利用卷积神经网络识别手写数字

1.测试数据准备

1.我们使用的测试数据，可以直接从keras.datasets.mnist导入

import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
plt.rcParams['figure.figsize']=(7,7)

from keras.datasets import mnist
from keras.models import Sequential
from keras.utils import np_utils
import keras
from keras.preprocessing.image import ImageDataGenerator
from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPool2D
from keras.layers.normalization import BatchNormalization

epochs = 20
input_shape = (28, 28, 1)
nb_classes = 10

(X_train, y_train), (X_test, y_test) = mnist.load_data()
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)

能看到一共有6W条训练数据和1W条测试数据

由于被墙的原因，mnist可能无法下载，可以换成本地数据导入的方式。下载数据请点击这里，数据导入方式换成

path='./mnist.npz'
f = np.load(path)
X_train, y_train = f['x_train'], f['y_train']
X_test, y_test = f['x_test'], f['y_test']
f.close()

2.查看数据

for i in range(9):
    plt.subplot(3, 3, i+1)
    plt.imshow(X_train[i], cmap='gray', interpolation='none')
    plt.title("Class {}".format(y_train[i]))

3.将数据格式转换为keras接受的数据结构

X_train = X_train.reshape(X_train.shape[0], 28, 28, 1)
X_test = X_test.reshape(X_test.shape[0], 28, 28, 1)
print('X_train shape:', X_train.shape)
print('X_test shape:', X_test.shape)

使用seaborn观察训练数据集的y值分布

sns.countplot(y_train)

Y_train = np_utils.to_categorical(y_train, nb_classes)
Y_test = np_utils.to_categorical(y_test, nb_classes)
print("Y_train:", Y_train.shape)

2.定义深度神经网络结构

model = Sequential()
model.add(Conv2D(32, kernel_size=(3,3), activation='relu', kernel_initializer='he_normal', input_shape=input_shape))
model.add(Conv2D(32, kernel_size=(3,3), activation='relu', kernel_initializer='he_normal'))
model.add(MaxPool2D(2,2))
model.add(Dropout(0.2))
model.add(Conv2D(64, kernel_size=(3,3), activation='relu', kernel_initializer='he_normal', padding='same'))
model.add(Conv2D(64, kernel_size=(3,3), activation='relu', kernel_initializer='he_normal', padding='same'))
model.add(MaxPool2D(2,2))
model.add(Dropout(0.25))
model.add(Conv2D(128, kernel_size=(3,3), activation='relu', kernel_initializer='he_normal', padding='same'))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.25))
model.add(Dense(nb_classes, activation='softmax'))

model.compile(loss=keras.losses.categorical_crossentropy,
             optimizer=keras.optimizers.Adam(),
             metrics=['accuracy'])
model.summary()

3.训练模型

完成数据训练并保存训练好的模型

datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)

datagen.fit(X_train)

filepath = 'model.hdf5'
from keras.callbacks import ModelCheckpoint

# monitor计算每一个模型validation_data的准确率
# save_best_only 只保存最好的一个模型
checkpointer = ModelCheckpoint(filepath, monitor='val_acc', save_best_only=True, mode='max')

# steps_per_epoch指定循环次数
h = model.fit_generator(datagen.flow(X_train, Y_train, batch_size=500),
                       steps_per_epoch=len(X_train)/1000, epochs=epochs,
                       validation_data=datagen.flow(X_test, Y_test, batch_size=len(X_test)),
                       validation_steps=1, callbacks=[checkpointer])

训练完成后可以看到模型在测试数据上的准确率已经达到了99.39%

我们可以用pyplot打印一下，随着epochs的执行，模型在训练数据和测试数据集的准确率的表现情况

history = h.history
accuracy = history['acc']
val_accuracy = history['val_acc']
loss = history['loss']
val_loss = history['val_loss']
epochs = range(len(accuracy))
plt.plot(epochs, accuracy, 'bo', label='Training Accuracy')
plt.plot(epochs, val_accuracy, 'b', label='Validation Accuracy')
plt.title('Training and validation accuracy')
plt.legend()
plt.show()

4.模型评估

from keras.models import load_model
model = load_model('model.hdf5')
score = model.evaluate(X_test, Y_test)
print('Test score:', score[0])
print('Test accuracy:', score[1])

可以看到一共有1W组数据，损失函数约为0.24，准确率达到了0.984，已经比较高了。

我们可以抽取一些预测数据，与真实值进行对比

predicted_classes = model.predict_classes(X_test)
correct_indices = np.nonzero(predicted_classes == y_test)[0]
incorrect_indices = np.nonzero(predicted_classes != y_test)[0]

选取9个预测正确的进行打印

plt.figure()
for i, correct in enumerate(correct_indices[:9]):
    plt.subplot(3,3,i+1)
    plt.imshow(X_test[correct].reshape(28, 28), cmap='gray', interpolation='none')
    plt.title("Predicted {}, Class {}".format(predicted_classes[correct], y_test[correct]))

当然我们也可以选择9个错误的进行打印

plt.figure()
for i, incorrect in enumerate(incorrect_indices[:9]):
    plt.subplot(3,3,i+1)
    plt.imshow(X_test[incorrect].reshape(28, 28), cmap='gray', interpolation='none')
    plt.title("Predicted {}, Class {}".format(predicted_classes[incorrect], y_test[incorrect]))

可以看到，有一些真实值为1的图片，被识别错误了，说明模型还有一些改进的空间。

所有源码、测试数据以及训练好的模型已上传至git，点击这里可直接查看，有疑问的同学请在博客下方留言或提Issuees，谢谢！