Convolutional Neural Network (CNN) mnist Digit Recognition - Tensorflow

大家好,我是老表～最近参与了 K同学啊 举行的 深度学习训练营,第一周任务：卷积神经网络（CNN）mnist数字识别,Here I share my own learning and implementation process,and summarizing reflections on the part of your doubts,希望对大家有所帮助.

使用环境：

• K80显卡
• Tensorflow 2.8
• opencv 4.6
  

I knew very little about deep learning before,So the goal I set for myself is：

• 跑通代码
• understand the code as much as possible
• Can use the trained model to predict their written words

导入相关包

# 导入包
import tensorflow as tf
from tensorflow.keras import datasets, layers, models
import matplotlib.pyplot as plt

设置GPU

# 设置 gpu
gpus = tf.config.list_physical_devices("GPU")
print(gpus)
if gpus:
    gpu0 = gpus[0] # 如果有多个GPU,仅使用第0个GPU
    tf.config.experimental.set_memory_growth(gpu0, True) # 设置GPU显存用量按需使用
    tf.config.set_visible_devices([gpu0],"GPU")

输出：

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

加载数据

# 加载数据集
(train_images, train_labels), (test_images, test_labels) = datasets.mnist.load_data()
# can also be loaded locally
# (train_images, train_labels), (test_images, test_labels) = datasets.mnist.load_data(path='/mnt/mnist.npz')
# 查看数据规模
train_images.shape, test_images.shape, train_labels.shape, test_labels.shape

输出：

((60000, 28, 28), (10000, 28, 28), (60000,), (10000,))

• 查看数据：

# To see the data
train_images[0][10]

输出：

array([  0,   0,   0,   0,   0,   0,   0,   0,   0,  14,   1, 154, 253,
        90,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,
         0,   0], dtype=uint8)

数据归一化

看了个b站upMaster's Video Introduction,It feels very clear,What I understand is that data normalization avoids a single feature having a big impact on the final result.

There are two ways to normalize：线性函数归一化（maximization）和零均值归一化（标准化）.

线性函数归一化（maximization）It is suitable for the situation with obvious edge,For example, the pixel in this case（0-255）,人的身高等.

零均值归一化（标准化）Suitable for data with outliers and more noise.

# 数据归一化 
# What is data normalization？作用？
# Normalize the value of the pixel to0到1的区间内
train_images, test_images = train_images / 255.0, test_images / 255.0
# max：255 min：0
# 归一化公式 (x-min)/(max-min)

• View normalized data

# View normalized data
train_images[0][10]

输出：

array([0.        , 0.        , 0.        , 0.        , 0.        ,
       0.        , 0.        , 0.        , 0.        , 0.05490196,
       0.00392157, 0.60392157, 0.99215686, 0.35294118, 0.        ,
       0.        , 0.        , 0.        , 0.        , 0.        ,
       0.        , 0.        , 0.        , 0.        , 0.        ,
       0.        , 0.        , 0.        ])

查看训练集数据

plt.figure(figsize=(20,10))
for i in range(20):
    plt.subplot(5,10,i+1)
    plt.xticks([])
    plt.yticks([])
    plt.grid(False)
    # cmap 参数作用？Set the image gradient
    plt.imshow(train_images[i], cmap=plt.cm.binary)
    plt.xlabel(train_labels[i])
plt.show()

调整数据 shape

# Adjust the data to the format we need
# Why adjust data？
train_images = train_images.reshape((60000, 28, 28, 1))
test_images = test_images.reshape((10000, 28, 28, 1))

train_images.shape,test_images.shape,train_labels.shape,test_labels.shape

输出：

((60000, 28, 28, 1), (10000, 28, 28, 1), (60000,), (10000,))

• To check the adjusted data

# To check the adjusted data
train_images[0][10]

输出：

array([[0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.05490196],
       [0.00392157],
       [0.60392157],
       [0.99215686],
       [0.35294118],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ],
       [0.        ]])

构建CNN网络

# 构建CNN网络模型
# every step？什么是cnn?
model = models.Sequential([
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)),
    layers.MaxPooling2D((2, 2)),
    layers.Conv2D(64, (3, 3), activation='relu'),
    layers.MaxPooling2D((2, 2)),
    
    layers.Flatten(),
    layers.Dense(64, activation='relu'),
    layers.Dense(10)
])

model.summary()

编译模型

# 编译模型
# optimizer loss metrics 含义和作用？
model.compile(optimizer ='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

训练模型

# 训练模型
history = model.fit(train_images, train_labels, epochs=10, 
                   validation_data=(test_images, test_labels))

预测

# Read a random image from the test set,并显示
plt.imshow(test_images[0])

• 预测测试集数据

# 使用训练好的 model 进行预测
pre = model.predict(test_images)
# 查看预测结果
pre[0]

输出：

array([-16.13687  ,  -1.1026648,  -3.5199716,  -4.108456 ,  -2.1431744,
       -11.160188 , -12.815703 ,  17.407137 , -13.253326 ,  -1.5100265],
      dtype=float32)

这里返回的是一个 array 数组,The value of each element corresponds to the number on the predicted picture is0-9的“概率”（and the element subscript exactly corresponds to the upper）,如上 数组最大值是17.407137,The corresponding subscript of the element is7,So according to the prediction result, the maximum number on the picture is predicted to be7.

