tensorflow2.0 Keras自定义层与网络

tensorflow2.0 Keras自定义层与网络

   使用keras通常的写法都是先调Sequential类（仅用于层的线性堆叠），如：

nn = Sequential([layers.Dense(128,activation='relu),
							layers.Dense(64,activation='relu),
							layers.Dense(10)])

   有时用已有的layer并不能满足我们的需要，比如在层中加入：

x = x-1

   这时就不得不加载自定义的Layer或Model，以适应各种需求，自定义层继承于keras.layers.Layer，自定义网络继承于keras.Model.
   自定义层或者网络时需定义 init 方法以及 call 方法：

class MyDenseLayer(layers.Layer):
	def __init__(self, input_dim, output_dim):
		super(MyDenseLayer, self).__init__()

		# add_variable已定义在父类中，用于添加trainable的变量
		self.kernel = self.add_variable('weight', [input_dim, output_dim]) 
		self.bias = self.add_variable('bias', [output_dim]) 
		
	#这里使用inputs作为输入参数，training用于控制训练或测试
	def call(self, inputs, training=None):
		out = inputs @ self.kernel + self.bias
		return out

   类似地，通过继承Model自定义网络实现fashion_mnist数据集的训练。继承Model父类后，仍然可以很方便地调用keras提供的compile、fit、predict等方法。

import tensorflow as tf
from tensorflow import keras
import datetime

# 自定义网络
class MyModel(keras.Model):
    def __init__(self):
        super(MyModel, self).__init__()
        self.layer1 = keras.layers.Dense(128, activation=tf.nn.relu)
        self.layer2 = keras.layers.Dense(64, activation=tf.nn.relu)
        self.layer3 = keras.layers.Dense(10)

    def call(self, inputs, training=False):
        inputs = tf.reshape(inputs, [-1, 28*28])
        inputs = self.layer1(inputs)
        inputs = self.layer2(inputs)
        inputs = self.layer3(inputs)

        return inputs

# 转换数据为tensor类型
def preprocessing(x,y):
    x = tf.cast(x, tf.float32) / 255
    y = tf.cast(y, tf.float32)
    return x,y

# 加载数据
(train_images,train_labels),(test_images,test_labels) = keras.datasets.fashion_mnist.load_data()

# 预处理数据并定义iterator
train_labels = tf.one_hot(tf.convert_to_tensor(train_labels),depth=10)
test_labels = tf.one_hot(tf.convert_to_tensor(test_labels),depth=10)
train_iterator = tf.data.Dataset.from_tensor_slices((train_images,train_labels))
test_iterator = tf.data.Dataset.from_tensor_slices((test_images,test_labels))

train_iterator = train_iterator.map(preprocessing).shuffle(train_images.shape[0]).batch(100)
test_iterator = test_iterator.map(preprocessing).shuffle(test_images.shape[0]).batch(100)
sample_x,sample_y = next(iter(train_iterator))
print(sample_x.shape, sample_y.shape)

model = MyModel()
model.compile(optimizer=keras.optimizers.Adam(1e-2),
              loss=tf.losses.CategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
model.fit(train_iterator.repeat(), epochs=10, steps_per_epoch=500, verbose=1,
              validation_data=test_iterator.repeat(),
              validation_steps=2)