Keras 实战: 复现 LeNet

keras复现LeNet

具体结构如下：

conv + subsampling + conv + subsampling + fully connected

具体解释下LeNet-5结构, 一共是7层，除去输入层实际上是7个feature map。

1:Input

32 * 32 * 1 的灰度图像，输入前做归一化处理

2:C1卷积层

输入：32 * 32 * 1

卷积核大小为 5,步长 1

卷积核的数目：6

输出feature map: , “(32 - 5 + 1) / 1 = 28，每个卷积核一个feature map”

3:S2池化层

输入: 28 * 28 * 6

池化尺寸: 2 * 2

输出: (28/2 = 14)

4:C3卷积层

输入: 14 * 14 * 6

卷积核大小为 5, 步长 1

卷积核的数目：16

输出feature map: , (14 -5 + 1) /1 = 10

5:S4池化层

输入: 101016

池化尺寸: 2*2

输出: (10/2 = 5)

6:C5全连接层(特殊的卷积)

实际上是将输入的feature map flatten 了一下, 然后和下一层全连

输入:

输出: 120维向量

7:F6全连接

输入: 120维向量

输出: 84 维向量

8:输出层

输入:84维向量

输出: 10维向量

接下来直接看Keras复现代码

代码复现

1:模型结构定义

这里我们使用Keras内置的mnist数据集，默认是28 * 28大小，而原论文中是32 * 32，为了和原论文保持一致，我们在网络结构中加入ZeroPadding2D, 在原始输入上下左右补0，其他和原文一致。

import numpy
import pandas
from keras.models import Sequential,Model
from keras.layers import Dense,Input,Conv2D,MaxPooling2D,Flatten,ZeroPadding2D
from keras import callbacks
from keras.wrappers.scikit_learn import KerasRegressor,KerasClassifier
from sklearn.model_selection import cross_val_score
from sklearn.model_selection import KFold,train_test_split
from sklearn.preprocessing import StandardScaler,LabelEncoder
from sklearn.pipeline import Pipeline
from keras.utils import np_utils
from keras.datasets import mnist
from keras.optimizers import RMSprop

def LeNet(input_tensor=None, input_shape=(32, 32, 1), activation='relu'):
    
    act = activation
     
    if input_tensor is None:
        input_tensor = Input(shape = input_shape)
    x = ZeroPadding2D(padding=(2,2), data_format='channels_last')(input_tensor)
    x = Conv2D(filters=6, kernel_size=5, strides=1, activation=act)(x)
    x = MaxPooling2D(pool_size=(2,2), strides=2)(x)
    x = Conv2D(filters=16, kernel_size=5, strides=1, activation=act)(x)
    x = MaxPooling2D(pool_size=2, strides=2)(x)
    x = Flatten()(x)
    x = Dense(units=120, activation=act)(x)
    x = Dense(units=84, activation=act)(x)
    x = Dense(units=10, activation='softmax')(x)

    model = Model(inputs=input_tensor, outputs=x)

    return model

if __name__ == '__main__':
    #load mnist dataset
    (x_train, y_train), (x_test, y_test) = mnist.load_data()
    x_train = np.expand_dims(x_train, axis=3)
    x_test = np.expand_dims(x_test, axis=3)
    print(x_train.shape)
    
    
    x_train,x_test = x_train.astype("float32"),x_test.astype("float32")
    mean = np.mean(x_train, axis=(0,1,2,3))
    std = np.std(x_train, axis=(0,1,2,3))
    # standardlization 
    x_train = (x_train - mean)/(std + 1e-7)
    x_test = (x_test - mean)/(std + 1e-7)
    
    # one-hot encoding
    y_train = np_utils.to_categorical(y_train,num_classes=10)
    y_test = np_utils.to_categorical(y_test,num_classes=10)
    model = LeNet(input_shape=(28,28,1))
    # tensorboard callbacks
    tensorboard_callback = callbacks.TensorBoard(log_dir='./logs', batch_size=10)
    model.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])
    model.fit(x_train, y_train, batch_size=10, validation_data=(x_test, y_test), epochs=42,
             callbacks=[tensorboard_callback])
    model.save('lenet.h5')

(60000, 28, 28, 1)



---------------------------------------------------------------------------

NameError                                 Traceback (most recent call last)

<ipython-input-83-06b24fb3476e> in <module>()
     22     model.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])
     23     model.fit(x_train, y_train, batch_size=10, validation_data=(x_test, y_test), epochs=42,
---> 24              callbacks=[tensorboard_callback])
     25 #     model.save('lenet.h5')
     26 


NameError: name 'tensorboard_callback' is not defined

模型测试

from keras.models import load_model
import cv2
import os
import numpy as np
from matplotlib import pyplot as plt

img_files = [f for f in os.listdir("./testSamples") if not f.startswith('.')]
model = load_model('*.h5')
model.summary()

for i,img_file in enumerate(img_file[:10]):
    img = cv2.imread(os.path.join('./testSamples',img_file),-1)
    plt.show(img)
    img = np.expand_dims(img, 2)
    img = np.expand_dims(img, 0)
    img = (img - np.mean(img)) / np.std(img)
    logits = model.predict(img, batch_size=1)
    label = np.argmax(logits)
    plt.subplot(1, 10, i+1)
    plt.title(str(label))
    plt.show()