【Keras-CNN】CIFAR-10

最新推荐文章于 2024-05-08 08:57:38 发布

bryant_meng

最新推荐文章于 2024-05-08 08:57:38 发布

阅读量1w

点赞数 15

分类专栏： PyTorch/Keras/Caffe/TensroFlow 文章标签： CNN Keras CIFAR10

本文链接：https://blog.csdn.net/bryant_meng/article/details/81077196

版权

PyTorch/Keras/Caffe/TensroFlow 专栏收录该内容

62 篇文章 9 订阅

订阅专栏

连载系列
https://blog.csdn.net/bryant_meng/article/details/79263542
3.5 Keras Demos（★★★★★）

CIFAR-10¹.（Canadian Institute for Advanced Research）是由 Alex Krizhevsky、Vinod Nair 与 Geoffrey Hinton 收集的一个用于图像识别的数据集，60000个32*32的彩色图像，50000个training data，10000个 test data 有10类，飞机、汽车、鸟、猫、鹿、狗、青蛙、马、船、卡车，每类6000张图。与MNIST相比，色彩、颜色噪点较多，同一类物体大小不一、角度不同、颜色不同。
这里写图片描述

1 Data preprocessing

1.1 下载和读入数据集

import numpy
from keras.datasets import cifar10
import numpy as np
np.random.seed(10)
# 导入数据集，如果没有就会自动下载
(x_img_train,y_label_train),(x_img_test, y_label_test)=cifar10.load_data()
print('train:',len(x_img_train))
print('test :',len(x_img_test))
print('train_image :',x_img_train.shape)
print('train_label :',y_label_train.shape)
print('test_image :',x_img_test.shape)
print('test_label :',y_label_test.shape)

Output

Using TensorFlow backend.
Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
170500096/170498071 [==============================] - 3835s 22us/step
train: 50000
test : 10000
train_image : (50000, 32, 32, 3)
train_label : (50000, 1)
test_image : (10000, 32, 32, 3)
test_label : (10000, 1)

看看图片的像素

print(x_img_test[0])

Output

[[[158 112  49]
  [159 111  47]
  [165 116  51]
  ...
  [137  95  36]
  [126  91  36]
  [116  85  33]]
  
   ...
   
  [[ 54 107 160]
  [ 56 105 149]
  [ 45  89 132]
  ...
  [ 24  77 124]
  [ 34  84 129]
  [ 21  67 110]]]

32个32*3，一行是一个RGB

1.2 可视化部分训练集

label_dict={0:"airplane",1:"automobile",2:"bird",3:"cat",4:"deer",
            5:"dog",6:"frog",7:"horse",8:"ship",9:"truck"}
import matplotlib.pyplot as plt
def plot_images_labels_prediction(images,labels,prediction,idx,num=10):
    fig = plt.gcf()
    fig.set_size_inches(12, 14) # 控制图片大小
    if num>25: num=25  #最多显示25张
    for i in range(0, num):
        ax=plt.subplot(5,5, 1+i)
        ax.imshow(images[idx],cmap='binary')
        title=str(i)+','+label_dict[labels[i][0]]# i-th张图片对应的类别
        if len(prediction)>0:
            title+='=>'+label_dict[prediction[i]]
        ax.set_title(title,fontsize=10) 
        ax.set_xticks([]);
        ax.set_yticks([])        
        idx+=1 
    plt.savefig('1.png')
    plt.show()

Output
这里写图片描述

Note，0-9类

print(y_label_train[0])
print(y_label_train[0][0])

Output

[6]
6

1.3 Image normalize

print(x_img_train[0][0][0]) #（50000，32，32，3）
x_img_train_normalize = x_img_train.astype('float32') / 255.0
x_img_test_normalize = x_img_test.astype('float32') / 255.0
print(x_img_train_normalize[0][0][0])

Output

[59 62 63]
[0.23137255 0.24313726 0.24705882]

1.4 One-Hot Encoding

print(y_label_train.shape)
print(y_label_train[:5])

Output

(50000, 1)
[[6]
 [9]
 [9]
 [4]
 [1]]

One-Hot Encoding

from keras.utils import np_utils
y_label_train_OneHot = np_utils.to_categorical(y_label_train)
y_label_test_OneHot = np_utils.to_categorical(y_label_test)
print(y_label_train_OneHot.shape)
print(y_label_train_OneHot[:5])

