【Deep Learning】tensorflow实现卷积神经网络(AlexNet)

一、实验要求

        1.使用卷积神经网络实现图片分类，数据集为OxFlowers17；

二、实验环境

        Anaconda2-4.3.1（Python2.7），tensorflow-cpu。

三、实验原理

3.1  数据读取

        已知数据集是连续80个样本为一个分类，有17个类别，所以一共只有1360个样本，所以首先用一个函数把每一类的样本分到一个文件夹，文件夹的名字从0开始命名，接着进行数据的读取，由于本次实验使用的数据集比较小，因而没有划分出验证集，根据官网给出的划分数据集的方法，把验证集也划分给训练集，以丰富训练集，最后得到1020个训练样本和340个测试样本。

        对应代码：loadData.py

# -*- coding: UTF-8 -*-

import os
import numpy as np
import scipy.io as sio
from PIL import Image

dir='../oxflower17/jpg/'

def to_categorical(y, nb_classes):
    y = np.asarray(y, dtype='int32')
    if not nb_classes:
        nb_classes = np.max(y)+1
    Y = np.zeros((1, nb_classes))
    Y[0,y] = 1.
    return Y

def build_class_directories(dir):
    dir_id = 0
    class_dir = os.path.join(dir, str(dir_id))
    if not os.path.exists(class_dir):
        os.mkdir(class_dir)
    for i in range(1, 1361):
        fname = "image_" + ("%.4i" % i) + ".jpg"
        os.rename(os.path.join(dir, fname), os.path.join(class_dir, fname))
        if i % 80 == 0 and dir_id < 16:
            dir_id += 1
            class_dir = os.path.join(dir, str(dir_id))
            os.mkdir(class_dir)

def get_input(resize=[224,224]):
    print 'Load data...'

    getJPG = lambda filePath: np.array(Image.open(filePath).resize(resize))

    dataSet=[];labels=[];choose=1
    classes = os.listdir(dir)
    for index, name in enumerate(classes):
        class_path = dir+ name + "/"
        if os.path.isdir(class_path):
            for img_name in os.listdir(class_path):
                img_path = class_path + img_name
                img_raw = getJPG(img_path)
                dataSet.append(img_raw)
                y=to_categorical(int(name),17)
                labels.append(y)

    datasplits = sio.loadmat('../oxflower17/datasplits.mat')
    keys = [x + str(choose) for x in ['val', 'trn', 'tst']]
    train_set, vall_set, test_set = [set(list(datasplits[name][0])) for name in keys]
    train_data, train_label,test_data ,test_label= [],[],[],[]

    for i in range(len(labels)):
        num = i + 1
        if num in test_set:
            test_data.append(dataSet[i])
            test_label.extend(labels[i])
        else:
            train_data.append(dataSet[i])
            train_label.extend(labels[i])
    train_data = np.array(train_data, dtype='float32')
    train_label = np.array(train_label, dtype='float32')
    test_data = np.array(test_data, dtype='float32')
    test_label = np.array(test_label, dtype='float32')
    return train_data, train_label,test_data ,test_label

def batch_iter(data, batch_size, num_epochs, shuffle=True):
    data = np.array(data)
    data_size = len(data)
    num_batches_per_epoch = int((len(data)-1)/batch_size) + 1
    for epoch in range(num_epochs):
        # Shuffle the data at each epoch
        if shuffle:
            shuffle_indices = np.random.permutation(np.arange(data_size))
            shuffled_data = data[shuffle_indices]
        else:
            shuffled_data = data
        for batch_num in range(num_batches_per_epoch):
            start_index = batch_num * batch_size
            end_index = min((batch_num + 1) * batch_size, data_size)
            yield shuffled_data[start_index:end_index]

if __name__=='__main__':
    #build_class_directories(os.path.join(dir))
    train_data, train_label,test_data, test_label=get_input()
    print len(train_data),len(test_data)

3.2  模型介绍

        本次实验使用的是AlexNet模型，AlexNet共有八层，其前五层是卷积层，后三层是全连接层。最后一个全连接层的输出具有17个输出单元。

        其中第一个卷积层