• 深度学习练手项目——DNN识别手写数字


    该案例主要目的是为了熟悉Keras基本用法,以及了解DNN基本流程。

    示例代码:

    import numpy as np
    import matplotlib.pyplot as plt
    from keras.models import Sequential
    from keras.datasets import mnist
    from keras.layers import Dense
    from keras.utils.np_utils import to_categorical
    
    #加载数据,训练60000条,测试10000条,X_train.shape=(60000,28,28)
    (X_train, y_train), (X_test, y_test) = mnist.load_data()
    #特征扁平化,缩放,标签独热
    X_train_flat = X_train.reshape(60000, 28*28)
    X_test_flat = X_test.reshape(10000, 28*28)
    X_train_norm = X_train_flat / 255 
    X_test_norm = X_test_flat / 255
    y_train_onehot = to_categorical(y_train, 10) #shape为(60000,10)
    y_test_onehot = to_categorical(y_test, 10) #shape为(10000,10)
    #构建模型
    model = Sequential()
    model.add(Dense(100, activation='relu', input_shape=(28*28,)))
    model.add(Dense(50, activation='relu'))
    model.add(Dense(10, activation='softmax'))
    #模型配置和训练
    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    model.fit(X_train_norm, y_train_onehot, epochs=5, batch_size=32, verbose=1)
    print("训练完毕!")
    

    训练结果为:

    继续在测试集上评估模型。

    #测试集上评估表现
    score = model.evaluate(X_test_norm, y_test_onehot)
    print("在测试集上评估完毕!")
    print("在测试集上表现:Loss={:.4f}, Accuracy={:.4f}".format(score[0], score[1]))
    #在测试集上预测
    y_pred_class = model.predict_classes(X_test_norm)  #shape=(10000,)
    print("预测完毕!")
    #查看预测效果,随机查看多张图片
    idx = 22  #随机设置
    count = 0
    fig1 = plt.figure(figsize = (10,7))
    for i in range(3):
        for j in range(5):
            count += 1
            ax = plt.subplot(3,5,count)
            plt.imshow(X_test[idx+count])  
            ax.set_title("predict:{} label:{}".format(y_pred_class[idx+count], 
                                                      y_test[idx+count]))
    fig1.savefig('images/look.jpg')
    

    运行结果为:


    为了了解模型预测错误原因,可查看预测错误的图片。

    #找出错误所在
    X_test_err = X_test[y_test!=y_pred_class]  #(num_errors, 28, 28)
    y_test_err = y_test[y_test!=y_pred_class]  #(num_errors,)
    y_pred_class_err = y_pred_class[y_test!=y_pred_class]
    #连续查看多张错误图片
    idx = -1
    count = 0
    fig2 = plt.figure(figsize = (10,7))
    for i in range(3):
        for j in range(5):
            count += 1
            ax = plt.subplot(3,5,count)
            plt.imshow(X_test_err[idx+count])  
            ax.set_title("predict:{} label:{}".format(y_pred_class_err[idx+count], 
                                                      y_test_err[idx+count]))
    fig2.savefig('images/errors.jpg')
    

    运行结果为:

  • 相关阅读:
    Jquery 的offset与position方法
    HTML精确定位:scrollLeft,scrollWidth,clientWidth,offsetWidth之完全详解
    js中对日期进行加减
    EditPlus 换行
    constructor属性,valueOf(),toSTring
    浅谈C#增加that关键字
    ASP.NET Global.ascx 事件大全
    Javascript:scrollWidth,clientWidth,offsetWidth的区别
    jquery val() 返回值怎么转换成 javascript 日期类型
    微软宣布Entity Framework 5的性能有了显著提升
  • 原文地址:https://www.cnblogs.com/inchbyinch/p/11515011.html
Copyright © 2020-2023  润新知