SqueezeNet

SqueezeNet

背景介绍

  SqueezeNet:是一种轻量级深度神经网络模型，在2017年发表于ICLR，作者来自Berkeley和Stanford，其只用1/50的参数量，可以达到与AlexNet相同的精度，其核心结构为Fire Module。

SqueezeNet特点

  引入Fire Module，根据降维思想，先通过1x1的卷积核对参数量进行压缩，然后采用了Inception的思想，进行多路融合。

SqueezeNet图像分析

TensorFlow2.0实现

from functools import reduce
import tensorflow.keras as keras


def compose(*funcs):
    if funcs:
        return reduce(lambda f, g: lambda *a, **kw: g(f(*a, **kw)), funcs)
    else:
        raise ValueError('Composition of empty sequence not supported.')


def fire_block(x, s1, e1, e3, name):
    x = keras.layers.Conv2D(s1, (1, 1), activation='relu', name='{}_conv'.format(name))(x)
    x1 = keras.layers.Conv2D(e1, (1, 1), activation='relu', name='{}_part1_conv'.format(name))(x)
    x2 = keras.layers.Conv2D(e3, (3, 3), padding='same', activation='relu', name='{}_part2_conv'.format(name))(x)
    x = keras.layers.Concatenate(name='{}_concatenate'.format(name))([x1, x2])

    return x


def squeezenet(input_shape):
    input_tensor = keras.layers.Input(input_shape, name='input')
    x = input_tensor

    x = compose(keras.layers.Conv2D(96, (7, 7), (2, 2), padding='same', activation='relu', name='conv1'),
                keras.layers.MaxPool2D((3, 3), (2, 2), name='maxpool'))(x)

    x = fire_block(x, 16, 64, 64, 'fire_block1_1')
    x = fire_block(x, 16, 64, 64, 'fire_block1_2')

    x = fire_block(x, 32, 128, 128, 'fire_block2_1')
    x = keras.layers.MaxPool2D((3, 3), (2, 2), name='fire_block2_maxpool')(x)
    x = fire_block(x, 32, 128, 128, 'fire_block2_2')

    x = fire_block(x, 48, 192, 192, 'fire_block3_1')
    x = fire_block(x, 48, 192, 192, 'fire_block3_2')

    x = fire_block(x, 64, 256, 256, 'fire_block4_1')
    x = keras.layers.MaxPool2D((3, 3), (2, 2), name='fire_block4_maxpool')(x)
    x = fire_block(x, 64, 256, 256, 'fire_block4_2')

    x = compose(keras.layers.Dropout(0.5, name='dropout'),
                keras.layers.Conv2D(1000, (1, 1), activation='relu', name='conv2'),
                keras.layers.GlobalAveragePooling2D(name='global_averagepool'),
                keras.layers.Softmax(name='softmax'))(x)

    model = keras.Model(input_tensor, x, name='SqueezeNet')

    return model


if __name__ == '__main__':

    model = squeezenet(input_shape=(224, 224, 3))
    model.build(input_shape=(None, 224, 224, 3))
    model.summary()

SqueezeNet小结

  SqueezeNet是一种简单的轻量级深度学习网络，从上图可以看出SqueezeNet模型的参数量只有1M，因此在某些特殊场合中能够发挥出很好的效果。