目前流行的深度學(xué)習(xí)庫有Caffe,，Keras，Theano,，本文采用谷歌開源的曾用來制作AlphaGo的深度學(xué)習(xí)系統(tǒng)Tensorflow,。

1：安裝Tensorflow

最早TensorFlow只支持mac和Linux系統(tǒng),，目前也支持windows系統(tǒng)，但要求python3.5 (64bit)版本,。TensorFlow有cpu和gpu版本,，由于本文使用服務(wù)器是NVIDIA顯卡，因此安裝gpu版本,，在cmd命令行鍵入

pip install --upgrade tensorflow-gpu

如果出現(xiàn)錯誤“Cannot remove entries from nonexistent file”,，執(zhí)行以下命令

“pip install --upgrade -I setuptools”，安裝成功出現(xiàn)以下界面

2：安裝CUDA庫

用gpu來運行Tensorflow還需要配置CUDA和cuDnn庫,，

用以下link下載win10(64bit)版本CUDA安裝包,，大小約為1.2G https://developer./cuda-downloads

安裝cuda_8.0.61_win10.exe,，完成后配置系統(tǒng)變量，在系統(tǒng)變量中的CUDA_PATH中,，加上三個路徑,， C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\bin

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\bin\lib\x64

3：安裝cuDnn庫

用以下link下載cuDnn庫

https://developer./cudnn

下載解壓后，為了在運行tensorflow的時候也能將這個庫加載進去,，我們要將解壓后的文件拷到CUDA對應(yīng)的文件夾下C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0

4：測試安裝

在Pycharm新建一個py文件鍵入

import tensorflow as tf

hello = tf.constant('Hello world, TensorFlow!')

sess = tf.Session()

print(sess.run(hello))

如果能夠輸出'Hello, TensorFlow!'像下面這樣就代表配置成功了,。

1: 導(dǎo)入必要的模塊

import sys

import cv2

import numpy as np

import tensorflow as tf

2: 定義CNN的基本組件

按照LeNet5 的定義，采用32*32 圖像輸入,，CNN的基本組件包括卷積C1,，采樣層S1、卷積C2,，采樣層S2,、全結(jié)合層1、分類層2

3：訓(xùn)練CNN

將輸入圖像縮小至32*32大小,，采用opencv中的resize函數(shù)

其變換參數(shù)有

CV_INTER_NN - 最近鄰插值, 

CV_INTER_LINEAR - 雙線性插值 (缺省使用) 

CV_INTER_AREA - 使用象素關(guān)系重采樣,。當(dāng)圖像縮小時候，該方法可以避免波紋出現(xiàn),。當(dāng)圖像放大時,，類似于 CV_INTER_NN 方法.. 

CV_INTER_CUBIC - 立方插值. 

output=cv2.resize(img,(32,32),interpolation=cv2.INTER_CUBIC)

核心代碼如下：

class CNNetwork:

    NUM_CLASSES = 2 #分兩類

IMAGE_SIZE = 28

IMAGE_PIXELS = IMAGE_SIZE*IMAGE_SIZE*3

def inference(images_placeholder, keep_prob):

    def weight_variable(shape):

      initial = tf.truncated_normal(shape, stddev=0.1)

      return tf.Variable(initial)

    def bias_variable(shape):

      initial = tf.constant(0.1, shape=shape)

      return tf.Variable(initial)

    def conv2d(x, W):

      return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

    def max_pool_2x2(x):

      return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],

                            strides=[1, 2, 2, 1], padding='SAME')

    x_image = tf.reshape(images_placeholder, [-1, 28, 28, 1])

    with tf.name_scope('conv1') as scope:

        W_conv1 = weight_variable([5, 5, 3, 32])

        b_conv1 = bias_variable([32])

        h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)

    with tf.name_scope('pool1') as scope:

        h_pool1 = max_pool_2x2(h_conv1)

    with tf.name_scope('conv2') as scope:

        W_conv2 = weight_variable([5, 5, 32, 64])

        b_conv2 = bias_variable([64])

        h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)

    with tf.name_scope('pool2') as scope:

        h_pool2 = max_pool_2x2(h_conv2)

    with tf.name_scope('fc1') as scope:

        W_fc1 = weight_variable([7*7*64, 1024])

        b_fc1 = bias_variable([1024])

        h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])

        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

        h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

    with tf.name_scope('fc2') as scope:

        W_fc2 = weight_variable([1024, NUM_CLASSES])

        b_fc2 = bias_variable([NUM_CLASSES])

    with tf.name_scope('softmax') as scope:

        y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

    return y_conv

if __name__ == '__main__':

    test_image = []

    filenames = []

    for i in range(1, len(sys.argv)):

        img = cv2.imread(sys.argv[i])

        img = cv2.resize(img, (28, 28))

        test_image.append(img.flatten().astype(np.float32)/255.0)

        filenames.append(sys.argv[i])

    test_image = np.asarray(test_image)

    images_placeholder = tf.placeholder("float", shape=(None, IMAGE_PIXELS))

    labels_placeholder = tf.placeholder("float", shape=(None, NUM_CLASSES))

    keep_prob = tf.placeholder("float")

    logits = inference(images_placeholder, keep_prob)

    sess = tf.InteractiveSession()

    saver = tf.train.Saver()

    sess.run(tf.initialize_all_variables())

    saver.restore(sess, "model.ckpt")

    for i in range(len(test_image)):

        pred = np.argmax(logits.eval(feed_dict={

            images_placeholder: [test_image[i]],

            keep_prob: 1.0 })[0])

        pred2 = logits.eval(feed_dict={

            images_placeholder: [test_image[i]],

            keep_prob: 1.0 })[0]

        print filenames[i],pred,"{0:10.8f}".format(pred2[0]),"{0:10.8f}".format(pred2[1])