# Shallow Layers
shallow_1 = Sequential()
shallow_1.add(MaxPooling2D(pool_size=4, strides=8, name='subsample_1'))
shallow_1.add(
Conv2D(96,
kernel_size=8,
strides=4,
activation='relu',
name='conv1'))
shallow_1.add(LRN(name='conv1_norm')) # ??
shallow_1.add(
MaxPooling2D(pool_size=7,
strides=4,
border_mode='valid',
name='pool1'))
shallow_1.add(Flatten())
shallow_2 = Sequential()
shallow_2.add(MaxPooling2D(pool_size=8, strides=8, name='subsample_2'))
shallow_2.add(
Conv2D(96,
kernel_size=8,
strides=4,
activation='relu',
name='conv2'))
shallow_2.add(LRN(name='conv2_norm')) # ??
shallow_2.add(MaxPooling2D(pool_size=3, strides=2, name='pool2'))
shallow_2.add(Flatten())
# concatenated shallow layer
shallow = Concatenate([shallow_1, shallow_2])
# TODO
# combine VGG16 and shallow layer output
shallow.add(Dense(4096))
shallow.add(LRN(name='shallow_norm', alpha=8191, n=8191, beta=0.5))
return model
def Discriminator(d): dx = Discriminator_x(d) dz = Discriminator_z() combined_tensor = Concatenate([dx, dz]) combined_tensor.add(Dropout(0.2)) combined_tensor.add(LeakyReLU(alpha = 0.1)) combined_tensor.add(Dense(units = 1, activation = 'linear')) return combined_tensor
import numpy as np
x = np.array([1,2,3])
y = np.array([1,2,3])
first = Sequential()
first.add(Dense(1, input_shape=(1,), activation='sigmoid'))
second = Sequential()
second.add(Dense(1, input_shape=(1,), activation='sigmoid'))
# result = Sequential()
merged = Concatenate([first, second])
# ada_grad = Adagrad(lr=0.1, epsilon=1e-08, decay=0.0)
# result.add(merged)
# result.compile(optimizer=ada_grad, loss=_loss_tensor, metrics=['accuracy'])
# merged.summary()
merged.add(Dense(2, activation="relu"))
merged.add(Dense(1, activation="linear"))
merged.summary()
# result.compile(optimizer='adam', loss='mse', metrics=['accuracy'])
# result.fit([x, x],[y, y], epochs=100, batch_size=1)
# loss, acc = result.evaluate([x, x],[y, y], batch_size=1)
# print("acc : ", acc)