碼上歡樂
相關內容    簡體    繁體

【轉】學習筆記：GoogLeNet

本文轉載自   查看原文   2015-07-06 21:11   14875 

原作者：lynnandwei 原文地址：http://blog.csdn.net/lynnandwei/article/details/44458033

GoogLeNet, 2014年ILSVRC挑戰賽冠軍，將Top5 的錯誤率降低到6.67%. 一個22層的深度網絡，論文在http://arxiv.org/pdf/1409.4842v1.pdf，題目為：Going deeper with convolutions。（每次看這么簡潔優雅的題目，就想吐槽國內寫paper的 八股文題目）。GoogLeNet這個名字也是挺有意思的，為了像開山鼻祖的LeNet網絡致敬，他們選擇了這樣的名字。

BVLC在caffe中給出了網絡的實現：https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet

模型下載地址：http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel

從論文里整理了一張這個22個層次的模型的圖出來（如果考慮pooling層是27層），先將模型跑了一遍結果，還是那只貓： 

直觀輸出是如下，覺得挺准確：

[287 281 285 282 283]
['n02127052 lynx, catamount' 'n02123045 tabby, tabby cat'
 'n02124075 Egyptian cat' 'n02123159 tiger cat' 'n02123394 Persian cat']

caffe的實現和原來論文的模型是有不同的：

not training with the relighting data-augmentation;
not training with the scale or aspect-ratio data-augmentation;
uses "xavier" to initialize the weights instead of "gaussian";
quick_solver.prototxt uses a different learning rate decay policy than the original solver.prototxt, that allows a much faster training (60 epochs vs 250 epochs);
The bundled model is the iteration 2,400,000 snapshot (60 epochs) using quick_solver.prototxt

但是准確度還是達到了 a top-1 accuracy 68.7% (31.3% error) and a top-5 accuracy 88.9% (11.1% error)

我們來分析一下這個模型的層次關系:

原始數據，輸入為224*224*3

第一層卷積層 conv1 ，pad是3，64個特征，7*7 步長為2，輸出特征為 112*112*64，然后進行relu，經過pool1（紅色的max pool） 進行pooling 3*3的核，步長為2， [（112 - 3+1）/2]+1 = 56  特征為56*56*64 ， 然后進行norm 

第二層卷積層 conv2， pad是1，3*3，192個特征，輸出為56*56*192，然后進行relu，進行norm，經過pool2進行pooling，3*3的核，步長為2 輸出為28*28*192 然后進行split 分成四個支線

 

[html]  view plain copy 在CODE上查看代碼片 派生到我的代碼片
 
  1. Setting up pool2/3x3_s2  
  2. I0319 23:50:37.405478  5765 net.cpp:103] Top shape: 10 192 28 28 (1505280)  
  3. I0319 23:50:37.405484  5765 net.cpp:113] Memory required for data: 174612480  
  4. I0319 23:50:37.405495  5765 net.cpp:67] Creating Layer pool2/3x3_s2_pool2/3x3_s2_0_split  
  5. I0319 23:50:37.405503  5765 net.cpp:394] pool2/3x3_s2_pool2/3x3_s2_0_split <- pool2/3x3_s2  
  6. I0319 23:50:37.405515  5765 net.cpp:356] pool2/3x3_s2_pool2/3x3_s2_0_split -> pool2/3x3_s2_pool2/3x3_s2_0_split_0  
  7. I0319 23:50:37.405531  5765 net.cpp:356] pool2/3x3_s2_pool2/3x3_s2_0_split -> pool2/3x3_s2_pool2/3x3_s2_0_split_1  
  8. I0319 23:50:37.405545  5765 net.cpp:356] pool2/3x3_s2_pool2/3x3_s2_0_split -> pool2/3x3_s2_pool2/3x3_s2_0_split_2  
  9. I0319 23:50:37.405557  5765 net.cpp:356] pool2/3x3_s2_pool2/3x3_s2_0_split -> pool2/3x3_s2_pool2/3x3_s2_0_split_3  
  10. I0319 23:50:37.405567  5765 net.cpp:96] Setting up pool2/3x3_s2_pool2/3x3_s2_0_split  
  11. I0319 23:50:37.405577  5765 net.cpp:103] Top shape: 10 192 28 28 (1505280)  
  12. I0319 23:50:37.405582  5765 net.cpp:103] Top shape: 10 192 28 28 (1505280)  
  13. I0319 23:50:37.405587  5765 net.cpp:103] Top shape: 10 192 28 28 (1505280)  
  14. I0319 23:50:37.405592  5765 net.cpp:103] Top shape: 10 192 28 28 (1505280)  
  15. I0319 23:50:37.405597  5765 net.cpp:113] Memory required for data: 198696960  
  16. I0319 23:50:37.405611  5765 net.cpp:67] Creating Layer inception_3a/1x1^M  


第三層開始時 inception module ，這個的思想受到使用不同尺度的Gabor過濾器來處理多尺度問題，inception module采用不同尺度的卷積核來處理問題。3a 包含 四個支線：

 

1： 64個1*1的卷積核（之后進行RULE計算） 變成28*28*64

2： 96個1*1的卷積核 作為3*3卷積核之前的reduce，變成28*28*96， 進行relu計算后，再進行128個3*3的卷積，pad為1， 28*28*128

