Keras Deep Learning - Building a CNN Model to Recognize MNIST Handwritten Digits

Continue to create, accelerate growth! This is the 5th day of my participation in the "Nuggets Daily New Plan · June Update Challenge", click to view the event details

Task and Model Analysis

In Convolutional Neural Networks ( Convolutional Neural Network, CNN) Fundamentals, we learned about the problems CNNof how . In this section, we will build a CNNmodel to recognize MNISThandwritten digits. We used the following strategies to build the CNNmodel :

The input shape is 28 x 28 x 1, the size of the convolution kernel used is3 x 3 x 1
- It should be noted that the size of the convolution kernel can be changed, but the number of channels cannot be changed, it needs to be the same as the number of input channels
- 10Use convolution kernels
After the input image goes through the convolutional layer, use the pooling layer:
- The output image size is halved
- Output obtained after flattening pooling
The flattened layer is connected to a hidden layer 1000with units
Finally, connect the hidden layer to the output layer, which has 10classes (including numbers 0-9) in the output layer

After building the model, we generate 1an , translate 1by pixels, and test the performance of the CNNmodel ; in 1Section , we have seen that a fully connected neural network cannot strive to predict this mean The category of the image.

CNN model construction and training

Next, Kerasimplement the CNNschema defined above using the implementation to learn MNISThow to use the CNNmodel on the data.

Load and preprocess data:

from keras.layers import Dense,Conv2D,MaxPool2D,Flatten
from keras.models import Sequential
from keras.datasets import mnist
from keras.utils import np_utils
import numpy as np

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train = x_train.reshape(x_train.shape[0], x_train.shape[1], x_train.shape[1], 1)
x_test = x_test.reshape(x_test.shape[0], x_test.shape[1], x_test.shape[1], 1)
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
y_train = np_utils.to_categorical(y_train)
y_test = np_utils.to_categorical(y_test)
num_classes = y_test.shape[1]
复制代码

The preprocessing steps are exactly the same as we used in Building Deep Feedforward Neural Networks .

Build and compile the model:

model = Sequential()
model.add(Conv2D(10, (3, 3), input_shape=(28, 28, 1), activation='relu'))
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(Dense(num_classes, activation='softmax'))
model.summary()

model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['acc'])
复制代码

Brief schema information for the model we initialized in the preceding code can be obtained:

model.summary()
复制代码

The brief schema information of the output model is as follows:

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 26, 26, 10)        100       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 13, 13, 10)        0         
_________________________________________________________________
flatten (Flatten)            (None, 1690)              0         
_________________________________________________________________
dense (Dense)                (None, 512)               865792    
_________________________________________________________________
dense_1 (Dense)              (None, 10)                5130      
=================================================================
Total params: 871,022
Trainable params: 871,022
Non-trainable params: 0
_________________________________________________________________
复制代码

There are 100a parameters in the convolutional layer, because there are10 convolution kernels in the convolutional layer, so there are a total of weight parameters and bias terms (in each convolution kernel ), a total of parameters. The max pooling layer does not have any parameters as it only needs to compute the maximum value in pooling kernel of size . It can be seen that using the model can greatly reduce the amount of network parameters.3 x 3 x 1901011002 x 2CNN

Finally, fit the model:

model.fit(x_train, y_train,
            validation_data=(x_test, y_test),
            epochs=10,
            batch_size=1024,
            verbose=1)
复制代码

The above 10model epochcan achieve 98％an :

Model training process monitoring

Next, generate 1an and shifted 1by pixels:

# 获取标签为1的所有图像输入
x_test1 = x_test[y_test[:, 1]==1]
# 利用所有标签为1的图像生成均值图像
pic = np.zeros((x_test.shape[1], x_test.shape[2]))
pic2 = np.copy(pic)
for i in range(x_test1.shape[0]):
    pic2 = x_test1[i, :, :, 0]
pic = pic + pic2
pic = (pic / x_test1.shape[0])
# 将均值图像中的每个像素向左平移一个像素
for i in range(pic.shape[0]):
    if i < 21:
        pic[:, i] = pic[:, i+1]
# 对平移后的图像进行预测
p = model.predict(pic.reshape(1, x_test.shape[1], x_test.shape[2], 1))
print(p)
c = np.argmax(p)
print('CNN预测结果：', c)
复制代码

The resulting model output is as follows:

[[1.3430370e-05 9.9989212e-01 2.0535077e-05 2.6301734e-07 4.3278211e-05
  5.9122913e-06 1.5874346e-05 6.2533190e-06 2.0079199e-06 4.1732173e-07]]
CNN预测结果： 1
复制代码

Looking at it, we can see that the predictions obtained using the CNNarchitecture 1.