def depthwiseConvBlock(x,
features_in,
features_out,
down_sample=False,
kernel_size=(3, 3),
final_activation='relu',
dilation_rate=[1, 1]):
strides = (2, 2) if down_sample else (1, 1)
x = DepthwiseConvolution2D(int(features_in),
kernel_size,
strides=strides,
padding='same',
use_bias=False,
dilation_rate=dilation_rate)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(features_out), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
if final_activation == 'leaky_relu':
x = LeakyReLU(alpha=0.3)(x)
else:
x = Activation(final_activation)(x)
return x
def _depthwise_convolution2D(input,
alpha,
deepwise_filter_size,
kernel_size,
strides,
padding='same',
bias=False):
x = DepthwiseConvolution2D(int(deepwise_filter_size * alpha),
kernel_size,
strides=strides,
padding=padding,
use_bias=bias)(input)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.1)(x)
return x
def MobileNet(input_tensor=None, input_shape=None, alpha=1, shallow=False, classes=1000):
"""Instantiates the MobileNet.Network has two hyper-parameters
which are the width of network (controlled by alpha)
and input size.
# Arguments
input_tensor: optional Keras tensor (i.e. output of `layers.Input()`)
to use as image input for the model.
input_shape: optional shape tuple, only to be specified
if `include_top` is False (otherwise the input shape
has to be `(224, 224, 3)` (with `channels_last` data format)
or `(3, 224, 244)` (with `channels_first` data format).
It should have exactly 3 inputs channels,
and width and height should be no smaller than 96.
E.g. `(200, 200, 3)` would be one valid value.
alpha: optional parameter of the network to change the
width of model.
shallow: optional parameter for making network smaller.
classes: optional number of classes to classify images
into.
# Returns
A Keras model instance.
"""
input_shape = _obtain_input_shape(input_shape,
default_size=224,
min_size=96,
data_format=K.image_data_format(),
include_top=True)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
x = Convolution2D(int(32 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(img_input)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(32 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(64 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
if not shallow:
for _ in range(5):
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(1024 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(1024 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(1024 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = GlobalAveragePooling2D()(x)
out = Dense(classes, activation='softmax')(x)
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
else:
inputs = img_input
model = Model(inputs, out, name='mobilenet')
return model
def mobilenet(input_tensor):
if input_tensor is None:
input_tensor = Input(shape=(300, 300, 3))
x = ZeroPadding2D(padding=((1, 1), (1, 1)),
name='conv1_padding')(input_tensor)
x = Convolution2D(32, (3, 3),
strides=(2, 2),
padding='valid',
use_bias=False,
name="conv0")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv0/bn")(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(32, (3, 3),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv1/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv1/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(64, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv1")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv1/bn")(x)
x = Activation('relu')(x)
print("conv1 shape: ", x.shape)
x = ZeroPadding2D(padding=((1, 1), (1, 1)), name='conv2_padding')(x)
x = DepthwiseConvolution2D(64, (3, 3),
strides=(2, 2),
padding='valid',
use_bias=False,
name="conv2/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv2/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(128, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv2")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv2/bn")(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(128, (3, 3),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv3/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv3/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(128, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv3")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv3/bn")(x)
x = Activation('relu')(x)
print("conv3 shape: ", x.shape)
x = ZeroPadding2D(padding=((1, 1), (1, 1)), name='conv3_padding')(x)
x = DepthwiseConvolution2D(128, (3, 3),
strides=(2, 2),
padding='valid',
use_bias=False,
name="conv4/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv4/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(256, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv4")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv4/bn")(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(256, (3, 3),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv5/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv5/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(256, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv5")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv5/bn")(x)
x = Activation('relu')(x)
print("conv5 shape: ", x.shape)
x = ZeroPadding2D(padding=((1, 1), (1, 1)), name='conv4_padding')(x)
x = DepthwiseConvolution2D(256, (3, 3),
strides=(2, 2),
padding='valid',
use_bias=False,
name="conv6/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv6/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(512, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv6")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv6/bn")(x)
x = Activation('relu')(x)
test = x
for i in range(5):
x = DepthwiseConvolution2D(512, (3, 3),
strides=(1, 1),
padding='same',
use_bias=False,
name=("conv" + str(7 + i) + "/dw"))(x)
x = BatchNormalization(momentum=0.99,
epsilon=0.00001,
name=("conv" + str(7 + i) + "/dw/bn"))(x)
x = Activation('relu')(x)
x = Convolution2D(512, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name=("conv" + str(7 + i)))(x)
x = BatchNormalization(momentum=0.99,
