本笔记记录使用自定义Layer和Model来做CIFAR10数据集的训练。
CIFAR10数据集下载：
https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
自定义的Layer和Model实现较为简单，参数量较少，并且没有卷积层和dropout等，最终准确率不高，仅做练习使用。
import tensorflow as tf
import numpy as np
from tensorflow import keras
from tensorflow.keras import datasets, layers, optimizers, Sequential, metricstf.__version__def preprocess(x, y):x = tf.cast(x, dtype=tf.float32) / 255y = tf.cast(y, dtype=tf.int32)return x,ybatchsize = 128
#CIFAR10数据集下载，可以直接使用网络下载
(x,y), (x_val, y_val) = datasets.cifar10.load_data()
#CIFAR10的标签（训练集）数据维度是[50000, 1],通过squeeze消除掉里面1的维度，变成[50000]
print("y.shape:", y.shape)
y = tf.squeeze(y)
print("squeezed y.shape:", y.shape)
y_val = tf.squeeze(y_val)
#进行onehot编码
y = tf.one_hot(y, depth=10)
y_val = tf.one_hot(y_val, depth=10)
print("Datasets: ", x.shape, " ", y.shape, " x.min():", x.min(), " x.max():", x.max())train_db = tf.data.Dataset.from_tensor_slices((x, y))
train_db = train_db.map(preprocess).shuffle(10000).batch(batchsize)
test_db = tf.data.Dataset.from_tensor_slices((x_val, y_val))
test_db = test_db.map(preprocess).batch(batchsize)sample = next(iter(train_db))
print("Batch:", sample[0].shape, sample[1].shape)#自定义Layer
class MyDense(layers.Layer):def __init__(self, input_dim, output_dim):super(MyDense, self).__init__()self.kernel = self.add_weight(name='w', shape=[input_dim, output_dim], initializer=tf.random_uniform_initializer(0, 1.0))self.bias = self.add_weight(name='b', shape=[output_dim], initializer=tf.random_uniform_initializer(0, 1.0))#self.kernel = self.add_weight(name='w', shape=[input_dim, output_dim])#self.bias = self.add_weight(name='b', shape=[output_dim])def call(self, inputs, training = None):x = inputs@self.kernel + self.biasreturn xclass MyNetwork(keras.Model):def __init__(self):super(MyNetwork, self).__init__()self.fc1 = MyDense(32 * 32 * 3, 512)self.fc2 = MyDense(512, 512)self.fc3 = MyDense(512, 256)self.fc4 = MyDense(256, 256)self.fc5 = MyDense(256, 10)def call(self, inputs, training = None):x = tf.reshape(inputs, [-1, 32 * 32 * 3])x = self.fc1(x)x = tf.nn.relu(x)x = self.fc2(x)x = tf.nn.relu(x)x = self.fc3(x)x = tf.nn.relu(x)x = self.fc4(x)x = tf.nn.relu(x)x = self.fc5(x)x = tf.nn.relu(x)#返回logitsreturn xtotal_epoches = 35
learn_rate = 0.001
network = MyNetwork()
network.compile(optimizer=optimizers.Adam(learning_rate=learn_rate),loss = tf.losses.CategoricalCrossentropy(from_logits=True),metrics=['Accuracy'])
network.fit(train_db, epochs=total_epoches, validation_data=test_db, validation_freq=1)运行结果：
