def get_alex_model(input_shape=(32, 32, 3), classes=10):
img_input = Input(input_shape)
he_init = he_normal()
x = bn_relu_conv(kernel_size=(3, 3),
filters=64,
kernel_initializer=my_init,
strides=(1, 1))(img_input)
x = bn_relu_conv(kernel_size=(3, 3),
filters=64,
kernel_initializer=my_init,
strides=(2, 2))(x)
x = bn_relu_conv(kernel_size=(3, 3),
filters=128,
kernel_initializer=my_init,
strides=(1, 1))(x)
x = bn_relu_conv(kernel_size=(3, 3),
filters=128,
kernel_initializer=my_init,
strides=(2, 2))(x)
x = bn_relu_conv(kernel_size=(3, 3),
filters=256,
kernel_initializer=my_init,
strides=(1, 1))(x)
x = bn_relu(x)
x = Flatten()(x)
dense = Dense(units=classes,
kernel_initializer="he_normal",
activation="softmax")(x)
model = Model(inputs=img_input, outputs=dense)
return model
def _bottleneck(input, nb_filters, init_subsample=(1, 1)):
conv_1_1 = bn_relu_conv(input,
nb_filters,
3,
3,
W_regularizer=l2(weight_decay),
subsample=init_subsample)
conv_3_3 = bn_relu_conv(conv_1_1,
nb_filters,
3,
3,
W_regularizer=l2(weight_decay))
return _shortcut(input, conv_3_3)
def _bottleneck(input, nb_filters, init_subsample=(1, 1)):
conv_1_1 = bn_relu_conv(input, nb_filters, 3, 3, W_regularizer=l2(weight_decay), subsample=init_subsample)
conv_3_3 = bn_relu_conv(conv_1_1, nb_filters, 3, 3, W_regularizer=l2(weight_decay))
return _shortcut(input, conv_3_3)