机器学习15卷积神经网络处理手写数字图片

 

1.手写数字数据集

• from sklearn.datasets import load_digits
• digits = load_digits()

from sklearn.datasets import load_digits
digits = load_digits()

 

2.图片数据预处理

• x：归一化MinMaxScaler()
• y：独热编码OneHotEncoder()或to_categorical
• 训练集测试集划分
• 张量结构

# 2、图片预处理
from sklearn.preprocessing import MinMaxScaler
from sklearn.preprocessing import OneHotEncoder
from sklearn.model_selection import train_test_split
#将x转化为浮点用于归一化
x = digits.data.astype(np.float32)
# 将Y_data变为一列用于onehot
y= digits.target.astype(np.float32).reshape(-1, 1)

# 将属性缩放到一个指定的最大和最小值（通常是1-0之间）
# x：归一化MinMaxScaler()
scaler = MinMaxScaler()
x = scaler.fit_transform(x)
print("归一化处理后的x：")
print(x)

# y：独热编码OneHotEncoder 张量结构todense
# 进行oe-hot编码
y = OneHotEncoder().fit_transform(y).todense()
print("one hot处理后的期望值：")
print(y)

#转换为图片的格式（batch, height, width, channels）
x = x.reshape(-1, 8, 8, 1)

# 训练集测试集划分
X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=0, stratify=y)
print('X_train.shape, X_test.shape, y_train.shape, y_test.shape：', X_train.shape, X_test.shape, y_train.shape, y_test.shape)

 

 

3.设计卷积神经网络结构

• 绘制模型结构图，并说明设计依据。

依据经典模型，在根据图片维数8乘8，选择4层卷积，3层池化。 为防止过拟合在其中加入Dropout层

 

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import MaxPool2D, Dropout, Flatten, Dense
model = Sequential()
ks = (3, 3)
# x
input_shape = x.shape[1:]
# 一层卷积，padding='same',tensorflow会对输入自动补0
model.add(Conv2D(filters=16, kernel_size=ks, padding='same', input_shape=input_shape, activation='relu'))
# 池化层1
model.add(MaxPool2D(pool_size=(2, 2)))
# 防止过拟合，随机丢掉连接
model.add(Dropout(0.25))
# 二层卷积
model.add(Conv2D(filters=32, kernel_size=ks, padding='same', activation='relu'))
# 池化层2
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
# 三层卷积
model.add(Conv2D(filters=64, kernel_size=ks, padding='same', activation='relu'))
# 四层卷积
model.add(Conv2D(filters=128, kernel_size=ks, padding='same', activation='relu'))
# 池化层3
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
# 平坦层
model.add(Flatten())
# 全连接层
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.25))
# 激活函数softmax
model.add(Dense(10, activation='softmax'))
print(model.summary())

 

 

 

 

 

 

 

4.模型训练

# 训练模型
# 损失函数:categorical_crossentropy,优化器:adam ,用准确率accuracy衡量模型
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
# 划分20%作为验证数据,每次训练300个数据,训练迭代150轮
train_history = model.fit(x=X_train, y=y_train, validation_split=0.2, batch_size=300, epochs=150, verbose=2)

 

import matplotlib.pyplot as plt
#自定义绘制函数
def show_train_history(train_history, train, validation):
    plt.plot(train_history.history[train])
    plt.plot(train_history.history[validation])
    plt.title('Train History')
    plt.ylabel('train')
    plt.xlabel('epoch')
    plt.legend(['train', 'validation'], loc='upper left')
    plt.show()

show_train_history(train_history, 'accuracy', 'val_accuracy',"准确值")

show_train_history(train_history, 'loss', 'val_loss',"损失值")

 

 

 

 

 

5.模型评价

• model.evaluate()
• 交叉表与交叉矩阵
• pandas.crosstab
• seaborn.heatmap

from boto import sns
import pandas as pd
# 5、模型评价
# model.evaluate()
score = model.evaluate(X_test, y_test)
print('score：', score)
# 预测值
y_pred = model.predict_classes(X_test)
print('y_pred：', y_pred[:10])
# 交叉表与交叉矩阵
y_test1 = np.argmax(y_test, axis=1).reshape(-1)
y_true = np.array(y_test1)[0]
# 交叉表查看预测数据与原数据对比
# pandas.crosstab
pd.crosstab(y_true, y_pred, rownames=['true'], colnames=['predict'])
# 交叉矩阵
# seaborn.heatmap
y_test1 = y_test1.tolist()[0]
a = pd.crosstab(np.array(y_test1), y_pred, rownames=['Lables'], colnames=['Predict'])
# 转换成属dataframe
df = pd.DataFrame(a)
sns.heatmap(df, annot=True, cmap="Reds", linewidths=0.2, linecolor='G')
plt.show()