• 快速抽取resnet_v2_152中间的特征层

    一、背景

            用resnet提取对象特征时，往往中间层也会包含一些细节信息，所以有时候会将中间层的特征抽取出来，本文将介绍如何抽取中间层的特征。

    二、准备代码

    1. import tensorflow as tf
    2.  
    3. from tensorflow.contrib import slim
    4. from tensorflow.contrib.slim.nets import resnet_v2
    5.  
    6. def resnet_arg_scope(self, is_training=True,  # 训练标记
    7.                      weight_decay=0.0001,  # 权重衰减速率
    8.                      batch_norm_decay=0.997,  # BN的衰减速率
    9.                      batch_norm_epsilon=1e-5,  # BN的epsilon默认1e-5
    10.                      batch_norm_scale=True):  # BN的scale默认值
    11.  
    12.     batch_norm_params = {  # 定义batch normalization（标准化）的参数字典
    13.         'is_training': is_training,
    14.         'decay': batch_norm_decay,
    15.         'epsilon': batch_norm_epsilon,
    16.         'scale': batch_norm_scale,
    17.         'updates_collections': tf.GraphKeys.UPDATE_OPS,
    18.     }
    19.  
    20.     with slim.arg_scope(  # 通过slim.arg_scope将[slim.conv2d]的几个默认参数设置好
    21.             [slim.conv2d],
    22.             weights_regularizer=slim.l2_regularizer(weight_decay),  # 权重正则器设置为L2正则
    23.             weights_initializer=slim.variance_scaling_initializer(),  # 权重初始化器
    24.             activation_fn=tf.nn.relu,  # 激活函数
    25.             normalizer_fn=slim.batch_norm,  # 标准化器设置为BN
    26.             normalizer_params=batch_norm_params):
    27.         with slim.arg_scope([slim.batch_norm], **batch_norm_params):
    28.             with slim.arg_scope([slim.max_pool2d], padding='SAME') as arg_sc:  # ResNet原论文是VALID模式，SAME模式可让特征对齐更简单
    29.                     return arg_sc
    30.  
    31. def _image_extract_resnet(image):
    32.     with slim.arg_scope(resnet_arg_scope(is_training=True)):
    33.         net, end_points = resnet_v2.resnet_v2_152(image, reuse=tf.AUTO_REUSE)

