通过 tensorflow 将归一化重现到 keras

Question

实际上我正在尝试在 keras 上重现一个 tensorflow 模型，我对这个话题真的很陌生。我想重现这些行

embedding = tf.layers.conv2d(conv6, 128, (16, 16), padding='VALID', name='embedding')
embedding = tf.reshape(embedding, (-1, 128))
embedding = embedding - tf.reduce_min(embedding, keepdims =True)
z_n = embedding/tf.reduce_max(embedding, keepdims =True)

我的实际代码是：

def conv_conv_pool(n_filters,
                   name,
                   pool=True,
                   activation=tf.nn.relu, padding='same', filters=(3,3)):
    """{Conv -> BN -> RELU}x2 -> {Pool, optional}
    Args:
        input_ (4-D Tensor): (batch_size, H, W, C)
        n_filters (list): number of filters [int, int]
        training (1-D Tensor): Boolean Tensor
        name (str): name postfix
        pool (bool): If True, MaxPool2D
        activation: Activaion functions
    Returns:
        net: output of the Convolution operations
        pool (optional): output of the max pooling operations
    """
    net = Sequential()
    for i, F in enumerate(n_filters):
        conv = Conv2D(
            filters = F,
            kernel_size = (3,3),
            padding = 'same',
            )
        net.add(conv)
        batch_norm = BatchNormalization()
        net.add(batch_norm)
        net.add(Activation('relu'))

    if pool is False:
        return net

    pool = Conv2D(
        filters = F,
        kernel_size = (3,3),
        strides = (2,2),
        padding = 'same',  
        )
    net.add(pool)
    batch_norm = BatchNormalization()
    net.add(batch_norm)
    net.add(Activation('relu'))
    return net


def model_keras():
    model = Sequential()
    model.add(conv_conv_pool(n_filters = [8, 8], name="1"))
    model.add(conv_conv_pool([32, 32], name="2"))
    model.add(conv_conv_pool([32, 32], name="3"))
    model.add(conv_conv_pool([64, 64], name="4"))
    model.add(conv_conv_pool([64, 64], name="5"))
    model.add(conv_conv_pool([128, 128], name="6", pool=False))
    return model

归一化应该在第 6 层之后。

我在考虑使用 lambda 层，这对吗？是的话应该怎么写？

Answer 1

我相信您想切换到使用 keras 作为 API 的 tensorflow 2。你需要 install/upgrade 到 tensorflow 2，然后你可以试试这个：

import tensorflow as tf

embedding = tf.keras.layers.conv2d(conv6, 128, (16, 16), padding='VALID', 
            name='embedding')
embedding = tf.keras.layers.reshape(embedding, (-1, 128))
embedding = embedding - tf.math.reduce_min(embedding, keepdims =True)
z_n = embedding/tf.math.reduce_max(embedding, keepdims =True)

Answer 2

如果你想使用 keras 层 api 你可以创建一个自定义层，你可以在这里找到如何做的文档 https://www.tensorflow.org/guide/keras/custom_layers_and_models, 你应该以这样的方式结束：

class NormalizationLayer(layers.Layer):

  def __init__(self, filters=128):
    super(NormalizationLayer, self).__init__()
    self.filters = filters

  def call(self, inputs):
    embedding = tf.keras.layers.conv2d(inputs, self.filters, (16, 16), padding='VALID', 
            name='embedding')
    embedding = tf.keras.layers.reshape(embedding, (-1, self.filters))
    embedding = embedding - tf.math.reduce_min(embedding, keepdims =True)
    z_n = embedding/tf.math.reduce_max(embedding, keepdims =True)
    return zn

Answer 3

我使用您在 Lambda 层中引入的规范化。我还做了一个更正（最小值和最大值是在同一个输入上计算的，而不是一个在输入上计算，另一个在转换上计算），但你也可以改变它。 norm_original 使用在所有通道上计算的最小值和最大值对 4D 输入进行归一化，并尝试 return 具有固定数量特征的 2D 输出这将产生错误，因为您正在修改批量维度

def norm_original(inp):

    embedding = tf.reshape(inp, (-1, inp.shape[-1]))
    embedding = embedding - tf.reduce_min(inp)
    embedding = embedding / tf.reduce_max(inp)

    return embedding

inp = Input((28,28,3))
x = Conv2D(128, 3, padding='same')(inp)
x = Lambda(norm_original)(x)

m = Model(inp, x)
m.compile('adam', 'mse')
m.summary()

X = np.random.uniform(0,1, (10,28,28,3))
y = np.random.uniform(0,1, (10,128))

m.fit(X,y, epochs=3) # error

为了避免这个错误，我提出了两种可能性。我也做了修改，按通道操作归一化（我保留它更合适）但你也可以修改它。

1) 您可以使用 min/max 规范化 4D 输入，然后将输出展平，将所有内容放在最后一个维度上。此解决方案不会交替批处理 dim

def norm(inp):
    ## this function operate normalization by channels
    embedding = inp - tf.reduce_min(inp, keepdims=True, axis=[0,1,2])
    embedding = embedding / tf.reduce_max(inp, keepdims=True, axis=[0,1,2])

    return embedding

inp = Input((28,28,3))
x = Conv2D(128, 3, padding='same')(inp)
x = Lambda(norm)(x)
x = Flatten()(x)
x = Dense(128, activation='relu')(x)

m = Model(inp, x)
m.compile('adam', 'mse')

X = np.random.uniform(0,1, (10,28,28,3))
y = np.random.uniform(0,1, (10,128))

m.fit(X,y, epochs=3)

2) 您可以使用 GlobalPooling 层来减少 4D 维度并重新生成 2D 形状，保留特征维度

inp = Input((28,28,3))
x = Conv2D(128, 3, padding='same')(inp)
x = Lambda(norm)(x)
x = GlobalMaxPool2D()(x) # u can also use GlobalAveragePooling2D

m = Model(inp, x)
m.compile('adam', 'mse')

X = np.random.uniform(0,1, (10,28,28,3))
y = np.random.uniform(0,1, (10,128))

m.fit(X,y, epochs=3)

通过 tensorflow 将归一化重现到 keras

Reproduce normalization by tensorflow to keras

python

normalization

keras

tensorflow