tf.keras--入门示例:Lenet手写字符分类

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/u010472607/article/details/82319295

运行环境:
Anaconda5.2
Python3.6.5
TensorFlow1.10

Model子类化的方式定义Lenet网络, 区别于一般的直接使用Sequential和Model搭建网络的方式, 代码如下

# -*- coding: utf-8 -*-
"""
tf.keras 标准Lenet
"""
import tensorflow as tf

from tensorflow.keras.models import Model, Sequential
from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, Activation, Flatten, Dense
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.datasets import mnist
from tensorflow.keras.utils import to_categorical
import numpy as np
import matplotlib.pyplot as plt

# tf.enable_eager_execution()

input_shape = 28 * 28
classes = 10

NB_EPOCH = 20
BATCH_SIZE = 200
VERBOSE = 1
OPTIMIZER = Adam()
VALIDATION_SPLIT = 0.2
IMG_ROWS, IMG_COLS = 28, 28
NB_CLASSES = 10
INPUT_SHAPE = (IMG_ROWS, IMG_COLS, 1) # 注意在TF中的数据格式 NHWC

# 加载数据,转换编码格式并归一化
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.astype("float32") / 255.
x_test = x_test.astype(np.float32) / 255.

# 扩展1维, 等效写法
x_train = x_train[:, :, :, None]
x_test = x_test[:, :, :, None]

print(x_train.shape, "train shape")
print(x_test.shape, "test shape")

y_train = to_categorical(y_train, NB_CLASSES)
y_test = to_categorical(y_test, NB_CLASSES)

# 创建网络
"""
Model 子类化

在__init__方法中创建层 并将其设置为示例的属性, 在call方法中前向传播计算

当启用eager执行时，模型子类化特别有用， 因为可以强制写入正向传递

关键点：
虽然Model子类化提供了灵活性，但其代价是更高的复杂性和更多的用户错误机会。
"""
class LeNet(Model):
    def __init__(self, input_shape=(28, 28, 1), num_classes=10):
        # super(LeNet, self).__init__(name="LeNet")
        self.num_classes = num_classes
        ''' 定义要用到的层 layers '''
        # 输入层
        img_input = Input(shape=input_shape)

        # Conv => ReLu => Pool
        x = Conv2D(filters=20, kernel_size=5, padding="same", activation="relu" ,name='block1_conv1')(img_input)
        x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='block1_pool')(x)
        # Conv => ReLu => Pool
        x = Conv2D(filters=50, kernel_size=5, padding="same", activation="relu", name='block1_conv2')(x)
        x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='block1_poo2')(x)
        # 压成一维
        x = Flatten(name='flatten')(x)
        # 全连接层
        x = Dense(units=500, activation="relu", name="f1")(x)
        # softmax分类器
        x = Dense(units=num_classes, activation="softmax", name="prediction")(x)

        # 调用Model类的Model(input, output, name="***")构造方法
        super(LeNet, self).__init__(img_input, x, name="LeNet")

    def call(self, inputs):
        # 前向传播计算
        # 使用在__init__方法中定义的层
        return self.output(inputs)


# model = LeNet(10)
# print(model.output_shape)
# model.summary()

# 初始化优化器和模型
model = LeNet(INPUT_SHAPE, NB_CLASSES)
model.summary()

model.compile(loss="categorical_crossentropy", optimizer=tf.train.RMSPropOptimizer(learning_rate=0.001),
              metrics=["accuracy"])
history = model.fit(x=x_train, y=y_train, batch_size=BATCH_SIZE, epochs=NB_EPOCH, verbose=VERBOSE,
                    validation_split=VALIDATION_SPLIT)
score = model.evaluate(x=x_test, y=y_test, verbose=VERBOSE)
print("test loss:", score[0])
print("test acc:", score[1])

# 列出历史数据
print(history.history.keys())
plt.plot(history.history['acc'])
plt.plot(history.history['val_acc'])
plt.title("model accuracy")
plt.ylabel("accuracy")
plt.xlabel("epoch")
plt.legend(["train", "test"], loc="upper left")
plt.show()

# 汇总损失函数历史数据
plt.plot(history.history["loss"])
plt.plot(history.history["val_loss"])
plt.title("model loss")
plt.ylabel("loss")
plt.xlabel("epoch")
plt.legend(["train", "test"], loc="upper left")
plt.show()