            以上代码是调用resnet152的过程，抽取特征则只涉及到end_points的信息。

    三、resnet-152的所有中间层名称及其大小

            直接对endpoint进行输出，然后整理一下就可以得到所有中间层的信息。整理的结果如下：

    1. ('resnet_v2_152/conv1', shape=(1, 112, 112, 64)), 
    2.  
    3. ('resnet_v2_152/block1/unit_1/bottleneck_v2/shortcut', shape=(1, 56, 56, 256)), 
    4. ('resnet_v2_152/block1/unit_1/bottleneck_v2/conv1', shape=(1, 56, 56, 64)), 
    5. ('resnet_v2_152/block1/unit_1/bottleneck_v2/conv2', shape=(1, 56, 56, 64)), 
    6. ('resnet_v2_152/block1/unit_1/bottleneck_v2/conv3', shape=(1, 56, 56, 256)), 
    7. ('resnet_v2_152/block1/unit_1/bottleneck_v2', shape=(1, 56, 56, 256)), 
    8. ('resnet_v2_152/block1/unit_2/bottleneck_v2/conv1', shape=(1, 56, 56, 64)), 
    9. ('resnet_v2_152/block1/unit_2/bottleneck_v2/conv2', shape=(1, 56, 56, 64)), 
    10. ('resnet_v2_152/block1/unit_2/bottleneck_v2/conv3', shape=(1, 56, 56, 256)), 
    11. ('resnet_v2_152/block1/unit_2/bottleneck_v2', shape=(1, 56, 56, 256)), 
    12. ('resnet_v2_152/block1/unit_3/bottleneck_v2/conv1', shape=(1, 56, 56, 64)), 
    13. ('resnet_v2_152/block1/unit_3/bottleneck_v2/conv2', shape=(1, 28, 28, 64)), 
    14. ('resnet_v2_152/block1/unit_3/bottleneck_v2/conv3', shape=(1, 28, 28, 256)), 
    15. ('resnet_v2_152/block1/unit_3/bottleneck_v2', shape=(1, 28, 28, 256)), 
    16. ('resnet_v2_152/block1', shape=(1, 28, 28, 256)), 
    17.  
    18. ('resnet_v2_152/block2/unit_1/bottleneck_v2/shortcut', shape=(1, 28, 28, 512)), 
    19. ('resnet_v2_152/block2/unit_1/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    20. ('resnet_v2_152/block2/unit_1/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    21. ('resnet_v2_152/block2/unit_1/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    22. ('resnet_v2_152/block2/unit_1/bottleneck_v2', shape=(1, 28, 28, 512)), 
    23. ('resnet_v2_152/block2/unit_2/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    24. ('resnet_v2_152/block2/unit_2/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    25. ('resnet_v2_152/block2/unit_2/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    26. ('resnet_v2_152/block2/unit_2/bottleneck_v2', shape=(1, 28, 28, 512)), 
    27. ('resnet_v2_152/block2/unit_3/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    28. ('resnet_v2_152/block2/unit_3/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    29. ('resnet_v2_152/block2/unit_3/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    30. ('resnet_v2_152/block2/unit_3/bottleneck_v2', shape=(1, 28, 28, 512)), 
    31. ('resnet_v2_152/block2/unit_4/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    32. ('resnet_v2_152/block2/unit_4/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    33. ('resnet_v2_152/block2/unit_4/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    34. ('resnet_v2_152/block2/unit_4/bottleneck_v2', shape=(1, 28, 28, 512)), 
    35. ('resnet_v2_152/block2/unit_5/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    36. ('resnet_v2_152/block2/unit_5/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    37. ('resnet_v2_152/block2/unit_5/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    38. ('resnet_v2_152/block2/unit_5/bottleneck_v2', shape=(1, 28, 28, 512)), 
    39. ('resnet_v2_152/block2/unit_6/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    40. ('resnet_v2_152/block2/unit_6/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    41. ('resnet_v2_152/block2/unit_6/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    42. ('resnet_v2_152/block2/unit_6/bottleneck_v2', shape=(1, 28, 28, 512)), 
    43. ('resnet_v2_152/block2/unit_7/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    44. ('resnet_v2_152/block2/unit_7/bottleneck_v2/conv2', shape=(1, 28, 28, 128)), 
    45. ('resnet_v2_152/block2/unit_7/bottleneck_v2/conv3', shape=(1, 28, 28, 512)), 
    46. ('resnet_v2_152/block2/unit_7/bottleneck_v2', shape=(1, 28, 28, 512)), 
    47. ('resnet_v2_152/block2/unit_8/bottleneck_v2/conv1', shape=(1, 28, 28, 128)), 
    48. ('resnet_v2_152/block2/unit_8/bottleneck_v2/conv2', shape=(1, 14, 14, 128)), 
    49. ('resnet_v2_152/block2/unit_8/bottleneck_v2/conv3', shape=(1, 14, 14, 512)), 
    50. ('resnet_v2_152/block2/unit_8/bottleneck_v2', shape=(1, 14, 14, 512)), 
    51. ('resnet_v2_152/block2', shape=(1, 14, 14, 512)), 
    52.  
    53. ('resnet_v2_152/block3/unit_1/bottleneck_v2/shortcut', shape=(1, 14, 14, 1024)), 
