TensorFlow学习1：尝试搭建CNN并实现MNIST数据库的训练

终于有时间静下心来开始TF的系统学习，这里推荐莫烦老师的视频，简洁明了，有Python和ML基础的同学很快就能学会

尝试CNN的搭建和数据的分批次训练

下面直接copy我的渣渣代码，有错误的地方请批评指正：
1 layers.py

import tensorflow as tf

# 定义全连接层
def fc_layer(inputs, in_size, out_size, activation=None, name=None):
    Weights = tf.Variable(tf.truncated_normal(shape=[in_size, out_size], stddev=0.1), dtype=tf.float32, name=name)
    bias = tf.Variable(tf.zeros(shape=[out_size], dtype=tf.float32) + 0.1)
    output = tf.matmul(inputs, Weights) + bias
    if activation != None:
        output = activation(output)
    return output

# 定义卷积层
def conv2d(inputs, kernel_size, stride, in_channel, out_channel, activation=None, padding='SAME', name=None):
    Weights = tf.Variable(tf.truncated_normal(shape=[kernel_size, kernel_size, in_channel, out_channel], stddev=0.1, name=name))
    bias = tf.Variable(tf.zeros(shape=[out_channel]) + 0.1)
    output = tf.nn.conv2d(input=inputs, filter=Weights, strides=[1, stride, stride, 1], padding=padding) + bias
    if activation!= None:
        output = activation(output)
    return output

# 定义池化层
def max_pool(inputs, kernel_size, stride, padding='SAME', name=None):
    return tf.nn.max_pool(value=inputs, ksize=[1, kernel_size, kernel_size, 1], strides=[1, stride, stride, 1], padding=padding, name=name)

# 定义dropout层
def dropout(inputs, keep_prob, name=None):
    return tf.nn.dropout(x=inputs, keep_prob=keep_prob, name=name)

tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=True, data_format=“NHWC”, name=None)

1. input：输入，是形状为[batch_size, h, w, channel]的tensor
2. filter：卷积核，需自行定义Variable，是形状为[kernel_size_h, kernel_size_w, in_channel, out_channel]的tensor
3. strides：卷积步长，通常设置为[1, stride_h, stride_w, 1]
4. padding：填充模式，分为SAME和VALID

tf.nn.max_pool(value, ksize, strides, padding, data_format=“NHWC”, name=None)

1. value：即输入
2. ksize：池化的滑动窗口大小，通常设置为[1, ksize_h, ksize_w, 1]
3. strides：滑动步长，通常设置为[1, stride_h, stride_w, 1]
4. padding：填充模式，分为SAME和VALID

tf.nn.dropout(x, keep_prob, noise_shape=None, seed=None, name=None)

1. x：输入
2. keep_prob：dropout中保留节点的比例（这一点与caffe不同，caffe中的dropout_ratio为drop的节点比例），如：
   layer { name: "drop7" type: "Dropout" bottom: "fc7" top: "fc7" dropout_param { dropout_ratio: 0.5 } }

2 data.py
这里使用常用的MNIST数据库和我经常做的鸟类数据库CUB200-2011

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
import numpy as np
import os
import cv2


def get_data(dataset, img_size=224):
    if dataset == 'MNIST':
        mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
        return mnist.train.images, mnist.train.labels, mnist.validation.images, mnist.validation.labels
    elif dataset =='CUB200-2011':
        return get_cub(img_size)
    else:
        return None


def get_cub(img_size):
    t_root = 'D:/dataset/CUB_200_2011/CUB_200_2011/images_split/train'
    v_root = 'D:/dataset/CUB_200_2011/CUB_200_2011/images_split/val'
    x_train, y_train, x_val, y_val = [], [], [], []
    for folder in os.listdir(t_root):
        t_path = os.path.join(t_root, folder)
        for file in os.listdir(t_path):
            label = int(folder[0:3]) - 1
            y_train.append(label)
            file_name = os.path.join(t_path, file)
            img = cv2.imread(file_name, 1)
            img = cv2.resize(img, dsize=(img_size, img_size))
            x_train.append(img)
    data_train = np.array(x_train).astype(np.float32)
    label_t = tf.one_hot(y_train, 200)
    with tf.Session() as sess:
        label_train = sess.run(label_t)
    print('train data read complete:%d'%data_train.shape[0])
    for folder in os.listdir(v_root):
        v_path = os.path.join(v_root, folder)
        for file in os.listdir(v_path):
            label = int(folder[0:3]) - 1
            y_val.append(label)
            file_name = os.path.join(v_path, file)
            img = cv2.imread(file_name, 1)
            img = cv2.resize(img, dsize=(img_size, img_size))
            x_val.append(img)
    data_val = np.array(x_val).astype(np.float32)
    label_v = tf.one_hot(y_val, 200)
    with tf.Session() as sess:
        label_val = sess.run(label_v)
    print('validation data read complete:%d'%data_val.shape[0])
    return data_train, label_train, data_val, label_val