部分输出

(60000, 28, 28, 1) train shape
(10000, 28, 28, 1) test shape
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 28, 28, 1)         0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 28, 28, 20)        520       
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 14, 14, 20)        0         
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 14, 14, 50)        25050     
_________________________________________________________________
block1_poo2 (MaxPooling2D)   (None, 7, 7, 50)          0         
_________________________________________________________________
flatten (Flatten)            (None, 2450)              0         
_________________________________________________________________
f1 (Dense)                   (None, 500)               1225500   
_________________________________________________________________
prediction (Dense)           (None, 10)                5010      
=================================================================
Total params: 1,256,080
Trainable params: 1,256,080
Non-trainable params: 0
_________________________________________________________________
Train on 48000 samples, validate on 12000 samples
Epoch 1/20
  200/48000 [..............................] - ETA: 17:09 - loss: 2.2920 - acc: 0.1050
 2000/48000 [>.............................] - ETA: 1:40 - loss: 2.2958 - acc: 0.0990 
 3800/48000 [=>............................] - ETA: 51s - loss: 2.2971 - acc: 0.1016 
 5800/48000 [==>...........................] - ETA: 32s - loss: 2.2942 - acc: 0.1102
 7800/48000 [===>..........................] - ETA: 23s - loss: 2.2914 - acc: 0.1181
 9800/48000 [=====>........................] - ETA: 17s - loss: 2.2877 - acc: 0.1272
12000/48000 [======>.......................] - ETA: 13s - loss: 2.2808 - acc: 0.1493
14200/48000 [=======>......................] - ETA: 11s - loss: 2.2671 - acc: 0.2039
16400/48000 [=========>....................] - ETA: 9s - loss: 2.2313 - acc: 0.2579 
18600/48000 [==========>...................] - ETA: 7s - loss: 2.1376 - acc: 0.3003
20800/48000 [============>.................] - ETA: 6s - loss: 2.0313 - acc: 0.3377
23000/48000 [=============>................] - ETA: 5s - loss: 1.8994 - acc: 0.3809
25200/48000 [==============>...............] - ETA: 4s - loss: 1.7820 - acc: 0.4192
27400/48000 [================>.............] - ETA: 3s - loss: 1.6768 - acc: 0.4542
29600/48000 [=================>............] - ETA: 3s - loss: 1.5772 - acc: 0.4874
31800/48000 [==================>...........] - ETA: 2s - loss: 1.4859 - acc: 0.5175
34000/48000 [====================>.........] - ETA: 2s - loss: 1.4071 - acc: 0.5435
36200/48000 [=====================>........] - ETA: 1s - loss: 1.3334 - acc: 0.5677
38000/48000 [======================>.......] - ETA: 1s - loss: 1.2792 - acc: 0.5852
40000/48000 [========================>.....] - ETA: 1s - loss: 1.2237 - acc: 0.6035
42000/48000 [=========================>....] - ETA: 0s - loss: 1.1731 - acc: 0.6200
44200/48000 [==========================>...] - ETA: 0s - loss: 1.1221 - acc: 0.6365
46400/48000 [============================>.] - ETA: 0s - loss: 1.0745 - acc: 0.6520
48000/48000 [==============================] - 6s 118us/step - loss: 1.0419 - acc: 0.6625 - val_loss: 0.1084 - val_acc: 0.9673
Epoch 2/20
...
Epoch 19/20
...
Epoch 20/20
  200/48000 [..............................] - ETA: 1s - loss: 2.3955e-06 - acc: 1.0000
 2400/48000 [>.............................] - ETA: 1s - loss: 1.6718e-04 - acc: 1.0000
 4600/48000 [=>............................] - ETA: 1s - loss: 8.8650e-05 - acc: 1.0000
 6800/48000 [===>..........................] - ETA: 0s - loss: 8.8033e-05 - acc: 1.0000
 9000/48000 [====>.........................] - ETA: 0s - loss: 0.0014 - acc: 0.9999    
11200/48000 [======>.......................] - ETA: 0s - loss: 0.0011 - acc: 0.9999
13400/48000 [=======>......................] - ETA: 0s - loss: 9.5830e-04 - acc: 0.9999
15600/48000 [========>.....................] - ETA: 0s - loss: 9.3345e-04 - acc: 0.9999
17800/48000 [==========>...................] - ETA: 0s - loss: 8.3938e-04 - acc: 0.9999
20200/48000 [===========>..................] - ETA: 0s - loss: 7.6238e-04 - acc: 0.9999
22400/48000 [=============>................] - ETA: 0s - loss: 7.5516e-04 - acc: 0.9999
24400/48000 [==============>...............] - ETA: 0s - loss: 7.1492e-04 - acc: 0.9999
26600/48000 [===============>..............] - ETA: 0s - loss: 8.5314e-04 - acc: 0.9998
28800/48000 [=================>............] - ETA: 0s - loss: 8.3093e-04 - acc: 0.9998
31000/48000 [==================>...........] - ETA: 0s - loss: 7.8067e-04 - acc: 0.9998
33200/48000 [===================>..........] - ETA: 0s - loss: 7.3853e-04 - acc: 0.9998
35400/48000 [=====================>........] - ETA: 0s - loss: 8.4584e-04 - acc: 0.9998
37600/48000 [======================>.......] - ETA: 0s - loss: 9.0971e-04 - acc: 0.9997
39800/48000 [=======================>......] - ETA: 0s - loss: 9.6476e-04 - acc: 0.9997
42000/48000 [=========================>....] - ETA: 0s - loss: 9.3635e-04 - acc: 0.9997
44200/48000 [==========================>...] - ETA: 0s - loss: 8.9145e-04 - acc: 0.9997
46400/48000 [============================>.] - ETA: 0s - loss: 8.5120e-04 - acc: 0.9997
48000/48000 [==============================] - 1s 26us/step - loss: 8.4154e-04 - acc: 0.9997 - val_loss: 0.0559 - val_acc: 0.9919
   32/10000 [..............................] - ETA: 1s
 1152/10000 [==>...........................] - ETA: 0s
 2368/10000 [======>.......................] - ETA: 0s
 3552/10000 [=========>....................] - ETA: 0s
 4864/10000 [=============>................] - ETA: 0s
 6176/10000 [=================>............] - ETA: 0s
 7488/10000 [=====================>........] - ETA: 0s
 8800/10000 [=========================>....] - ETA: 0s
10000/10000 [==============================] - 0s 41us/step
test loss: 0.04930802768102574
test acc: 0.9923
dict_keys(['val_loss', 'val_acc', 'loss', 'acc'])

loss曲线

acc曲线