    54. ('resnet_v2_152/block3/unit_1/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    55. ('resnet_v2_152/block3/unit_1/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    56. ('resnet_v2_152/block3/unit_1/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    57. ('resnet_v2_152/block3/unit_1/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    58. ('resnet_v2_152/block3/unit_2/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    59. ('resnet_v2_152/block3/unit_2/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    60. ('resnet_v2_152/block3/unit_2/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    61. ('resnet_v2_152/block3/unit_2/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    62. ('resnet_v2_152/block3/unit_3/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    63. ('resnet_v2_152/block3/unit_3/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    64. ('resnet_v2_152/block3/unit_3/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    65. ('resnet_v2_152/block3/unit_3/bottleneck_v2', shape=(1, 14, 14, 1024)),
    66. ('resnet_v2_152/block3/unit_4/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    67. ('resnet_v2_152/block3/unit_4/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    68. ('resnet_v2_152/block3/unit_4/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    69. ('resnet_v2_152/block3/unit_4/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    70. ('resnet_v2_152/block3/unit_5/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    71. ('resnet_v2_152/block3/unit_5/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    72. ('resnet_v2_152/block3/unit_5/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    73. ('resnet_v2_152/block3/unit_5/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    74. ('resnet_v2_152/block3/unit_6/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    75. ('resnet_v2_152/block3/unit_6/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    76. ('resnet_v2_152/block3/unit_6/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    77. ('resnet_v2_152/block3/unit_6/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    78. ('resnet_v2_152/block3/unit_7/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    79. ('resnet_v2_152/block3/unit_7/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    80. ('resnet_v2_152/block3/unit_7/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    81. ('resnet_v2_152/block3/unit_7/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    82. ('resnet_v2_152/block3/unit_8/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    83. ('resnet_v2_152/block3/unit_8/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    84. ('resnet_v2_152/block3/unit_8/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    85. ('resnet_v2_152/block3/unit_8/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    86. ('resnet_v2_152/block3/unit_9/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    87. ('resnet_v2_152/block3/unit_9/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    88. ('resnet_v2_152/block3/unit_9/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    89. ('resnet_v2_152/block3/unit_9/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    90. ('resnet_v2_152/block3/unit_10/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    91. ('resnet_v2_152/block3/unit_10/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    92. ('resnet_v2_152/block3/unit_10/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    93. ('resnet_v2_152/block3/unit_10/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    94. ('resnet_v2_152/block3/unit_11/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    95. ('resnet_v2_152/block3/unit_11/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    96. ('resnet_v2_152/block3/unit_11/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    97. ('resnet_v2_152/block3/unit_11/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    98. ('resnet_v2_152/block3/unit_12/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    99. ('resnet_v2_152/block3/unit_12/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    100. ('resnet_v2_152/block3/unit_12/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    101. ('resnet_v2_152/block3/unit_12/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    102. ('resnet_v2_152/block3/unit_13/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    103. ('resnet_v2_152/block3/unit_13/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    104. ('resnet_v2_152/block3/unit_13/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    105. ('resnet_v2_152/block3/unit_13/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    106. ('resnet_v2_152/block3/unit_14/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    107. ('resnet_v2_152/block3/unit_14/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    108. ('resnet_v2_152/block3/unit_14/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    109. ('resnet_v2_152/block3/unit_14/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    110. ('resnet_v2_152/block3/unit_15/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    111. ('resnet_v2_152/block3/unit_15/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    112. ('resnet_v2_152/block3/unit_15/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    