ShuffleNetV2网络结构训练模型（图像分类）

• • 1.ShuffleNetV2网络结构讲解
  • 2.重新搭建ShuffleNetV2模型结构（便于训练）
  • • （1）关于数据集和训练的整个代码可以参考我这篇文章
  • 3.训练结果
  • 4.图像测试

https://blog.csdn.net/Keep_Trying_Go/article/details/124774129

https://mydreamambitious.blog.csdn.net/article/details/123966676
注：只要将代码中的网络结构换成下面的这个ShuffleNetV2结构即可（如果采用的是面向对象的方式设计的网络结构的话，在训练数据的时候需要做一些调整）并且再将代码中的一些名称修改一下即可。
需要修改的名称如下：

# #通道重排
def Shuffle_Channels(inputs):
    # 获取图片batch，高度，宽度，通道数
    batch, h, w, c = inputs.shape
    # 对通道进行变形
    input_reshape = tf.reshape(inputs, [-1, h, w, 2, c // 2])
    # 退通道进行转置
    input_transposed = tf.transpose(input_reshape, [0, 1, 2, 4, 3])
    # 再对通道进行变形并且将通道分为两组
    output_reshape = tf.reshape(input_transposed, [2, -1, h, w, c // 2])
    return output_reshape[0], output_reshape[1]


def ShuffleNetV2UnitC(inputs, inp, oup):
    x0, x1 = Shuffle_Channels(inputs)

    inp = inp // 2
    output_channel = oup - inp

    x = layers.Conv2D(oup // 2, kernel_size=[1, 1], strides=[1, 1], padding='same')(x1)
    x = layers.BatchNormalization()(x)
    x = layers.Activation('relu')(x)

    x = layers.DepthwiseConv2D(kernel_size=[3, 3], strides=[1, 1], padding='same')(x)
    x = layers.BatchNormalization()(x)

    x = layers.Conv2D(output_channel, kernel_size=[1, 1], strides=[1, 1], padding='same')(x)
    x = layers.BatchNormalization()(x)
    x = layers.Activation('relu')(x)

    x_out = tf.concat([
        x0, x
    ], axis=3)
    return x_out


def ShuffleNetV2UnitD(inputs, inp, oup):
    output_channel = oup - inp

    # right
    x = layers.Conv2D(oup // 2, kernel_size=[1, 1], strides=[1, 1], padding='same')(inputs)
    x = layers.BatchNormalization()(x)
    x = layers.Activation('relu')(x)

    x = layers.DepthwiseConv2D(kernel_size=[3, 3], strides=[2, 2], padding='same')(x)
    x = layers.BatchNormalization()(x)

    x = layers.Conv2D(output_channel, kernel_size=[1, 1], strides=[1, 1], padding='same')(x)
    x = layers.BatchNormalization()(x)
    x = layers.Activation('relu')(x)

    # left
    x0 = layers.DepthwiseConv2D(kernel_size=[3, 3], strides=[2, 2], padding='same')(inputs)
    x0 = layers.BatchNormalization()(x0)

    x0 = layers.Conv2D(inp, kernel_size=[1, 1], strides=[1, 1], padding='same')(x0)
    x0 = layers.BatchNormalization()(x0)
    x0 = layers.Activation('relu')(x0)

    x_out = tf.concat([
        x0, x
    ], axis=3)
    return x_out


def ShuffleNetV2(inputFilter, numlayers, StageFilter, input_shape=(224, 224, 3)):
    inputs = keras.Input(input_shape)

    x = layers.Conv2D(inputFilter, kernel_size=[3, 3], strides=[2, 2], padding='same')(inputs)
    x = layers.BatchNormalization()(x)
    x = layers.Activation('relu')(x)

    x = layers.MaxPool2D(pool_size=[3, 3], strides=[2, 2], padding='same')(x)

    def StageX(inputs, numlayers, inputFilter, outFilter):
        x = ShuffleNetV2UnitD(inputs, inp=inputFilter, oup=outFilter)

        for i in range(numlayers):
            inputFilter = outFilter
            x = ShuffleNetV2UnitC(x, inp=inputFilter, oup=outFilter)
        return x

    x = StageX(x, numlayers[0], inputFilter, StageFilter[0])
    x = StageX(x, numlayers[1], StageFilter[0], StageFilter[1])
    x = StageX(x, numlayers[2], StageFilter[1], StageFilter[2])

    x = layers.Conv2D(1024, kernel_size=[1,1], strides=[1, 1], padding='same')(x)

    x = layers.GlobalAveragePooling2D()(x)
    x = layers.Dense(2)(x)
    x = layers.Dropout(0.5)(x)
    x = layers.Activation('softmax')(x)

    model = Model(inputs=inputs, outputs=x)

    return model


model_shufflenetv2 = ShuffleNetV2(inputFilter=24, numlayers=[3, 7, 3], StageFilter=[116, 232, 464],
                                  input_shape=(224, 224, 3))
# model_shufflenetv2.summary()

由于数据集不是很多，只有近1000张图片，并且之包含两个类别（猫-cat,狗-dot）所以训练的最后效果并不怎么样（中间还需要调参），读者可以根据自己的需要收集图片进行训练。

只需要将flask中加载的权重文件换成自己训练的就可以了（flask这部分代码在从Github下载的文件中）。
https://github.com/KeepTryingTo/-.git