張量流中連接vgg19的問題

Question

我有一個問題讓我的VGG19神經網絡連接。下面是整個圖表：

class VGG(object): 
    def __init__(self, n_classes=252): 

     self.input_data = tf.placeholder(dtype=tf.float32, name='input_data', shape=(5, 224, 224, 3)) 

     with tf.variable_scope("group1"): 
      with tf.variable_scope("conv1"): 
       self.weights1 = tf.get_variable('weights',[3, 3, 3, 64], initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias1 = tf.get_variable('bias', [64], initializer=tf.constant_initializer(0.0)) 
       self.convolve1 = tf.nn.conv2d(self.input_data, self.weights1, [1, 1, 1, 1], padding="SAME") 
       self.conv1 = tf.nn.elu(tf.nn.bias_add(self.convolve1, self.bias1)) 
      with tf.variable_scope("conv2"): 
       self.weights2= tf.get_variable('weights', [3, 3, 64, 64], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias2 = tf.get_variable('bias', [64], initializer=tf.constant_initializer(0.0)) 
       self.convolve2 = tf.nn.conv2d(self.conv1, self.weights2, [1, 1, 1, 1], padding="SAME") 
       self.conv2 = tf.nn.elu(tf.nn.bias_add(self.convolve2, self.bias2)) 

      self.pool1 = tf.nn.max_pool(self.conv2, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME', name='pool1') 

     with tf.variable_scope("group2"): 
      with tf.variable_scope("conv3"): 
       self.weights3 = tf.get_variable('weights', [3, 3, 64, 128], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias3 = tf.get_variable('bias', [128], initializer=tf.constant_initializer(0.0)) 
       self.convolve3 = tf.nn.conv2d(self.pool1, self.weights3, [1, 1, 1, 1], padding="SAME") 
       self.conv3 = tf.nn.elu(tf.nn.bias_add(self.convolve3, self.bias3)) 
      with tf.variable_scope("conv4"): 
       self.weights4 = tf.get_variable('weights', [3, 3, 128, 128], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias4 = tf.get_variable('bias', [128], initializer=tf.constant_initializer(0.0)) 
       self.convolve4 = tf.nn.conv2d(self.conv3, self.weights4, [1, 1, 1, 1], padding="SAME") 
       self.conv4 = tf.nn.elu(tf.nn.bias_add(self.convolve4, self.bias4)) 
      self.pool2 = tf.nn.max_pool(self.conv4, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME', name='pool2') 

     with tf.variable_scope("group3"): 
      with tf.variable_scope("conv5"): 
       self.weights5 = tf.get_variable('weights', [3, 3, 128, 256], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias5 = tf.get_variable('bias', [256], initializer=tf.constant_initializer(0.0)) 
       self.convolve5 = tf.nn.conv2d(self.pool2, self.weights5, [1, 1, 1, 1], padding="SAME") 
       self.conv5 = tf.nn.elu(tf.nn.bias_add(self.convolve5, self.bias5)) 
      with tf.variable_scope("conv6"): 
       self.weights6 = tf.get_variable('weights', [3, 3, 256, 256], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias6 = tf.get_variable('bias', [256], initializer=tf.constant_initializer(0.0)) 
       self.convolve6 = tf.nn.conv2d(self.conv5, self.weights6, [1, 1, 1, 1], padding="SAME") 
       self.conv6 = tf.nn.elu(tf.nn.bias_add(self.convolve6, self.bias6)) 
      with tf.variable_scope("conv7"): 
       self.weights7 = tf.get_variable('weights', [3, 3, 256, 256], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias7 = tf.get_variable('bias', [256], initializer=tf.constant_initializer(0.0)) 
       self.convolve7 = tf.nn.conv2d(self.conv6, self.weights7, [1, 1, 1, 1], padding="SAME") 
       self.conv7 = tf.nn.elu(tf.nn.bias_add(self.convolve7, self.bias7)) 
      with tf.variable_scope("conv8"): 
       self.weights8 = tf.get_variable('weights', [3, 3, 256, 256], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias8 = tf.get_variable('bias', [256], initializer=tf.constant_initializer(0.0)) 
       self.convolve8 = tf.nn.conv2d(self.conv7, self.weights8, [1, 1, 1, 1], padding="SAME") 
       self.conv8 = tf.nn.elu(tf.nn.bias_add(self.convolve8, self.bias8)) 
      self.pool3 = tf.nn.max_pool(self.conv8, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME', name='pool3') 

     with tf.variable_scope("group4"): 
      with tf.variable_scope("conv9"): 
       self.weights9 = tf.get_variable('weights', [3, 3, 256, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias9 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve9 = tf.nn.conv2d(self.pool3, self.weights9, [1, 1, 1, 1], padding="SAME") 
       self.conv9 = tf.nn.elu(tf.nn.bias_add(self.convolve9, self.bias9)) 
      with tf.variable_scope("conv10"): 
       self.weights10 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias10 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve10 = tf.nn.conv2d(self.conv9, self.weights10, [1, 1, 1, 1], padding="SAME") 
       self.conv10 = tf.nn.elu(tf.nn.bias_add(self.convolve10, self.bias10)) 
      with tf.variable_scope("conv11"): 
       self.weights11 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias11 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve11 = tf.nn.conv2d(self.conv10, self.weights11, [1, 1, 1, 1], padding="SAME") 
       self.conv11 = tf.nn.elu(tf.nn.bias_add(self.convolve11, self.bias11)) 
      with tf.variable_scope("conv12"): 
       self.weights12 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias12 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve12 = tf.nn.conv2d(self.conv11, self.weights12, [1, 1, 1, 1], padding="SAME") 
       self.conv12 = tf.nn.elu(tf.nn.bias_add(self.convolve12, self.bias12)) 
      self.pool4 = tf.nn.max_pool(self.conv12, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME', name="pool4") 

