def cnn_seq(num_filters=16,
kernel_size=3,
strides=1,
activation=ReLU,
batch_normalization=True,
conv_first=True):
seq = Sequential()
seq.add(
Conv2D(num_filters,
kernel_size=kernel_size,
strides=strides,
padding='same',
kernel_regularizer=l2(5e-3)))
if batch_normalization:
seq.add(BatchNormalization_v2())
seq.add(activation())
seq.add(
Conv2D(num_filters,
kernel_size=kernel_size,
strides=1,
padding='same',
kernel_regularizer=l2(5e-3)))
if batch_normalization:
seq.add(BatchNormalization_v2())
return seq
def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(ReLU())
model.add(Flatten())
model.add(Dense(2500, kernel_initializer='He'))
model.add(ReLU())
model.add(Dense(1500, kernel_initializer='He'))
model.add(ReLU())
model.add(Dense(10, kernel_initializer='He'))
model.add(Softmax())
model.summary()
model.compile(Adam(), 'categorical_crossentropy', 'accuracy')
return model
def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(Conv2D(32, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(MaxPooling2D(2, 2, stride=2))
model.add(Conv2D(64, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(Conv2D(64, kernel_size=(3, 3), padding='same'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(MaxPooling2D(2, 2, stride=2))
model.add(Flatten())
model.add(Dense(512, kernel_initializer='He'))
model.add(BN_LAYER())
model.add(ReLU())
model.add(Dense(10, kernel_initializer='He'))
model.add(Softmax())
model.summary()
model.compile(Adam(), loss='categorical_crossentropy', metric='accuracy')
return model
def make_model():
model = Sequential()
model.add(Input(shape=input_shape))
model.add(
Conv2D(16,
kernel_size=3,
strides=1,
padding='same',
kernel_regularizer=l2(1e-4)))
model.add(BatchNormalization_v2())
model.add(ReLU())
add_residual_block(model, num_filters=16)
add_residual_block(model, num_filters=16)
add_residual_block(model, num_filters=16)
add_residual_block(model, num_filters=32, strides=2, cnn_shortcut=True)
add_residual_block(model, num_filters=32)
add_residual_block(model, num_filters=32)
add_residual_block(model, num_filters=64, strides=2, cnn_shortcut=True)
add_residual_block(model, num_filters=64)
add_residual_block(model, num_filters=64)
model.add(AveragePooling2DAll())
model.add(Flatten())
model.add(Dense(10, kernel_initializer='He'))
model.add(Softmax())
model.summary()
model.compile(Adam(lr=0.001, decay=1e-4), 'categorical_crossentropy',
'accuracy')
return model
def add_residual_block(model,
num_filters=16,
kernel_size=3,
strides=1,
activation=ReLU,
cnn_shortcut=False,
batch_normalization=True,
conv_first=True):
if cnn_shortcut:
shortcut = Conv2D(num_filters,
kernel_size=1,
strides=strides,
padding='same',
kernel_regularizer=l2(5e-3))
else:
shortcut = Same()
cnn_road = cnn_seq(num_filters, kernel_size, strides, activation,
batch_normalization, conv_first)
model.add(Separate())
model.add([shortcut, cnn_road])
model.add(Add())
model.add(ReLU())
return model
def make_my_yolo(input_shape, num_anchors, num_classes):
ori_c, ori_w, ori_h = input_shape
model = Sequential()
model.add(Input(shape=input_shape))
add_conv2d_bn_leaky(model, 16, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 32, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 64, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 128, 3)
model.add(MaxPooling2D(2, 2))
model.add(Conv2D(num_anchors * (4 + 1 + num_classes), 1, 1))
model.add(
Reshape(
(num_anchors, (4 + 1 + num_classes), ori_w // 16, ori_h // 16)))
model.add(Transpose((0, 1, 3, 4, 2)))
mask = np.ones((3, 7, 7, 7), dtype=bool)
mask[..., 2] = False
mask[..., 3] = False
model.add(MaskedSigmoid(mask))
return model
def add_conv2d_bn_leaky(model, num_filters=16, kernel_size=3, strides=1):
model.add(
Conv2D(num_filters,
kernel_size,
strides,
padding='same',
use_bias=False))
model.add(BatchNormalization_v2())
model.add(LeakyReLU(0.1))
def add_resnet_layer(model, num_filters=16, kernel_size=3, strides=1,
activation='relu', batch_normalization=True, conv_first=True):
conv = Conv2D(num_filters, kernel_size, strides=strides, padding='same', kernel_regularizer=l2(l2_lambda))
if conv_first:
model.add(conv)
if batch_normalization:
model.add(BatchNormalization_v2())
if activation is not None:
model.add(Activation(activation))
else:
if batch_normalization:
model.add(BatchNormalization_v2())
if activation is not None:
model.add(Activation(activation))
model.add(conv)
return model
def make_darknet19(input_shape):
model = Sequential()
model.add(Input(shape=input_shape))
add_conv2d_bn_leaky(model, 32, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 64, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 128, 3)
add_conv2d_bn_leaky(model, 64, 1)
add_conv2d_bn_leaky(model, 128, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 256, 3)
add_conv2d_bn_leaky(model, 128, 1)
add_conv2d_bn_leaky(model, 256, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 512, 3)
add_conv2d_bn_leaky(model, 256, 1)
add_conv2d_bn_leaky(model, 512, 3)
add_conv2d_bn_leaky(model, 256, 1)
add_conv2d_bn_leaky(model, 512, 3)
model.add(MaxPooling2D(2, 2))
add_conv2d_bn_leaky(model, 1024, 3)
add_conv2d_bn_leaky(model, 512, 1)
add_conv2d_bn_leaky(model, 1024, 3)
add_conv2d_bn_leaky(model, 512, 1)
add_conv2d_bn_leaky(model, 1024, 3)
add_conv2d_bn_leaky(model, 1000, 1)
model.add(Conv2D(5 * (4 + 1 + 1), 1, 1))
return model