Output

(50000, 10)
[[0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
 [0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]]

2 two convolutional layer network

2.1 Build Model

第一节的数据预处理流程合起来如下

from keras.datasets import cifar10
import numpy as np
np.random.seed(10)
# 载入数据集
(x_img_train,y_label_train),(x_img_test,y_label_test)=cifar10.load_data()
print("train data:",'images:',x_img_train.shape,
      " labels:",y_label_train.shape) 
print("test  data:",'images:',x_img_test.shape ,
      " labels:",y_label_test.shape) 
# 归一化
x_img_train_normalize = x_img_train.astype('float32') / 255.0
x_img_test_normalize = x_img_test.astype('float32') / 255.0
# One-Hot Encoding
from keras.utils import np_utils
y_label_train_OneHot = np_utils.to_categorical(y_label_train)
y_label_test_OneHot = np_utils.to_categorical(y_label_test)
y_label_test_OneHot.shape

Output

Using TensorFlow backend.
train data: images: (50000, 32, 32, 3)  labels: (50000, 1)
test  data: images: (10000, 32, 32, 3)  labels: (10000, 1)
(10000, 10)

建立模型 input - 卷积 - relu - drop out - 池化 - FC - drop out - output

from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv2D, MaxPooling2D, ZeroPadding2D
# 搭面包架子
model = Sequential()
# 加面包：卷积层1 和 池化层1
model.add(Conv2D(filters=32,kernel_size=(3,3),
                 input_shape=(32, 32,3), 
                 activation='relu', 
                 padding='same'))
model.add(Dropout(rate=0.25))
model.add(MaxPooling2D(pool_size=(2, 2))) # 16* 16
# 加面包：卷积层2 和 池化层2
model.add(Conv2D(filters=64, kernel_size=(3, 3), 
                 activation='relu', padding='same'))
model.add(Dropout(0.25))
model.add(MaxPooling2D(pool_size=(2, 2))) # 8 * 8
#Step3	建立神經網路(平坦層、隱藏層、輸出層)
model.add(Flatten()) # FC1,64个8*8转化为1维向量
model.add(Dropout(rate=0.25))
model.add(Dense(1024, activation='relu')) # FC2 1024
model.add(Dropout(rate=0.25))
model.add(Dense(10, activation='softmax')) # Output 10
print(model.summary())

Output

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 32, 32, 32)        896       
_________________________________________________________________
dropout_1 (Dropout)          (None, 32, 32, 32)        0         
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 16, 16, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 16, 16, 64)        18496     
_________________________________________________________________
dropout_2 (Dropout)          (None, 16, 16, 64)        0         
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 8, 8, 64)          0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 4096)              0         
_________________________________________________________________
dropout_3 (Dropout)          (None, 4096)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 1024)              4195328   
_________________________________________________________________
dropout_4 (Dropout)          (None, 1024)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 10)                10250     
=================================================================
Total params: 4,224,970
Trainable params: 4,224,970
Non-trainable params: 0
_________________________________________________________________
None

参数计算
con1 滤波器大小 3*3，输入3通道，32个滤波器，输出32通道
3 × 3 × 3 × 32 + 32 = 896

con2 滤波器大小 3*3，输入32通道，64个滤波器，输出64通道
32 × 3 × 3 × 64 + 64 = 18496

FC1
输入8864，输出 1024
8864*1024 + 1024 = 4195328

FC2（output layer）
输入
1024，输出10
1024*10+10 = 10250

2.2 Training process

训练前设置下需要使用的GPU号

import os  
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"   
os.environ["CUDA_VISIBLE_DEVICES"]="0"

0 表示0号GPU

model.compile(loss='categorical_crossentropy',
              optimizer='adam', metrics=['accuracy'])

train_history=model.fit(x_img_train_normalize, y_label_train_OneHot,
                        validation_split=0.2,
                        epochs=10, batch_size=128, verbose=1)