• Write a function to convert predictions to numbers

# Write a function to convert predictions to numbers
import numpy as np

def get_pre_data(pre_array):
    return np.argmax(pre_array, axis=0)

调用函数：

get_pre_data(pre[0])

# 输出：7

• Encapsulate the prediction function Predict a single image in the test set

通过前面,我们不难看出,model.predict(test_images)is to predict all the test set images at once,In reality, we definitely want to predict one by one.,If violence changes the parameter totest_images[0]You will encounter the following error.

ValueError: Exception encountered when calling layer "conv2d" (type Conv2D).
    
    Negative dimension size caused by subtracting 3 from 1 for '{
    {node sequential/conv2d/Conv2D}} = Conv2D[T=DT_FLOAT, data_format="NHWC", dilations=[1, 1, 1, 1], explicit_paddings=[], padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](sequential/ExpandDims, sequential/conv2d/Conv2D/ReadVariableOp)' with input shapes: [?,28,1,1], [3,3,1,32].
    
    Call arguments received:
      • inputs=tf.Tensor(shape=(None, 28, 1, 1), dtype=float32)

大概意思是在conv2d中图像 shape 出现了问题,The reason is because the data passed into the modelshape应该为x, 28, 28, 1,而test_images[0]的 shape 是 28, 28, 1,知道原因,改起来就简单了,具体代码如下.

# Encapsulate the prediction function Predict a single image in the test set
def pre_image(image):
    # 传入 image shape 是 (28, 28, 1) needs to be transformed into (1, 28, 28, 1)
    image = image.reshape((1, 28, 28, 1))
    # 使用训练好的 model 进行预测 returns a two-digit
    pre_array = model.predict(image)[0]
    # 获取预测结果
    return get_pre_data(pre_array)

调用函数：

pre_image(test_images[0])

# 输出：7

• Predict the words we write ourselves

First you need to save the model.

# 模型保存
model.save('/mnt/tf_data/mnist_model.h5')

Then write a number by yourself（Write a number on toilet paper2）：

Then write a model prediction function,大概步骤就是：加载模型-读取图片-图片处理-预测.

Here we should pay special attention to the following two points：

• We do before the training data normalization,Then the predicted data must also be normalized accordingly.
• When reading data, you need to specify cv2.IMREAD_GRAYSCALE,Otherwise, the default reading is three channels（Data processing is troublesome）

代码如下：

# To predict their written words
import tensorflow as tf
import cv2

def pre_image(image_path, model):
    # 读取本地图片
    # Reading pictures in grayscale,默认是三通道（Data processing is troublesome）
    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    # print('图片大小', img.shape)
    # 转变成 28,28
    img = cv2.resize(img,(28,28))
    # print('图片大小', img.shape)
    # 显示图片 
    plt.imshow(img)
    # 数据归一化 （How to normalize the model before, how to normalize now）
    img = img/255
    # 将图片 shape 改成 (1,28,28,1) Also data type last and training set/测试集一样
    img = img.reshape((1,28,28,1)).astype('float64')
    # 使用训练好的 model 进行预测 returns a two-digit
    pre_array = model.predict(img)[0]
    # 获取预测结果
    return get_pre_data(pre_array)

# 加载模型
model = tf.keras.models.load_model('/mnt/tf_data/mnist_model.h5')
image_path = '/mnt/tf_data/2.jpeg'
# 预测
print('预测结果：', pre_image(image_path, model))

总结

The project ran smoothly,And solved some of my doubts by consulting the information,Finally, I also use the model I trained myself,Predicted the value written by yourself,感觉很棒.

However, there are still many problems unresolved,比如：Understand the specific role of normalization？How to affect the convergence efficiency of gradient descent when solving？Why should the data be transformed into(60000, 28, 28, 1)？CNNHow the network model works？what does the compiled model do？等.

I will find time to continue my study later,Looking forward to learning more with you.

参考

• 卷积神经网络（CNN）mnistNumber Recognition Tutorial https://mtyjkh.blog.csdn.net/article/details/116920825

• Why do you need to normalize numeric features?？ https://b23.tv/RnN2VoQ

• Why do we need to normalize data in machine learning?？ https://cloud.tencent.com/developer/article/1456997