def compile(input_shape):
X_input = Input(input_shape)
X = Conv2D(64,(7,7),strides=(1,1),padding='valid')(X_input)
X = BatchNormalization()(X)
X1 = Conv2D(64,(7,7),strides=(1,1),padding='valid')(X_input)
X1 = BatchNormalization()(X1)
X = layers.Maximum()([X,X1])
X = Conv2D(64,(4,4),strides=(1,1),padding='valid',activation='relu')(X)
X2 = Conv2D(160,(13,13),strides=(1,1),padding='valid')(X_input)
X2 = BatchNormalization()(X2)
X21 = Conv2D(160,(13,13),strides=(1,1),padding='valid')(X_input)
X21 = BatchNormalization()(X21)
X2 = layers.Maximum()([X2,X21])
X3 = Conv2D(64,(3,3),strides=(1,1),padding='valid')(X)
X3 = BatchNormalization()(X3)
X31 = Conv2D(64,(3,3),strides=(1,1),padding='valid')(X)
X31 = BatchNormalization()(X31)
X = layers.Maximum()([X3,X31])
X = Conv2D(64,(2,2),strides=(1,1),padding='valid',activation='relu')(X)
X = Concatenate()([X2,X])
X = Conv2D(5,(21,21),strides=(1,1),padding='valid')(X)
X = Activation('softmax')(X)
model = Model(inputs = X_input, outputs = X)
model.compile(optimizer='adam', loss='categorical_crossentropy')
return model
def create_maxout_model(input_shape):
"""
Create a maxout model with three convolutional maxout layers and one dense
maxout layer.
Args:
input_shape (tuple):
shape of the images to run on; i.e. (rows, cols, channels)
Returns:
the compiled keras model, ready to be trained.
"""
inputs = layers.Input(shape = input_shape, name='input')
x = layers.Dropout(0.2, name='dropout_1')(inputs)
# First maxout layer
x = layers.Maximum(name='maxout_1')([
layers.Conv2D(96, (8,8), activation='relu', padding='same', name='conv_1_{}'.format(i))(x)
for i in range(2)
])
x = layers.MaxPool2D(name='maxpool_1')(x)
x = layers.Dropout(0.2, name='dropout_2')(x)
# Second maxout layer
x = layers.Maximum(name='maxout_2')([
layers.Conv2D(192, (8,8), activation='relu', padding='same', name='conv_2_{}'.format(i))(x)
for i in range(2)
])
x = layers.MaxPool2D(name='maxpool_2')(x)
x = layers.Dropout(0.2, name='dropout_3')(x)
# Third maxout layer
x = layers.Maximum(name='maxout_3')([
layers.Conv2D(192, (5,5), activation='relu', padding='same', name='conv_3_{}'.format(i))(x)
for i in range(2)
])
x = layers.MaxPool2D(name='maxpool_3')(x)
x = layers.Flatten(name='flatten')(x)
x = layers.Dropout(0.2, name='dropout_4')(x)
# Dense maxout layer
x = layers.Maximum(name='maxout_5')([
layers.Dense(500, activation='relu', name='dense_1_{}'.format(i))(x)
for i in range(5)
])
x = layers.Dropout(0.2, name='dropout_5')(x)
predictions = layers.Dense(10, activation='softmax', name='dense_2')(x)
model = Model(inputs = inputs, outputs = predictions)
model.compile(loss='categorical_crossentropy',
optimizer = optimizers.adadelta(),
metrics=['accuracy'])
return model
def two_path(x_input):
x = Conv2D(64,(7,7),strides=(1,1),padding='valid')(x_input)
x = BatchNormalization()(x)
x1 = Conv2D(64,(7,7),strides=(1,1),padding='valid')(x_input)
x1 = BatchNormalization()(x1)
x = layers.Maximum()([x,x1])
x = Conv2D(64,(4,4),strides=(1,1),padding='valid',activation='relu')(x)
x2 = Conv2D(160,(13,13),strides=(1,1),padding='valid')(x_input)
x2 = BatchNormalization()(x2)