参数说明参考【Keras-MLP】MNIST 5.3 节

verbose：日志显示，0为不在标准输出流输出日志信息，1为输出进度条记录，2为每个epoch输出一行记录

Output

Train on 40000 samples, validate on 10000 samples
Epoch 1/10
40000/40000 [==============================] - 12s 290us/step - loss: 1.4855 - acc: 0.4641 - val_loss: 1.2901 - val_acc: 0.5711
Epoch 2/10
40000/40000 [==============================] - 5s 134us/step - loss: 1.1340 - acc: 0.5968 - val_loss: 1.1208 - val_acc: 0.6347
Epoch 3/10
40000/40000 [==============================] - 5s 127us/step - loss: 0.9778 - acc: 0.6561 - val_loss: 1.0021 - val_acc: 0.6681
Epoch 4/10
40000/40000 [==============================] - 5s 121us/step - loss: 0.8719 - acc: 0.6945 - val_loss: 0.9609 - val_acc: 0.6908
Epoch 5/10
40000/40000 [==============================] - 5s 133us/step - loss: 0.7807 - acc: 0.7255 - val_loss: 0.8714 - val_acc: 0.7039
Epoch 6/10
40000/40000 [==============================] - 5s 131us/step - loss: 0.6964 - acc: 0.7564 - val_loss: 0.8542 - val_acc: 0.7130
Epoch 7/10
40000/40000 [==============================] - 5s 127us/step - loss: 0.6222 - acc: 0.7831 - val_loss: 0.8217 - val_acc: 0.7238
Epoch 8/10
40000/40000 [==============================] - 5s 132us/step - loss: 0.5545 - acc: 0.8055 - val_loss: 0.8121 - val_acc: 0.7255
Epoch 9/10
40000/40000 [==============================] - 5s 136us/step - loss: 0.4914 - acc: 0.8285 - val_loss: 0.7473 - val_acc: 0.7467
Epoch 10/10
40000/40000 [==============================] - 5s 125us/step - loss: 0.4284 - acc: 0.8515 - val_loss: 0.7794 - val_acc: 0.7354

2.3 可视化训练过程

import matplotlib.pyplot as plt
def show_train_history(train_acc,test_acc):
    plt.plot(train_history.history[train_acc])
    plt.plot(train_history.history[test_acc])
    plt.title('Train History')
    plt.ylabel('Accuracy')
    plt.xlabel('Epoch')
    plt.legend(['train', 'test'], loc='upper left')
    plt.savefig("1.png")
    plt.show()

Output

show_train_history('acc','val_acc')

这里写图片描述

show_train_history('loss','val_loss')

这里写图片描述

2.4 模型评估

print(model.metrics_names)
scores = model.evaluate(x_img_test_normalize, 
                        y_label_test_OneHot, verbose=0)
print(scores)
print(scores[1])

Output

['loss', 'acc']
[0.7902285717010498, 0.7278]
0.7278

2.5 预测结果

前10个为例，predict_classes（预测类别，也即是softmax输出的最大值）

prediction=model.predict_classes(x_img_test_normalize)
prediction[:10]

Output

array([3, 8, 8, 0, 6, 6, 1, 6, 3, 1])

可视化部分预测结果

label_dict={0:"airplane",1:"automobile",2:"bird",3:"cat",4:"deer",
            5:"dog",6:"frog",7:"horse",8:"ship",9:"truck"}
import matplotlib.pyplot as plt
def plot_images_labels_prediction(images,labels,prediction,
                                  idx,num=10):
    fig = plt.gcf()
    fig.set_size_inches(12, 14)
    if num>25: num=25 
    for i in range(0, num):
        ax=plt.subplot(5,5, 1+i)
        ax.imshow(images[idx],cmap='binary')
                
        title=str(i)+','+label_dict[labels[i][0]]
        if len(prediction)>0:
            title+='=>'+label_dict[prediction[i]]
            
        ax.set_title(title,fontsize=10) 
        ax.set_xticks([]);ax.set_yticks([])        
        idx+=1 
    plt.show()

调用

plot_images_labels_prediction(x_img_test,y_label_test,prediction,0,10)

Output
这里写图片描述

2.6 查看预测概率

model.predict，结果为每个样本的 softmax 输出的10个值，预测结果为10个值中最大值对应的那个类别

Predicted_Probability = model.predict(x_img_test_normalize)

label_dict={0:"airplane",1:"automobile",2:"bird",3:"cat",4:"deer",
            5:"dog",6:"frog",7:"horse",8:"ship",9:"truck"}
            
def show_Predicted_Probability(y,prediction,x_img,Predicted_Probability,i):
    print('label:',label_dict[y[i][0]],'predict:',label_dict[prediction[i]])
    plt.figure(figsize=(2,2))
    plt.imshow(np.reshape(x_img_test[i],(32, 32,3)))
    plt.show()
    for j in range(10):
        print(label_dict[j]+' Probability:%1.9f'%(Predicted_Probability[i][j]))

