| 序号 | 网络结构 | loss和acc | ||
| 2 | model = Sequential() model.add(Dense(units = 121,input_dim = 28 * 28)) model.add(Activation('relu')) model.add(Dense(units = 81)) model.add(Activation('relu')) model.add(Dense(units = 10)) model.add(Activation('softmax')) | loss: 0.11605372323654592 acc: 0.9649999737739563 | ||
| 3 | model = Sequential() model.add(Dense(units = 121,input_dim = 28 * 28)) model.add(Activation('relu')) model.add(Dense(units = 108)) model.add(Activation('relu')) model.add(Dense(units = 10)) model.add(Activation('softmax')) | loss: 0.11722589706927537 acc: 0.964600026607513 | ||
| 4 | model = Sequential() model.add(Dense(units = 121,input_dim = 28 * 28)) model.add(Activation('relu')) model.add(Dense(units = 81)) model.add(Activation('relu')) model.add(Dense(units = 108)) model.add(Activation('relu')) model.add(Dense(units = 10)) model.add(Activation('softmax')) | loss: 0.10050583745818585 acc: 0.9702000021934509 | ||
| 5 | model = Sequential() model.add(Dense(units = 121,input_dim = 28 * 28)) model.add(Activation('relu')) model.add(Dense(units = 81)) model.add(Activation('relu')) model.add(Dense(units = 108)) model.add(Activation('relu')) model.add(Dense(units = 138)) model.add(Activation('relu')) model.add(Dense(units = 169)) model.add(Activation('relu')) model.add(Dense(units = 10)) model.add(Activation('softmax')) | loss: 0.11574177471690346 acc: 0.9668999910354614 | ||
| 6 | model = Sequential() model.add(Dense(units = 121,input_dim = 28 * 28)) model.add(Activation('relu')) model.add(Dropout(0.5)) model.add(Dense(units = 81)) model.add(Activation('relu')) model.add(Dropout(0.25)) model.add(Dense(units = 10)) model.add(Activation('softmax')) model.compile( loss = 'categorical_crossentropy', optimizer = keras.optimizers.SGD(lr=0.01,momentum=0.9,nesterov=True), metrics = ['accuracy'] ) | loss: 0.08351417180134449 acc: 0.9753000140190125 | ||
| 以上都是多层感知机，以下都是卷积神经网络 | ||||
| 7 | def read_images(filename,items): file_image = open(filename,'rb') file_image.seek(16) data = file_image.read(items * 28 * 28) X = np.zeros(items * 28 * 28) for i in range(items * 28 * 28): X[i] = data[i] / 255 file_image.close() return X.reshape(-1,28 ,28,1) X_train = read_images('D:/dl4cv/datesets/mnist/train-images.idx3-ubyte',60000) X_test = read_images('D:/dl4cv/datesets/mnist/t10k-images.idx3-ubyte',10000) y_train = keras.utils.to_categorical(y_train,10) y_test = keras.utils.to_categorical(y_test,10) model = Sequential() model.add(Conv2D(32,kernel_size = (3,3),activation='relu',input_shape=(28,28,1))) model.add(MaxPooling2D(pool_size =(2,2))) model.add(Conv2D(64,(3,3),activation='relu')) model.add(MaxPooling2D(pool_size =(2,2))) model.add(Flatten()) model.add(Dense(128,activation='relu')) model.add(Dropout(0.5)) model.add(Dense(10,activation='softmax')) model.compile( loss = keras.losses.categorical_crossentropy, optimizer = keras.optimizers.Adadelta(), metrics = ['accuracy'] ) #plot_model(model, to_file='model1.png', show_shapes=True) model.fit( X_train, y_train, batch_size=128, epochs=10, verbose=1, validation_data =(X_test,y_test) ) | loss: 0.020046546813212628 acc: 0.993399977684021 注意事项：一般是dense后才链接dropout 中间用relu,最后用softmax flatten一般在最后的地方。 | ||
| 8 | model = Sequential() model.add(Conv2D(6,kernel_size = (5,5),strides = 1,activation ='relu', input_shape = (28,28,1))) #filters: 整数，输出空间的维度 （即卷积中滤波器的输出数量） model.add(MaxPooling2D(pool_size = (2,2),strides = 2)) model.add(Conv2D(16,kernel_size = (5,5),strides = 1,activation ='relu')) #卷积核越小，filters越长 model.add(MaxPooling2D(pool_size = (2,2),strides = 2)) model.add(Flatten()) #model.add(Dense(84,activation='relu')) 这里添加之后，可以提高0.1左右 model.add(Dense(10,activation='softmax')) #输出10类 | 原始的letnet loss: 0.05069914509201189 acc: 0.9839000105857849 | ||
LeNet5网络虽然很小，但是包含了深度学习的基本模块：卷积层、池化层、全连接层。LeNet5共有七层，不包含输入，每层都包含可训练参数，每个层有多个Feature Map，每个Feature Map通过一种卷积滤波器提取输入的一种特征，然后每Feature Map有多个神经元。
输入： 32∗32
32∗32的手写字体图片，这些手写字体包含0-9数字，也就是相当于10个类别的图片。
输出： 分类结果，0-9之间的一个数（softmax）
2.2 各层结构及参数
1. INPUT（输入层）
32∗32
2. C1(卷积层)
选取6个5∗5
3. S2（池化层）
4. C3（卷积层）
选取61个5∗5
5. S4（池化层）
6. C5（卷积层）
总共120个feature map，每个feature map与S4层所有的feature map相连接，卷积核大小为5∗5
7. F6（全连接层）
F6相当于MLP（Multi-Layer Perceptron，多层感知机）中的隐含层，有84个节点，所以有84∗(120+1)=10164
8. Output（输出层）
全连接层，共有10个节点，采用的是径向基函数（RBF）的网络连接方式。