Resnet实现CIFAR-10分类

复现何凯明Resnet论文正确率0.9+结果，但是正确率还是没有达到0.9以上，因为 101-layers 和 152-layers 的残差块结构和 34-layers 是有一些区别的。

import tensorflow as tf
import  os
import numpy as np
import pickle

# 文件存放目录
CIFAR_DIR = "./cifar-10-batches-py"


def load_data(filename):
    '''read data from data file'''
    with open(filename, 'rb') as f:
        data = pickle.load(f, encoding='bytes') # python3 需要添加上encoding='bytes'
        return data[b'data'], data[b'labels'] # 并且 在 key 前需要加上 b

class CifarData:
    def __init__(self, filenames, need_shuffle):
        '''参数1:文件夹 参数2:是否需要随机打乱'''
        all_data = []
        all_labels = []

        for filename in filenames:
            # 将所有的数据,标签分别存放在两个list中
            data, labels = load_data(filename)
            all_data.append(data)
            all_labels.append(labels)

        # 将列表 组成 一个numpy类型的矩阵!!!!
        self._data = np.vstack(all_data)
        # 对数据进行归一化, 尺度固定在 [-1, 1] 之间
        self._data = self._data / 127.5 - 1
        # 将列表,变成一个 numpy 数组
        self._labels = np.hstack(all_labels)
        # 记录当前的样本 数量
        self._num_examples = self._data.shape[0]
        # 保存是否需要随机打乱
        self._need_shuffle = need_shuffle
        # 样本的起始点
        self._indicator = 0
        # 判断是否需要打乱
        if self._need_shuffle:
            self._shffle_data()

    def _shffle_data(self):
        # np.random.permutation() 从 0 到 参数,随机打乱
        p = np.random.permutation(self._num_examples)
        # 保存 已经打乱 顺序的数据
        self._data = self._data[p]
        self._labels = self._labels[p]

    def next_batch(self, batch_size):
        '''return batch_size example as a batch'''
        # 开始点 + 数量 = 结束点
        end_indictor = self._indicator + batch_size
        # 如果结束点大于样本数量
        if end_indictor > self._num_examples:
            if self._need_shuffle:
                # 重新打乱
                self._shffle_data()
                # 开始点归零,从头再来
                self._indicator = 0
                # 重新指定 结束点. 和上面的那一句,说白了就是重新开始
                end_indictor = batch_size # 其实就是 0 + batch_size, 把 0 省略了
            else:
                raise Exception("have no more examples")
        # 再次查看是否 超出边界了
        if end_indictor > self._num_examples:
            raise Exception("batch size is larger than all example")

        # 把 batch 区间 的data和label保存,并最后return
        batch_data = self._data[self._indicator:end_indictor]
        batch_labels = self._labels[self._indicator:end_indictor]
        self._indicator = end_indictor
        return batch_data, batch_labels

# 拿到所有文件名称
train_filename = [os.path.join(CIFAR_DIR, 'data_batch_%d' % i) for i in range(1, 6)]
# 拿到标签
test_filename = [os.path.join(CIFAR_DIR, 'test_batch')]

# 拿到训练数据和测试数据
train_data = CifarData(train_filename, True)
test_data = CifarData(test_filename, False)

def residual_block(x, output_channel, is_training):
    '''
    定义残差块儿
    :param x: 输入tensor
    :param output_channel: 输出的通道数
    :return: tensor
    需要注意的是:每经过一个stage,通道数就要 * 2
    在同一个stage中,通道数是没有变化的
    '''
    input_channel = x.get_shape().as_list()[-1] # 拿出 输入 tensor 的 最后一维:也就是通道数
    if input_channel * 2 == output_channel:
        increase_dim = True
        strides = (2, 2) #
    elif input_channel == output_channel:
        increase_dim = False
        strides = (1, 1)
    else:
        raise Exception("input channel can't match output channel")


    conv1 = tf.layers.conv2d(x,
                              output_channel,
                              (3, 3),
                              strides = strides,
                              padding = 'same',
                              activation = None,
                              name = 'conv1'
                             )
    bn = tf.layers.batch_normalization(conv1, training = is_training)
    conv1 = tf.nn.relu(bn)

    conv2 = tf.layers.conv2d(conv1,
                              output_channel,
                              (3, 3),
                              strides = (1, 1), # 因为 上一层 卷积已经进行过降采样,故这里不需要
                              padding = 'same',
                              activation = None,
                              name = 'conv2'
                             )
    bn = tf.layers.batch_normalization(conv2, training = is_training)
    conv2 = tf.nn.relu(bn)

    if increase_dim: # 需要使用降采样
        # pooled_x 数据格式 [ None, image_width, image_height, channel ]
        # 要求格式 [ None, image_width, image_height, channel * 2 ]
        pooled_x = tf.layers.average_pooling2d(x,
                                                (2, 2), # size
                                                (2, 2), # stride
                                                padding = 'valid'
                                               )
        '''
        如果输出通道数是输入的两倍的话,需要增加通道数量.
        maxpooling 只能降采样,而不能增加通道数,
        所以需要单独增加通道数
        '''
        padded_x = tf.pad(pooled_x, # 参数 2 ,在每一个通道上 加 pad
                           [
                               [ 0, 0 ],
                               [ 0, 0 ],
                               [ 0, 0 ],
                               [input_channel // 2, input_channel // 2] # 实际上就是 2倍input_channel,需要均分开
                            ]
                          )
    else:
        padded_x = x

    output_x = conv2 + padded_x   # 就是 公式: H(x) = F(x) + x
    return  output_x

def res_net(x, num_residual_blocks, num_filter_base, class_num, is_training):
    '''
    残差网络主程序
    :param