调用

show_Predicted_Probability(y_label_test,prediction,x_img_test,Predicted_Probability,0)

Output
这里写图片描述

调用

show_Predicted_Probability(y_label_test,prediction,x_img_test,Predicted_Probability,3)

Output
这里写图片描述

2.7 显示混淆矩阵

print(y_label_test)
print(y_label_test.reshape(-1))# 转化为1维数组

Output

[[3]
 [8]
 [8]
 ...
 [5]
 [1]
 [7]]
[3 8 8 ... 5 1 7]

混淆矩阵

import pandas as pd
print(label_dict)
pd.crosstab(y_label_test.reshape(-1),prediction,rownames=['label'],colnames=['predict'])

这里写图片描述

分析：
对角线
frog 预测最准确，856/1000，说明最不容易被混淆
cat 预测最不准，458/1000，说明最容易被混淆

非对角线
把 cat 预测成 dog 最多，190
把 dog 预测成 cat 第三多，108 ，猫狗很容易混淆

CIFAR-10大致分为两类

动物类：2，3，4，5，6，7
交通工具类：0，1，8，9

表中第3列，第11列可以看出
这里写图片描述

动物很少被错分为 car 和 track

3 six convolutional layer network

第二节建立的神经网络准确率才0.7278，不太好，我们希望能通过更多次的卷积运算提高准确率

3.1 data preprocessing

# 选GPU
import os  
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"   
os.environ["CUDA_VISIBLE_DEVICES"]="0"  
# Simple CNN model for the CIFAR-10 Dataset

import numpy
from keras.datasets import cifar10
import numpy as np
np.random.seed(10)
#载入数据集
(X_img_train, y_label_train), (X_img_test, y_label_test) = cifar10.load_data()
print("train data:",'images:',X_img_train.shape," labels:",y_label_train.shape) 
print("test  data:",'images:',X_img_test.shape ," labels:",y_label_test.shape) 

#标准化
X_img_train_normalize = X_img_train.astype('float32') / 255.0
X_img_test_normalize = X_img_test.astype('float32') / 255.0

# One-Hot Encoding
from keras.utils import np_utils
y_label_train_OneHot = np_utils.to_categorical(y_label_train)
y_label_test_OneHot = np_utils.to_categorical(y_label_test)

Output

Using TensorFlow backend.
train data: images: (50000, 32, 32, 3)  labels: (50000, 1)
test  data: images: (10000, 32, 32, 3)  labels: (10000, 1)

y_label_test_OneHot.shape

Output

(10000, 10)

3.2 building model

from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv2D, MaxPooling2D, ZeroPadding2D
#蛋糕架子
model = Sequential()
#卷積層1與池化層1，开始加蛋糕
model.add(Conv2D(filters=32,kernel_size=(3, 3),input_shape=(32, 32,3),
                 activation='relu', padding='same'))
model.add(Dropout(0.3))
model.add(Conv2D(filters=32, kernel_size=(3, 3),
                 activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#卷積層2與池化層2
model.add(Conv2D(filters=64, kernel_size=(3, 3),
                 activation='relu', padding='same'))
model.add(Dropout(0.3))
model.add(Conv2D(filters=64, kernel_size=(3, 3), 
                 activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#卷積層3與池化層3
model.add(Conv2D(filters=128, kernel_size=(3, 3), 
                 activation='relu', padding='same'))
model.add(Dropout(0.3))
model.add(Conv2D(filters=128, kernel_size=(3, 3), 
                 activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#建立神經網路(平坦層、隱藏層、輸出層)
model.add(Flatten())
model.add(Dropout(0.3))
model.add(Dense(2500, activation='relu'))
model.add(Dropout(0.3))
model.add(Dense(1500, activation='relu'))
model.add(Dropout(0.3))
model.add(Dense(10, activation='softmax'))
print(model.summary())

