4.1.3 卷积神经网络实战

• 使用神经网路进行图像分类

  使用TensorFlow封装的函数简化定义模型的过程

  # (3072, 10)
  w = tf.get_variable('w', [x.get_shape()[-1], 10],
                      initializer=tf.random_normal_initializer(0, 1))
  # (10, )
  b = tf.get_variable('b', [10],
                     initializer=tf.constant_initializer(0.0))
  # [None, 3072] * [3072, 1] = [None, 1]
  y_ = tf.matmul(x, w) + b
  

  上面这段代码可以替换为

  y_ = tf.layers.dense(hidden, 10)
  

  dense函数时一个全连接的api，我们可以用他构建更多的层次

  下面的代码构建了一个三个隐含层的神经网络，前两层有100个神经元，第三层有50个，激活函数都是relu

  # activation选择激活函数
  hidden1 = tf.layers.dense(x, 100, activation=tf.nn.relu)
  hidden2 = tf.layers.dense(hidden1, 100, activation=tf.nn.relu)
  hidden3 = tf.layers.dense(hidden2, 50, activation=tf.nn.relu)
  
  y_ = tf.layers.dense(hidden3, 10)
  

  修改完代码后重新跑我们的测试，模型准确率达到了百分之50

  [Train] Step: 500, loss: 2.14457, acc: 0.25000
  [Train] Step: 1000, loss: 1.38850, acc: 0.45000
  [Train] Step: 1500, loss: 1.48442, acc: 0.45000
  [Train] Step: 2000, loss: 1.30306, acc: 0.70000
  [Train] Step: 2500, loss: 1.81453, acc: 0.35000
  [Train] Step: 3000, loss: 1.25715, acc: 0.55000
  [Train] Step: 3500, loss: 1.24998, acc: 0.55000
  [Train] Step: 4000, loss: 1.52799, acc: 0.45000
  [Train] Step: 4500, loss: 1.40961, acc: 0.40000
  [Train] Step: 5000, loss: 1.29267, acc: 0.65000
  (10000, 3072)
  (10000,)
  [Test ] Step: 5000, acc: 0.46650
  [Train] Step: 5500, loss: 1.61286, acc: 0.20000
  [Train] Step: 6000, loss: 1.14901, acc: 0.45000
  [Train] Step: 6500, loss: 1.59980, acc: 0.60000
  [Train] Step: 7000, loss: 1.80693, acc: 0.40000
  [Train] Step: 7500, loss: 1.60266, acc: 0.45000
  [Train] Step: 8000, loss: 1.46613, acc: 0.65000
  [Train] Step: 8500, loss: 1.77019, acc: 0.45000
  [Train] Step: 9000, loss: 1.57591, acc: 0.50000
  [Train] Step: 9500, loss: 0.96180, acc: 0.75000
  [Train] Step: 10000, loss: 1.08688, acc: 0.70000
  (10000, 3072)
  (10000,)
  [Test ] Step: 10000, acc: 0.49750
  

• 使用卷积神经网络进行图像分类

  全连接有api，池化核卷积当然也有

  # conv1：神经元图,feature map,输出图像
  conv1 = tf.layers.conv2d(x_image,
                           32, # output channel number
                           (3,3), # kernal size
                           padding = 'same', # same 代表输出图像的大小没有变化，valid 代表不做padding
                           activation = tf.nn.relu,
                           name = 'conv1'
                           )
  # 16*16
  pooling1 = tf.layers.max_pooling2d(conv1,
                                     (2, 2), # kernal size
                                     (2, 2), # stride
                                     name = 'pool1' # name为了给这一层做一个命名，这样会让图打印出来的时候会是一个有意义的图
                                    )
  

  这样就构建了一个卷积层一个池化层，我们可以再复制两遍，修改输入，这样就有了三层卷积层三层池化层

  最后再加一层全连接

  conv2 = tf.layers.conv2d(pooling1,
                           32, # output channel number
                           (3,3), # kernal size
                           padding = 'same', # same 代表输出图像的大小没有变化，valid 代表不做padding
                           activation = tf.nn.relu,
                           name = 'conv2'
                           )
  # 8*8
  pooling2 = tf.layers.max_pooling2d(conv2,
                                     (2, 2), # kernal size
                                     (2, 2), # stride
                                     name = 'pool2' # name为了给这一层做一个命名，这样会让图打印出来的时候会是一个有意义的图
                                    )
  
  conv3 = tf.layers.conv2d(pooling2,
                           32, # output channel number
                           (3,3), # kernal size
                           padding = 'same', # same 代表输出图像的大小没有变化，valid 代表不做padding
                           activation = tf.nn.relu,
                           name = 'conv3'
                           )
  # 4*4*32
  pooling3 = tf.layers.max_pooling2d(conv3,
                                     (2, 2), # kernal size
                                     (2, 2), # stride
                                     name = 'pool3' # name为了给这一层做一个命名，这样会让图打印出来的时候会是一个有意义的图
                                    )
  
  # [None, 4*4*42] 将三通道的图形转换成矩阵
  flatten = tf.layers.flatten(pooling3)
  y_ = tf.layers.dense(flatten, 10)
  

  这样卷积神经网络结构就搭建完了，可以看到，我们在完成数据准备和测试结果之后，使用TensorFlow的API构建网络结构还是非常简单的。

  使用卷积神经网络，10000次训练进行测试，分类结果达到了百分之69

  [Train] Step: 500, loss: 1.24817, acc: 0.60000
  [Train] Step: 1000, loss: 1.24423, acc: 0.50000
  [Train] Step: 1500, loss: 1.15608, acc: 0.55000
  [Train] Step: 2000, loss: 0.89077, acc: 0.85000
  [Train] Step: 2500, loss: 0.91770, acc: 0.60000
  [Train] Step: 3000, loss: 1.09620, acc: 0.55000
  [Train] Step: 3500, loss: 0.83352, acc: 0.70000
  [Train] Step: 4000, loss: 1.00452, acc: 0.60000
  [Train] Step: 4500, loss: 1.13865, acc: 0.60000
  [Train] Step: 5000, loss: 0.63163, acc: 0.85000
  (10000, 3072)
  (10000,)
  [Test ] Step: 5000, acc: 0.64850
  [Train] Step: 5500, loss: 1.29329, acc: 0.55000
  [Train] Step: 6000, loss: 1.14539, acc: 0.65000
  [Train] Step: 6500, loss: 0.48069, acc: 0.80000
  [Train] Step: 7000, loss: 1.02633, acc: 0.65000
  [Train] Step: 7500, loss: 0.93267, acc: 0.70000
  [Train] Step: 8000, loss: 0.97426, acc: 0.70000
  [Train] Step: 8500, loss: 0.97432, acc: 0.75000
  [Train] Step: 9000, loss: 0.84112, acc: 0.70000
  [Train] Step: 9500, loss: 0.79695, acc: 0.70000
  [Train] Step: 10000, loss: 0.64198, acc: 0.80000
  (10000, 3072)
  (10000,)
  [Test ] Step: 10000, acc: 0.69950