epsilon=0.00001,
name=("conv" + str(7 + i) + "/bn"))(x)
x = Activation('relu')(x)
# print ("conv11 shape: ", x.shape)
conv11 = x
x = ZeroPadding2D(padding=((1, 1), (1, 1)), name='conv5_padding')(x)
x = DepthwiseConvolution2D(512, (3, 3),
strides=(2, 2),
padding='valid',
use_bias=False,
name="conv12/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv12/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(1024, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv12")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv12/bn")(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(1024, (3, 3),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv13/dw")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001,
name="conv13/dw/bn")(x)
x = Activation('relu')(x)
x = Convolution2D(1024, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
name="conv13")(x)
x = BatchNormalization(momentum=0.99, epsilon=0.00001, name="conv13/bn")(x)
x = Activation('relu')(x)
conv13 = x
# print ("conv13 shape: ", x.shape)
# model = Model(inputs=input_tensor, outputs=x)
return [conv11, conv13, test]
def VGG_with_MobileNet(input_tensor=None,
input_shape=None,
alpha=1,
shallow=False,
classes=10):
input_shape = _obtain_input_shape(input_shape,
default_size=224,
min_size=96,
data_format=K.image_data_format(),
include_top=True)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
""" Input and 3x3 conv 64 filters"""
x = Convolution2D(int(64 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(img_input)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 64 filters and maxpooling by 2"""
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3),
strides=(2, 2),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(64 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 128 filters"""
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 128 filters and maxpooling by 2"""
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(2, 2),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 256 filters"""
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 256 filters"""
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 256 filters and maxpooling by 2"""
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(2, 2),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.05)(x)
""" 3x3 conv 512 filters"""
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 512 filters"""
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 512 filters and maxpooling by 2"""
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(2, 2),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.1)(x)
""" 3x3 conv 512 filters"""
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 512 filters"""
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
""" 3x3 conv 512 filters and maxpooling by 2"""
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(2, 2),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.2)(x)
""" 7x7 conv 512 filters"""
x = DepthwiseConvolution2D(int(512 * alpha), (7, 7),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.3)(x)
""" 7x7 conv 512 filters"""
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
W_regularizer=l2(0.001),
use_bias=False)(x)
x = BatchNormalization()(x)
#x = Activation('relu')(x)
x = LeakyReLU(alpha=0.01)(x)
x = Dropout(0.4)(x)
x = GlobalAveragePooling2D()(x)
out = Dense(classes, activation='softmax')(x)
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
else:
inputs = img_input
model = Model(inputs, out, name='VGG_with_MobileNet')
model.load_weights('weights.h5')
adam = Adam(lr=0.0001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.000001)
model.compile(adam, loss='categorical_crossentropy', metrics=['accuracy'])
return model
def MobileNetDih4(alpha=1.0, shape=[75,75,2]):
img_input = Input(shape)
xd = []
for d in Dih4(img_input, name='1'):
x1 = Convolution2D(int(32 * alpha), (3, 3), strides=(2, 2), padding='valid', use_bias=True)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = DepthwiseConvolution2D(int(32 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(64 * alpha/8), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='2'):
x1 = DepthwiseConvolution2D(int(64 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=True)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(128 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(128 * alpha/8), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='3'):
x1 = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=True)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(256 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(256 * alpha/8), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='4'):
x1 = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=True)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(512 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = DepthwiseConvolution2D(int(512 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(512 * alpha/8), (1, 1), strides=(1, 1), padding='same', use_bias=True)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
x = GlobalAveragePooling2D()(x)
x = Dropout(0.)(x)
x1 = Dense(128)(x)
x2 = Dense(128)(x)
x = Maximum()([x1, x2])
x = BatchNormalization()(x)
out = Dense(1, activation='sigmoid')(x)
model = Model(img_input, out, name='mobilenet')
return model
def MobileNet(alpha=1.0, shape=[16, 224, 224, 3], nframe=16):
img_input = Input(shape)
x = TimeDistributed(
Convolution2D(int(32 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False))(img_input)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
DepthwiseConvolution2D(int(32 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
Convolution2D(int(64 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
DepthwiseConvolution2D(int(64 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = TimeDistributed(
Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False))(x)
x = TimeDistributed(BatchNormalization())(x)
x = TimeDistributed(Activation('relu'))(x)