     with tf.variable_scope("group5"): 
      with tf.variable_scope("conv13"): 
       self.weights13 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias13 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve13 = tf.nn.conv2d(self.pool4, self.weights13, [1, 1, 1, 1], padding="SAME") 
       self.conv13 = tf.nn.elu(tf.nn.bias_add(self.convolve13, self.bias13)) 

      with tf.variable_scope("conv14"): 
       self.weights14 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias14 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve14 = tf.nn.conv2d(self.conv13, self.weights14, [1, 1, 1, 1], padding="SAME") 
       self.conv14 = tf.nn.elu(tf.nn.bias_add(self.convolve14, self.bias14)) 
      with tf.variable_scope("conv15"): 
       self.weights15 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias15 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve15 = tf.nn.conv2d(self.conv14, self.weights15, [1, 1, 1, 1], padding="SAME") 
       self.conv15 = tf.nn.elu(tf.nn.bias_add(self.convolve15, self.bias15)) 
      with tf.variable_scope("conv16"): 
       self.weights16 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias16 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve16 = tf.nn.conv2d(self.conv15, self.weights16, [1, 1, 1, 1], padding="SAME") 
       self.conv16 = tf.nn.elu(tf.nn.bias_add(self.convolve16, self.bias16)) 
      self.pool5 = tf.nn.max_pool(self.conv16, [1, 2, 2, 1], [1, 2, 2, 1], "SAME", name="pool5") 

     with tf.variable_scope("group6"): 
      self.fc6 = tf.reshape(self.pool5, [-1, 4096], 'fc6') 
      with tf.variable_scope("fc7"): 
       self.weights17 = tf.get_variable('weights', [4096, 4096], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias17 = tf.get_variable('bias', [4096], initializer=tf.constant_initializer(0.0)) 
       self.fc7 = tf.nn.elu(tf.nn.bias_add(tf.matmul(self.fc6, self.weights17), self.bias17)) 
      with tf.variable_scope("logits"): 
       self.weights17 = tf.get_variable('weights', [4096, n_classes], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias17 = tf.get_variable('bias', [n_classes], initializer=tf.constant_initializer(0.0)) 
       self.logits = tf.nn.elu(tf.nn.bias_add(tf.matmul(self.fc6, self.weights17), self.bias17)) 

     self.softmax = tf.nn.softmax(self.logits, 'softmax') 

我已經經歷過並重複閱讀論文並搜索了其他Tensorflow實現。在這一點上，我認爲我對事業失明。從self.pool5重塑成self.fc6時，我得到了一個錯誤的形狀錯誤。當它應該是（？，8，8，512）時，self.pool5以形狀（？，7，7，512）結束。任何幫助將不勝感激。

謝謝。

Answer 1

我覺得你的問題是，你試圖重塑pool5成完全連接層4096的大小。但是那些4096僅僅是fc層具有的神經元的數量，而不是其輸入的數量。我試圖相應地更改您的代碼。

 with tf.variable_scope("conv16"): 
       self.weights16 = tf.get_variable('weights', [3, 3, 512, 512], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias16 = tf.get_variable('bias', [512], initializer=tf.constant_initializer(0.0)) 
       self.convolve16 = tf.nn.conv2d(self.conv15, self.weights16, [1, 1, 1, 1], padding="SAME") 
       self.conv16 = tf.nn.elu(tf.nn.bias_add(self.convolve16, self.bias16)) 
      self.pool5 = tf.nn.max_pool(self.conv16, [1, 2, 2, 1], [1, 2, 2, 1], "SAME", name="pool5") 
      self.pool5_flatten = tf.reshape(self.pool5, [-1, 25088]) 

     with tf.variable_scope("group6"): 
      with tf.variable_scope("fc7"): 
       self.weights17 = tf.get_variable('weights', [25088, 4096], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias17 = tf.get_variable('bias', [4096], initializer=tf.constant_initializer(0.0)) 
       self.fc7 = tf.nn.elu(tf.nn.bias_add(tf.matmul(self.pool5_flatten, self.weights17), self.bias17)) 
      with tf.variable_scope("logits"): 
       self.weights17 = tf.get_variable('weights', [4096, n_classes], 
               initializer=tf.random_normal_initializer(stddev=1e-2)) 
       self.bias17 = tf.get_variable('bias', [n_classes], initializer=tf.constant_initializer(0.0)) 
       self.logits = tf.nn.elu(tf.nn.bias_add(tf.matmul(self.fc7, self.weights17), self.bias17)) 

     self.softmax = tf.nn.softmax(self.logits, 'softmax') 

Answer 2

這是完全正常的，你的實現如果沒問題。

VGG使用5最大池與步幅2和padding='SAME'，這將特徵映射大小降低高度2和寬度2（如果它可以被2整除）。

您指定大小224x224的輸入圖像：

self.input_data = tf.placeholder(dtype=tf.float32, name='input_data', shape=(5, 224, 224, 3)) 

因此，功能地圖的大小是由2^5=32分割，你到底得到形狀的輸出：[?, 7, 7, feature_size]（224/32 = 7）。

---

如果你真的想與VGG建築形狀[?, 8, 8, feature_size]的輸出尺寸，你應該形狀的圖像開始256 x 256