x21 = Conv2D(160,(13,13),strides=(1,1),padding='valid')(x_input)
x21 = BatchNormalization()(x21)
x2 = layers.Maximum()([x2,x21])
x3 = Conv2D(64,(3,3),strides=(1,1),padding='valid')(x)
x3 = BatchNormalization()(x3)
x31 = Conv2D(64,(3,3),strides=(1,1),padding='valid')(x)
x31 = BatchNormalization()(x31)
x = layers.Maximum()([x3,x31])
x = Conv2D(64,(2,2),strides=(1,1),padding='valid',activation='relu')(x)
x = Concatenate()([x2,x])
return x
def test_merge_maximum():
i1 = layers.Input(shape=(4, 5))
i2 = layers.Input(shape=(4, 5))
o = layers.maximum([i1, i2])
assert o._keras_shape == (None, 4, 5)
model = models.Model([i1, i2], o)
max_layer = layers.Maximum()
o2 = max_layer([i1, i2])
assert max_layer.output_shape == (None, 4, 5)
x1 = np.random.random((2, 4, 5))
x2 = np.random.random((2, 4, 5))
out = model.predict([x1, x2])
assert out.shape == (2, 4, 5)
assert_allclose(out, np.maximum(x1, x2), atol=1e-4)
def squeeze_excite_block(input, ratio=0.25):
cse = channel_squeeze_excite_block(input, ratio)
sse = spatial_squeeze_excite_block(input)
output = layers.Maximum()([cse, sse])
return output
def three_path(X_input):
"""
Three pathway block.
:param X_input: Input to the block
:return: Output of the block
"""
# Small Path
X0 = Conv2D(64, (5, 5),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X_input)
X0 = LeakyReLU(alpha=0.3)(X0)
X0 = BatchNormalization()(X0)
X0 = Conv2D(64, (5, 5),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X0)
X0 = LeakyReLU(alpha=0.3)(X0)
X0 = BatchNormalization()(X0)
X0 = Conv2D(64, (5, 5),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X0)
X0 = LeakyReLU(alpha=0.3)(X0)
X0 = BatchNormalization()(X0)
# Local path
X = Conv2D(64, (7, 7),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X_input)
X = LeakyReLU(alpha=0.3)(X)
X = BatchNormalization()(X)
X1 = Conv2D(64, (7, 7),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X_input)
X1 = LeakyReLU(alpha=0.3)(X1)
X1 = BatchNormalization()(X1)
X = layers.Maximum()([X, X1])
X = Conv2D(64, (4, 4),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X)
X = LeakyReLU(alpha=0.3)(X)
X = BatchNormalization()(X)
X3 = Conv2D(64, (3, 3),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X)
X3 = LeakyReLU(alpha=0.3)(X3)
X3 = BatchNormalization()(X3)
X31 = Conv2D(64, (3, 3),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X)
X31 = LeakyReLU(alpha=0.3)(X31)
X31 = BatchNormalization()(X31)
X = layers.Maximum()([X3, X31])
X = Conv2D(64, (2, 2),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X)
X = LeakyReLU(alpha=0.3)(X)
X = BatchNormalization()(X)
# Global path
X2 = Conv2D(160, (13, 13),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X_input)
X2 = LeakyReLU(alpha=0.3)(X2)
X2 = BatchNormalization()(X2)
# X21 = Conv2D(160, (13, 13), strides=(1, 1), padding='valid', activation='relu', kernel_regularizer=l1_l2(0.01, 0.01))(X_input)
X21 = Conv2D(160, (13, 13),
strides=(1, 1),
padding='valid',
kernel_regularizer=l1_l2(0.01, 0.01))(X_input)
X21 = LeakyReLU(alpha=0.3)(X21)
X21 = BatchNormalization()(X21)
X2 = layers.Maximum()([X2, X21])
# Merging the two paths
X = Concatenate()([X2, X, X0])
return X
