卷积神经网络的keras实现

深度残差网络的代码实现

深度残差网络是中微软kaiming he等提出的，在ILSVRC 2015中的classification，detection，localization任务取得了冠军，且在COCO比赛上也夺冠【论文】。在这里对深度残差网络使用keras实现，这里使用的数据集是notmnist，字母A-J，十个字母的识别。(数据集小，渣渣笔记本上运行)

1. 读取数据

import os
import numpy as np
import cv2
import matplotlib.pyplot as plt
import tqdm

train_path='../datasets/notmnist/notMNIST_large/'
test_path='../datasets/notmnist/notMNIST_small/'
img_size=28
num_channels=3

# 读取文件夹图像
def read_letters(folder, size=None):
    image_names=os.listdir(folder,)
    images=[]
    for index, img_name in enumerate(image_names):
        if size and index>size-1: break
        image=cv2.imread(folder+img_name)
        if not image is None: 
            images.append(image)
    return np.array(images)

A=read_letters(train_path+'A/', 100)
print('complete!')
plt.imshow(A[66])
plt.show()

complete!

def concate(x, x_):
    if not x is None: 
        x=np.concatenate([x, x_], axis=0)
    else: x=x_
    return x

def get_data(train_size=None, test_size=None):
    letters=['A', 'B','C','D','E','F','G','H','I','J']
    x_train=y_train=x_test=y_test=None
    for index, letter in tqdm.tqdm(enumerate(letters)):
        x_=read_letters(train_path+letter+'/', train_size)
        x_train=concate(x_train, x_)
        y_=np.ndarray([x_.shape[0]], dtype=np.uint8)
        y_[:]=index
        y_train=concate(y_train, y_) 
        x_=read_letters(test_path+letter+&#