Output

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 32, 32, 32)        896       
_________________________________________________________________
dropout_1 (Dropout)          (None, 32, 32, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 32, 32, 32)        9248      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 16, 16, 32)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 16, 16, 64)        18496     
_________________________________________________________________
dropout_2 (Dropout)          (None, 16, 16, 64)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 16, 16, 64)        36928     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 8, 8, 64)          0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 8, 8, 128)         73856     
_________________________________________________________________
dropout_3 (Dropout)          (None, 8, 8, 128)         0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 8, 8, 128)         147584    
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 4, 4, 128)         0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 2048)              0         
_________________________________________________________________
dropout_4 (Dropout)          (None, 2048)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 2500)              5122500   
_________________________________________________________________
dropout_5 (Dropout)          (None, 2500)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 1500)              3751500   
_________________________________________________________________
dropout_6 (Dropout)          (None, 1500)              0         
_________________________________________________________________
dense_3 (Dense)              (None, 10)                15010     
=================================================================
Total params: 9,176,018
Trainable params: 9,176,018
Non-trainable params: 0
_________________________________________________________________
None

输入
conv - drop out - conv - maxpooling（3次）

fc - drop out（2次）
输出

参数计算
con1 滤波器大小 3*3，输入3通道，32个滤波器，输出32通道
3 × 3 × 3 × 32 + 32 = 896

con2 滤波器大小 3*3，输入32通道，32个滤波器，输出32通道
32 × 3 × 3 × 32 + 32 = 9248

con3 滤波器大小 3*3，输入32通道，64个滤波器，输出64通道
32 × 3 × 3 × 64 + 64 = 18496

con4 滤波器大小 3*3，输入64通道，64个滤波器，输出64通道
64 × 3 × 3 × 64 + 64 = 36928

con5 滤波器大小 3*3，输入64通道，128个滤波器，输出128通道
64 × 3 × 3 × 128 + 128 = 73856

con6 滤波器大小 3*3，输入128通道，128个滤波器，输出128通道
128 × 3 × 3 × 128 + 128 = 147584

FC1

输入 $4 * 4 * 128 = 2048$ ，输出 2500
2048* 2048 + 2500 = 5122500

FC2
输入 2500，输出1500
2500*1500 + 1500 = 3751500

FC3（output layer）
输入1500，输出10
1500*10+10 = 15010

3.3 training process

model.compile(loss='categorical_crossentropy', optimizer='adam',
              metrics=['accuracy'])
              
train_history=model.fit(X_img_train_normalize, y_label_train_OneHot,
                        validation_split=0.2,
                        epochs=50, batch_size=300, verbose=1)

参数说明参考【Keras-MLP】MNIST 5.3 节
或者 Keras中文文档

verbose：日志显示，0为不在标准输出流输出日志信息，1为输出进度条记录，2为每个epoch输出一行记录

epochs = 50，训练时间会比较长，我用一张TITAN xp训练，一个epochs要5-6s

Output（截取部分）

Train on 40000 samples, validate on 10000 samples
Epoch 1/50
40000/40000 [==============================] - 6s 140us/step - loss: 0.6132 - acc: 0.7841 - val_loss: 0.6855 - val_acc: 0.7628
Epoch 2/50
40000/40000 [==============================] - 5s 130us/step - loss: 0.5631 - acc: 0.8018 - val_loss: 0.6670 - val_acc: 0.7720
Epoch 3/50
40000/40000 [==============================] - 5s 126us/step - loss: 0.5222 - acc: 0.8134 - val_loss: 0.6741 - val_acc: 0.7752
Epoch 4/50
40000/40000 [==============================] - 5s 136us/step - loss: 0.4852 - acc: 0.8264 - val_loss: 0.6401 - val_acc: 0.7815
Epoch 5/50
40000/40000 [==============================] - 5s 130us/step - loss: 0.4547 - acc: 0.8364 - val_loss: 0.6573 - val_acc: 0.7816
………………
………………
………………
Epoch 45/50
40000/40000 [==============================] - 5s 131us/step - loss: 0.1225 - acc: 0.9586 - val_loss: 0.7927 - val_acc: 0.7978
Epoch 46/50
40000/40000 [==============================] - 5s 133us/step - loss: 0.1185 - acc: 0.9609 - val_loss: 0.7982 - val_acc: 0.8025
Epoch 47/50
40000/40000 [==============================] - 6s 140us/step - loss: 0.1280 - acc: 0.9565 - val_loss: 0.8781 - val_acc: 0.7910
Epoch 48/50
40000/40000 [==============================] - 5s 132us/step - loss: 0.1241 - acc: 0.9581 - val_loss: 0.7653 - val_acc: 0.8053
Epoch 49/50
40000/40000 [==============================] - 5s 134us/step - loss: 0.1256 - acc: 0.9562 - val_loss: 0.7894 - val_acc: 0.7962
Epoch 50/50
40000/40000 [==============================] - 5s 136us/step - loss: 0.1190 - acc: 0.9596 - val_loss: 0.8189 - val_acc: 0.8019