113. ('resnet_v2_152/block3/unit_15/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    114. ('resnet_v2_152/block3/unit_16/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    115. ('resnet_v2_152/block3/unit_16/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    116. ('resnet_v2_152/block3/unit_16/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    117. ('resnet_v2_152/block3/unit_16/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    118. ('resnet_v2_152/block3/unit_17/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    119. ('resnet_v2_152/block3/unit_17/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    120. ('resnet_v2_152/block3/unit_17/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    121. ('resnet_v2_152/block3/unit_17/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    122. ('resnet_v2_152/block3/unit_18/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    123. ('resnet_v2_152/block3/unit_18/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    124. ('resnet_v2_152/block3/unit_18/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    125. ('resnet_v2_152/block3/unit_18/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    126. ('resnet_v2_152/block3/unit_19/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    127. ('resnet_v2_152/block3/unit_19/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    128. ('resnet_v2_152/block3/unit_19/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    129. ('resnet_v2_152/block3/unit_19/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    130. ('resnet_v2_152/block3/unit_20/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    131. ('resnet_v2_152/block3/unit_20/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    132. ('resnet_v2_152/block3/unit_20/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    133. ('resnet_v2_152/block3/unit_20/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    134. ('resnet_v2_152/block3/unit_21/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    135. ('resnet_v2_152/block3/unit_21/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    136. ('resnet_v2_152/block3/unit_21/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    137. ('resnet_v2_152/block3/unit_21/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    138. ('resnet_v2_152/block3/unit_22/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    139. ('resnet_v2_152/block3/unit_22/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    140. ('resnet_v2_152/block3/unit_22/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    141. ('resnet_v2_152/block3/unit_22/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    142. ('resnet_v2_152/block3/unit_23/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    143. ('resnet_v2_152/block3/unit_23/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    144. ('resnet_v2_152/block3/unit_23/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    145. ('resnet_v2_152/block3/unit_23/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    146. ('resnet_v2_152/block3/unit_24/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    147. ('resnet_v2_152/block3/unit_24/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    148. ('resnet_v2_152/block3/unit_24/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    149. ('resnet_v2_152/block3/unit_24/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    150. ('resnet_v2_152/block3/unit_25/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    151. ('resnet_v2_152/block3/unit_25/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    152. ('resnet_v2_152/block3/unit_25/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    153. ('resnet_v2_152/block3/unit_25/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    154. ('resnet_v2_152/block3/unit_26/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    155. ('resnet_v2_152/block3/unit_26/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    156. ('resnet_v2_152/block3/unit_26/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    157. ('resnet_v2_152/block3/unit_26/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    158. ('resnet_v2_152/block3/unit_27/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    159. ('resnet_v2_152/block3/unit_27/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    160. ('resnet_v2_152/block3/unit_27/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    161. ('resnet_v2_152/block3/unit_27/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    162. ('resnet_v2_152/block3/unit_28/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    163. ('resnet_v2_152/block3/unit_28/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    164. ('resnet_v2_152/block3/unit_28/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    165. ('resnet_v2_152/block3/unit_28/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    