Tensorflow2构建模型-境界2

1.导入CIFAR-10数据集

CIFAR-10是由 Hinton 的学生 Alex Krizhevsky 和 Ilya Sutskever 整理的一个用于识别普适物体的小型数据集。一共包含 10 个类别的 RGB 彩色图片:飞机( a叩lane )、汽车( automobile )、鸟类( bird )、猫( cat )、鹿( deer )、狗( dog )、蛙类( frog )、马( horse )、船( ship )和卡车( truck )。图片的尺寸为 32×32,3个通道 ,数据集中一共有 50000 张训练圄片和 10000 张测试图片。 CIFAR-10数据集有3个版本,这里使用python版本。

1.1 导入需要的库

1.2 定义批量导入数据的函数

1.3 定义加载数据函数

1.4 加载数据

运行结果
loading data\cifar-10-batches-py\data_batch_1
loading data\cifar-10-batches-py\data_batch_2
loading data\cifar-10-batches-py\data_batch_3
loading data\cifar-10-batches-py\data_batch_4
loading data\cifar-10-batches-py\data_batch_5
finished loadding CIFAR-10 data

1.5 可视化加载数据

运行结果

2 .数据预处理并设置超参数

3.使用tf.data构建数据管道

4.定义卷积层及池化层

5.构建模型

6.定义训练模型函数

自定义训练过程:
(1)打开一个遍历各epoch的for循环
(2)对于每个epoch,打开一个分批遍历数据集的 for 循环
(3)对于每个批次,打开一个 GradientTape() 作用域
(4)在此作用域内,调用模型(前向传递)并计算损失
(5)在作用域之外,检索模型权重相对于损失的梯度
(6)根据梯度使用优化器来更新模型的权重
(7)评估模型指标

train_model(model,train_data,training_steps,display_step)
运行结果
step =1000,loss = 1.2807,accuracy =0.5781 ,times=0.0080
step =2000,loss = 1.1832,accuracy =0.6562 ,times=0.0080
step =3000,loss = 0.9727,accuracy =0.6562 ,times=0.0080
step =4000,loss = 1.0398,accuracy =0.6406 ,times=0.0050
step =5000,loss = 0.8615,accuracy =0.6406 ,times=0.0156
step =6000,loss = 0.7207,accuracy =0.7188 ,times=0.0000
step =7000,loss = 1.0945,accuracy =0.5938 ,times=0.0090
step =8000,loss = 0.7337,accuracy =0.7656 ,times=0.0080
step =9000,loss = 0.5792,accuracy =0.7812 ,times=0.0080
step =10000,loss = 0.7154,accuracy =0.7500 ,times=0.0080
step =11000,loss = 0.6398,accuracy =0.8125 ,times=0.0156
step =12000,loss = 0.6413,accuracy =0.7500 ,times=0.0080
step =13000,loss = 0.5555,accuracy =0.7812 ,times=0.0156
step =14000,loss = 0.6729,accuracy =0.8281 ,times=0.0156
step =15000,loss = 0.5163,accuracy =0.7500 ,times=0.0156
step =16000,loss = 0.5521,accuracy =0.7969 ,times=0.0000
step =17000,loss = 0.4475,accuracy =0.8594 ,times=0.0156
step =18000,loss = 0.3158,accuracy =0.8594 ,times=0.0080
step =19000,loss = 0.3829,accuracy =0.9062 ,times=0.0090
step =20000,loss = 0.2731,accuracy =0.9062 ,times=0.0080

7.可视化运行结果

8.测试型

运行结果
Test accuracy:0.712640

发表评论