用第二节的show_train_history函数，查看下训练过程

import matplotlib.pyplot as plt
def show_train_history(train_acc,test_acc):
    plt.plot(train_history.history[train_acc])
    plt.plot(train_history.history[test_acc])
    plt.title('Train History')
    plt.ylabel('Accuracy')
    plt.xlabel('Epoch')
    plt.legend(['train', 'test'], loc='upper left')
    plt.show()

调用

show_train_history('acc','val_acc')

Output
这里写图片描述

show_train_history('loss','val_loss')

这里写图片描述

哈哈。过拟合了，说明网络设计的相对复杂。自己去DIY吧

3.4 模型的评估

scores = model.evaluate(X_img_test_normalize, 
                        y_label_test_OneHot,verbose=0)
scores[1]

结果

0.7882

4 模型的保存

第三节中，程序训练需要花费很长时间，跑更复杂的模型和更大的图片数据集时候，往往需要十几个小时或者数天的时间，有时候可能因为某些原因导致计算机宕机，这样之前的训练就前功尽弃了，解决的方法是，每次程序执行完训练后，将模型保存一下。下次程序训练之前，先加载模型权重，再继续训练。

4.1 h5

build model以后，载入与训练模型，第一次没有保存与训练模型肯定报错

try:
    model.load_weights("SaveModel/cifarCnnModel.h5")
    print("载入模型成果！继续训练模型")
except :    
    print("载入模型失败！开始训练一个新模型")

Output

载入模型失败！开始训练一个新模型

开始训练

model.compile(loss='categorical_crossentropy',
              optimizer='adam', metrics=['accuracy'])
train_history=model.fit(x_img_train_normalize, y_label_train_OneHot,
                        validation_split=0.2,
                        epochs=5, batch_size=128, verbose=1)

保存模型，在当前目录下，新建一个文件夹，然后把模型保存进去

import os
os.mkdir("SaveModel")
model.save_weights("SaveModel/cifarCnnModel.h5")
print("Saved model to disk")

Output

Saved model to disk

这样就把训练模型保存为 h5 格式了，当然也可以保存为其它格式，例如 json 和 yaml 格式

下次训练的时候，就可以先载入模型的权重了

try:
    model.load_weights("SaveModel/cifarCnnModel.h5")
    print("载入模型成果！继续训练模型")
except :    
    print("载入模型失败！开始训练一个新模型")

Output

载入模型成果！继续训练模型

记得训练完了要保存哟

model.save_weights("SaveModel/cifarCnnModel.h5")
print("Saved model to disk")

4.2 json

model_json = model.to_json()
with open("SaveModel/cifarCnnModelnew.json", "w") as json_file:
    json_file.write(model_json)

4.3 yaml

model_yaml = model.to_yaml()
with open("SaveModel/cifarCnnModelnew.yaml", "w") as yaml_file:
    yaml_file.write(model_yaml)

声明

代码源于《TensorFlow+Keras深度学习人工智能实践应用》林大贵版，引用、转载请注明出处，谢谢，如果对书本感兴趣，买一本看看吧！！！

CIFAR-10官网. ↩︎

bryant_meng

关注

15
点赞
踩
80

收藏

觉得还不错? 一键收藏
7
评论
【Keras-CNN】CIFAR-10

参考声明CIFAR-10是由Alex Krizhevsky、Vinod Nair 与 Geoffrey Hinton 收集的一个用于图像识别的数据集，有10类，飞机、汽车、鸟、猫、鹿、狗、青蛙、马、船、卡车。与MNIST相比，色彩、颜色噪点较多，同一类物体大小不一、角度不同、颜色不同。参考【1】CIFAR-10官网声明声明：代码源于《Te...
复制链接

扫一扫