166. ('resnet_v2_152/block3/unit_29/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    167. ('resnet_v2_152/block3/unit_29/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    168. ('resnet_v2_152/block3/unit_29/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    169. ('resnet_v2_152/block3/unit_29/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    170. ('resnet_v2_152/block3/unit_30/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    171. ('resnet_v2_152/block3/unit_30/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    172. ('resnet_v2_152/block3/unit_30/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    173. ('resnet_v2_152/block3/unit_30/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    174. ('resnet_v2_152/block3/unit_31/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    175. ('resnet_v2_152/block3/unit_31/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    176. ('resnet_v2_152/block3/unit_31/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    177. ('resnet_v2_152/block3/unit_31/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    178. ('resnet_v2_152/block3/unit_32/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    179. ('resnet_v2_152/block3/unit_32/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    180. ('resnet_v2_152/block3/unit_32/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    181. ('resnet_v2_152/block3/unit_32/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    182. ('resnet_v2_152/block3/unit_33/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    183. ('resnet_v2_152/block3/unit_33/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    184. ('resnet_v2_152/block3/unit_33/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    185. ('resnet_v2_152/block3/unit_33/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    186. ('resnet_v2_152/block3/unit_34/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    187. ('resnet_v2_152/block3/unit_34/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    188. ('resnet_v2_152/block3/unit_34/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)), 
    189. ('resnet_v2_152/block3/unit_34/bottleneck_v2', shape=(1, 14, 14, 1024)), 
    190. ('resnet_v2_152/block3/unit_35/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    191. ('resnet_v2_152/block3/unit_35/bottleneck_v2/conv2', shape=(1, 14, 14, 256)), 
    192. ('resnet_v2_152/block3/unit_35/bottleneck_v2/conv3', shape=(1, 14, 14, 1024)),
    193. ('resnet_v2_152/block3/unit_35/bottleneck_v2',  shape=(1, 14, 14, 1024)), 
    194. ('resnet_v2_152/block3/unit_36/bottleneck_v2/conv1', shape=(1, 14, 14, 256)), 
    195. ('resnet_v2_152/block3/unit_36/bottleneck_v2/conv2', shape=(1, 7, 7, 256)), 
    196. ('resnet_v2_152/block3/unit_36/bottleneck_v2/conv3', shape=(1, 7, 7, 1024)), 
    197. ('resnet_v2_152/block3/unit_36/bottleneck_v2', shape=(1, 7, 7, 1024)), 
    198. ('resnet_v2_152/block3', shape=(1, 7, 7, 1024)), 
    199.  
    200. ('resnet_v2_152/block4/unit_1/bottleneck_v2/shortcut', shape=(1, 7, 7, 2048)), 
    201. ('resnet_v2_152/block4/unit_1/bottleneck_v2/conv1', shape=(1, 7, 7, 512)), 
    202. ('resnet_v2_152/block4/unit_1/bottleneck_v2/conv2', shape=(1, 7, 7, 512)), 
    203. ('resnet_v2_152/block4/unit_1/bottleneck_v2/conv3', shape=(1, 7, 7, 2048)), 
    204. ('resnet_v2_152/block4/unit_1/bottleneck_v2', shape=(1, 7, 7, 2048)), 
    205. ('resnet_v2_152/block4/unit_2/bottleneck_v2/conv1', shape=(1, 7, 7, 512)), 
    206. ('resnet_v2_152/block4/unit_2/bottleneck_v2/conv2', shape=(1, 7, 7, 512)), 
    207. ('resnet_v2_152/block4/unit_2/bottleneck_v2/conv3', shape=(1, 7, 7, 2048)), 
    208. ('resnet_v2_152/block4/unit_2/bottleneck_v2', shape=(1, 7, 7, 2048)), 
    209. ('resnet_v2_152/block4/unit_3/bottleneck_v2/conv1', shape=(1, 7, 7, 512)), 
    210. ('resnet_v2_152/block4/unit_3/bottleneck_v2/conv2', shape=(1, 7, 7, 512)), 
    211. ('resnet_v2_152/block4/unit_3/bottleneck_v2/conv3', shape=(1, 7, 7, 2048)), 
    212. ('resnet_v2_152/block4/unit_3/bottleneck_v2', shape=(1, 7, 7, 2048)), 
    213. ('resnet_v2_152/block4', shape=(1, 7, 7, 2048))])

    如果想要调用中间的某一层可以如下设置：

    1. def _image_extract_resnet(self, image, label, driven=False):
    2.     # 图像特征处理模块
    3.     with slim.arg_scope(self.resnet_arg_scope(is_training=True)):
    4.         net, end_points = resnet_v2.resnet_v2_152(image, reuse=tf.AUTO_REUSE)
    5.  
    6.         # 想要别的层的话替换[]中的内容就可以了
    7.         feature = end_points['resnet_v2_152/conv1']
    8.  
    9.     return net, feature

  • 相关阅读:
    JAVA实现SAP接口
    【ChatGPT】无需代理使用ChatGPT
    傻妞and青龙面板对接短信登录【2022.7.4】
    拾壹博客拆解改造，页面元素替换（二）
    Animate动画
    Java线程未捕获异常处理 UncaughtExceptionHandler
    【接口测试】微信小程序接口自动化的实现方案
    vsCode设置文件注释和函数注释使用koroFileHeader插件
    磷脂-聚乙二醇-醛基 DSPE-PEG-Aldehyde DSPE-PEG-CHO MW：5000
    某建筑集团中层管理人员竞聘上岗项目成功案例纪实
  • 原文地址：https://blog.csdn.net/z704630835/article/details/126059288