1.手写数字数据集
- from sklearn.datasets import load_digits
- digits = load_digits()
代码：
1 from sklearn.datasets import load_digits 2 digits = load_digits()
2.图片数据预处理
- x：归一化MinMaxScaler()
代码：
1 #对x集归一化;对y进行热编码 2 import matplotlib.pyplot as plt 3 import numpy as np 4 from sklearn.datasets import load_digits #小数据集为8*8/大的为28*28 5 from sklearn.model_selection import train_test_split 6 from sklearn.preprocessing import MinMaxScaler 7 from sklearn.preprocessing import OneHotEncoder 8 import tensorflow as tf 9 from sklearn.metrics import accuracy_score 10 11 X_data = digits.data.astype(np.float32) 12 Y_data = digits.target.astype(np.float32).reshape(-1,1) #将数据转为一列 13 14 #将属性xSHUJHU数据进行归一放在最大与最小之间（0,1） 15 scaler = MinMaxScaler() 16 X_data = scaler.fit_transform(X_data) #归一 17 print("看效果",X_data)
- y：独热编码OneHotEncoder()或to_categorical
代码：
from sklearn.preprocessing import OneHotEncoder Y = OneHotEncoder().fit_transform(Y_data).todense() print('one-hot_Y:') print(Y) # 热编码有一说一
- 训练集测试集划分
代码：
1 #先对归一的数据集转为图片格式 2 X=X_data.reshape(-1,8,8,1) 3 4 #训练集测试集划分 5 from sklearn.model_selection import train_test_split #测试与训练2/8分 6 X_tain,X_test,y_train,y_test = train_test_split(X,Y,test_size=0.2,random_state=0,stratify=Y) 7 print(X_tain.shape,X_test.shape,y_train.shape,y_test.shape)
- 张量结构
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。
读入手写数字识别经过输入层，经过最小维度的卷积层，经过池化，越往下卷积核的数目越多，提出的特征也越多；反复后最终到达全链接层。即有两个连续的卷积-池化就会知道第二个卷积是在获得图片的压缩版。
from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense,Dropout,Conv2D,MaxPool2D,Flatten model = Sequential() # 建立模型 #设定卷积核 ks = (2, 2) #输入层（1层）此处input_shape需要指定训练集的数据，往下之后是会自动推导 model.add(Conv2D(filters=8, kernel_size=ks, padding='same', input_shape=X_tain.shape[1:], activation='relu')) model.add(MaxPool2D(pool_size=(2, 2))) # 池化层 model.add(Dropout(0.25)) # 防止过拟合丢掉1/4的链接 model.add(Conv2D(filters=16, kernel_size=ks, padding='same', activation='relu'))# 第2卷积层 model.add(MaxPool2D(pool_size=(2, 2)))# 池化 model.add(Dropout(0.25))# 防止过拟合丢掉1/4的链接 model.add(Conv2D(filters=324, kernel_size=ks, padding='same', activation='relu'))# 第三卷积层 model.add(Conv2D(filters=64, kernel_size=ks, padding='same', activation='relu'))# 第四卷积层 model.add(MaxPool2D(pool_size=(2, 2)))# 池化 model.add(Dropout(0.25)) model.add(Flatten())# 平坦层 model.add(Dense(64, activation='relu'))# 全连接层 model.add(Dropout(0.25)) model.add(Dense(10, activation='softmax'))# 激活函数softmax # 输出模型各层的参数状况 print(model.summary())
4.模型训练
- model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
- train_history = model.fit(x=X_train,y=y_train,validation_split=0.2, batch_size=300,epochs=10,verbose=2)
代码：
# 4 模型训练 model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) train_history = model.fit(x=X_tain,y=y_train,validation_split=0.2, batch_size=128,epochs=10,verbose=2) # 定义训练参数可视化 import matplotlib.pyplot as plt # &matplotlib inline 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','validations'],loc='upper left') plt.show() # 查看准确率 show_train_history(train_history,'accuracy','val_accuracy') # 查看损失率 show_train_history(train_history,'lose','val_lose')
准确率
错误率
5.模型评价
- model.evaluate()
- 交叉表与交叉矩阵
- pandas.crosstab
- seaborn.heatmap
实验代码：
# 5.模型评估 score = model.evaluate(X_test,y_test) print(score) print("查看历史训练过程",train_history.history) # 预测值 y_pre = model.predict_classes(X_test) print("y_pre:",y_pre[:10]) print("y_test:",y_test[:10]) y_test1 = np.argmax(y_test,axis=1).reshape(-1) # print(y_test1) y_true = np.array(y_test1)[0] # 交叉表与交叉矩阵 import pandas as pd pd.crosstab(y_true,y_pre,rownames=['true'],colnames=['pre']) # 交叉矩阵 import seaborn as sns y_test1 = y_test1.tolist()[0] a = pd.crosstab(np.array(y_test1), y_pre, rownames=['Lables'], colnames=['Predict']) df = pd.DataFrame(a) # 转换成属dataframe sh = sns.heatmap(df, annot=True, cmap="Reds", linewidths=0.2, linecolor='G') plt.show(sh) plt.savefig('digits_heatmap.peg')