x = ConvLSTM2D(int(128 * alpha),
kernel_size=(7, 7),
activation='relu',
dropout=0.35,
recurrent_dropout=0.,
return_sequences=False)(x)
x = GlobalAveragePooling2D()(x)
out = Dense(6, activation='softmax')(x)
model = Model(img_input, out, name='mobilenet')
return model
def MobileNet(alpha=1.0, shape=[75, 75, 2]):
img_input = Input(shape)
x = Convolution2D(int(32 * alpha), (3, 3),
strides=(2, 2),
padding='valid',
use_bias=False)(img_input)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(32 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(64 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
for _ in range(5):
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = GlobalAveragePooling2D()(x)
x = Dropout(0.35)(x)
x1 = Dense(128)(x)
x2 = Dense(128)(x)
x = Maximum()([x1, x2])
x = BatchNormalization()(x)
out = Dense(1, activation='sigmoid')(x)
model = Model(img_input, out, name='mobilenet')
return model
def MobileNet(input_tensor=None,
input_shape=None,
alpha=1,
shallow=False,
classes=1000):
input_shape = _obtain_input_shape(input_shape,
default_size=224,
min_size=96,
data_format=K.image_data_format(),
require_flatten=True)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
x = Convolution2D(int(32 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(img_input)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(32 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(64 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
if not shallow:
for _ in range(5):
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(1024 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(1024 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(1024 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = GlobalAveragePooling2D()(x)
out = Dense(classes, activation='softmax')(x)
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
else:
inputs = img_input
model = Model(inputs, out, name='mobilenet')
return model
def MobileNetDR(config, alpha=1.0, shape=[224, 224, 3]):
img_input = Input(shape)
#x = Dih4URandom()(img_input)
x = Convolution2D(int(32 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(img_input)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(32 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(64 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dih4URandom()(x)
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(512 * alpha), (3, 3),
strides=(2, 2),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(1024 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(1024 * alpha), (3, 3),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(1024 * alpha), (1, 1),
strides=(1, 1),
padding='same',
use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = GlobalAveragePooling2D()(x)
x = Dropout(0.)(x)
out = []
for l in config:
out.append(Dense(config[l], activation='softmax', name=str(l))(x))
model = Model(img_input, out, name='mobilenet')
return model
def build_model():
shallow=True
input_tensor=None
input_shape=None
classes = 20
input_tensor = Input(shape=(target_size, target_size, 3))
alpha=1
x = Convolution2D(int(32 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(input_tensor)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(32 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(64 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(64 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(128 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(256 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Convolution2D(int(512 * alpha), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = GlobalAveragePooling2D()(x)
out = Dense(classes, activation='softmax')(x)
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
else:
inputs = img_input
model = Model(inputs, out, name='mobilenet')
return model
def MobileNetDih4(config, alpha=1.0, shape=[224,224,3]):
img_input = Input(shape)
x = Convolution2D(int(32 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(img_input)
x = BatchNormalization()(x)
x = Activation('relu')(x)
xd = []
for d in Dih4(x, name='1'):
x1 = DepthwiseConvolution2D(int(32 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(64 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='2'):
x1 = DepthwiseConvolution2D(int(64 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(128 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='3'):
x1 = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(128 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='4'):
x1 = DepthwiseConvolution2D(int(128 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(256 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='5'):
x1 = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(256 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='6'):
x1 = DepthwiseConvolution2D(int(256 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(512 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='7'):
x1 = DepthwiseConvolution2D(int(512 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(512 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
xd = []
for d in Dih4(x, name='8'):
x1 = DepthwiseConvolution2D(int(512 * alpha), (3, 3), strides=(2, 2), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(1024 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
for d in Dih4(x, name='9'):
x1 = DepthwiseConvolution2D(int(1024 * alpha), (3, 3), strides=(1, 1), padding='same', use_bias=False)(d)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
x1 = Convolution2D(int(1024 * alpha/4), (1, 1), strides=(1, 1), padding='same', use_bias=False)(x1)
x1 = BatchNormalization()(x1)
x1 = Activation('relu')(x1)
xd.append(x1)
x = concatenate(xd, axis=-1)
x = GlobalAveragePooling2D()(x)
x = Dropout(0.35)(x)
x = Dense(256)(x)
x = BatchNormalization()(x)
x = Activation('tanh')(x)
out = []
for l in config:
out.append(Dense(config[l], activation='linear', name=str(l))(x))
model = Model(img_input, out, name='mobilenet_dih')
return model