这里需要注意是：

1. tf.one_hot()返回的依然是tensor，需要通过会话run一下才能给出ndarray格式的数据标签
2. cv2.imread(filename， flag)中第二个参数flag的设置见博客opencv中imread第二个参数的含义

3 LeNet-5.py
这里也没有完全按照LeNet的结构去设置，总之是一个比较简单的网络结构~~

import tensorflow as tf
from datas import *
from layers import *


def get_accu(v_xs, v_ys):
    global pred
    y_pre = sess.run(pred, feed_dict={xs:v_xs, ys:v_ys, keep_prob:1})
    correct_pred = tf.equal(tf.argmax(y_pre, axis=1), tf.argmax(v_ys, axis=1))
    accu = tf.reduce_mean(tf.cast(correct_pred, dtype=tf.float32))
    result = sess.run(accu, feed_dict={xs:v_xs, ys:v_ys, keep_prob:1})
    return result


max_iter = 51
batch_size = 100
img_size = 28
img_channel = 1
classes = 10

# x_train, y_train, x_val, y_val = get_data('CUB200-2011', img_size=img_size)
x_train, y_train, x_val, y_val = get_data('MNIST', img_size=img_size)

# placeholder for input data
xs = tf.placeholder(dtype=tf.float32, shape=[None, 784])
ys = tf.placeholder(dtype=tf.float32, shape=[None, classes])
keep_prob = tf.placeholder(tf.float32)
x_image = tf.reshape(xs, shape=[-1, 28, 28, 1])

# construct the network
conv1 = conv2d(x_image, kernel_size=5, stride=1, in_channel=img_channel, out_channel=6, padding='SAME', activation=tf.nn.relu, name='conv1')
pool1 = max_pool(conv1, 2, 2, padding='SAME', name='pool1')
conv2 = conv2d(pool1, kernel_size=5, stride=1, in_channel=6, out_channel=16, padding='SAME', activation=tf.nn.relu, name='conv2')
pool2 = max_pool(conv2, 2, 2, padding='SAME', name='pool2')
flatten = tf.reshape(pool2, shape=[-1, 7*7*16], name='flatten')
fc1 = fc_layer(flatten, in_size=7*7*16, out_size=256, activation=tf.nn.relu, name='fc1')
dropout = dropout(fc1, keep_prob=keep_prob, name='dropout')
pred = fc_layer(dropout, in_size=256, out_size=classes, activation=tf.nn.softmax, name='pred')

# define loss and optimizer
loss = -tf.reduce_mean(tf.reduce_sum(ys * tf.log(tf.clip_by_value(pred, 1e-4, 1))))
optimizer = tf.train.AdamOptimizer(1e-4)
train_op = optimizer.minimize(loss)

with tf.Session() as sess:
    # initialize variables
    sess.run(tf.global_variables_initializer())
    for step in range(max_iter):
        # batch_x, batch_y= mnist.train.next_batch(batch_size)
        for i in range(x_train.shape[0]//batch_size):
            batch_x = x_train[batch_size * i : batch_size * (i + 1)]
            batch_y = y_train[batch_size * i: batch_size * (i + 1)]
            sess.run(train_op, feed_dict={xs:batch_x, ys:batch_y, keep_prob:0.5})
            if i%20 == 0:
                print(sess.run(loss, feed_dict={xs:batch_x, ys:batch_y, keep_prob:0.5}))
        train_accu = get_accu(x_train, y_train)
        val_accu = get_accu(x_val, y_val)
        print('epoch:%d train accu:%f validation accu:%f'%(step + 1, train_accu, val_accu))

这里需要注意的一点是,tf.reshape()中，shape参数的第一维应设置为-1而不是None，即[-1, 7×7×16]
程序没有使用mnist.train.next_batch(batch_size)生成训练batch，而是手动去生成，但该方法需要将数据库全部load进内存，随后自行按批次送入网络训练。接下来应当学习tf.train.slice_input_producer 和 tf.train.batch 函数的使用

运行结果

在MNIST上使用上述简单CNN架构训练50个epoch，结果如下：

epoch:45 train accu:0.995364 validation accu:0.990600
epoch:46 train accu:0.996000 validation accu:0.989600
epoch:47 train accu:0.995964 validation accu:0.990200
epoch:48 train accu:0.996145 validation accu:0.990600
epoch:49 train accu:0.996291 validation accu:0.990400
epoch:50 train accu:0.996309 validation accu:0.991000

最终在验证集上可以得到99.1%的准确率