3：16個1*1的卷積核 作為5*5卷積核之前的reduce，變成28*28*16， 進行relu計算后，再進行32個5*5的卷積，pad為2，變成28*28*32

4：pool層，3*3的核，pad為1，輸出還是28*28*192，然后進行32個1*1的卷積，變成28*28*32。

將四個結果進行連接，輸出為28*28*256

 

 

 

 

然后將3a的結果又分成四條支線，開始建立3b的inception module

3b 

 

1：128個1*1的卷積核（之后進行RULE計算） 變成28*28*128

2：128個1*1的卷積核 作為3*3卷積核之前的reduce，變成28*28*128， 再進行192個3*3的卷積，pad為1， 28*28*192，進行relu計算

3：32個1*1的卷積核 作為5*5卷積核之前的reduce，變成28*28*32， 進行relu計算后，再進行96個5*5的卷積，pad為2，變成28*28*96

4：pool層，3*3的核，pad為1，輸出還是28*28*256，然后進行64個1*1的卷積，變成28*28*64。

 

將四個結果進行連接，輸出為28*28*480

同理依次推算，數據變化如下表：

 

一部分輸出結果如下：

I0319 22:27:51.257917 5080 net.cpp:208] This network produces output prob

 

[html]  view plain copy 在CODE上查看代碼片 派生到我的代碼片
 
  1. I0319 22:27:51.258116  5080 net.cpp:467] Collecting Learning Rate and Weight Decay.  
  2. I0319 22:27:51.258162  5080 net.cpp:219] Network initialization done.  
  3. I0319 22:27:51.258167  5080 net.cpp:220] Memory required for data: 545512320  
  4. I0319 22:27:51.345417  5080 net.cpp:702] Ignoring source layer data  
  5. I0319 22:27:51.345443  5080 net.cpp:702] Ignoring source layer label_data_1_split  
  6. I0319 22:27:51.345448  5080 net.cpp:705] Copying source layer conv1/7x7_s2  
  7. I0319 22:27:51.345542  5080 net.cpp:705] Copying source layer conv1/relu_7x7  
  8. I0319 22:27:51.345548  5080 net.cpp:705] Copying source layer pool1/3x3_s2  
  9. I0319 22:27:51.345553  5080 net.cpp:705] Copying source layer pool1/norm1  
  10. I0319 22:27:51.345558  5080 net.cpp:705] Copying source layer conv2/3x3_reduce  
  11. I0319 22:27:51.345600  5080 net.cpp:705] Copying source layer conv2/relu_3x3_reduce  
  12. I0319 22:27:51.345607  5080 net.cpp:705] Copying source layer conv2/3x3  
  13. I0319 22:27:51.346571  5080 net.cpp:705] Copying source layer conv2/relu_3x3  
  14. I0319 22:27:51.346580  5080 net.cpp:705] Copying source layer conv2/norm2  
  15. I0319 22:27:51.346585  5080 net.cpp:705] Copying source layer pool2/3x3_s2  
  16. I0319 22:27:51.346590  5080 net.cpp:705] Copying source layer pool2/3x3_s2_pool2/3x3_s2_0_split  
  17. I0319 22:27:51.346595  5080 net.cpp:705] Copying source layer inception_3a/1x1  
  18. I0319 22:27:51.346706  5080 net.cpp:705] Copying source layer inception_3a/relu_1x1  
  19. I0319 22:27:51.346712  5080 net.cpp:705] Copying source layer inception_3a/3x3_reduce  
  20. I0319 22:27:51.346879  5080 net.cpp:705] Copying source layer inception_3a/relu_3x3_reduce  
  21. I0319 22:27:51.346885  5080 net.cpp:705] Copying source layer inception_3a/3x3  
  22. I0319 22:27:51.347844  5080 net.cpp:705] Copying source layer inception_3a/relu_3x3  
  23. I0319 22:27:51.347851  5080 net.cpp:705] Copying source layer inception_3a/5x5_reduce  
  24. I0319 22:27:51.347885  5080 net.cpp:705] Copying source layer inception_3a/relu_5x5_reduce  
  25. I0319 22:27:51.347892  5080 net.cpp:705] Copying source layer inception_3a/5x5  
  26. I0319 22:27:51.348008  5080 net.cpp:705] Copying source layer inception_3a/relu_5x5  
  27. I0319 22:27:51.348014  5080 net.cpp:705] Copying source layer inception_3a/pool  
  28. I0319 22:27:51.348019  5080 net.cpp:705] Copying source layer inception_3a/pool_proj  
  29. I0319 22:27:51.348080  5080 net.cpp:705] Copying source layer inception_3a/relu_pool_proj  
  30. I0319 22:27:51.348085  5080 net.cpp:705] Copying source layer inception_3a/output  
  31. I0319 22:27:51.348091  5080 net.cpp:705] Copying source layer inception_3a/output_inception_3a/output_0_split  
  32. I0319 22:27:51.348096  5080 net.cpp:705] Copying source layer inception_3b/1x1  
  33. I0319 22:27:51.348398  5080 net.cpp:705] Copying source layer inception_3b/relu_1x1  
  34. I0319 22:27:51.348405  5080 net.cpp:705] Copying source layer inception_3b/3x3_reduce  
  35. I0319 22:27:51.348700  5080 net.cpp:705] Copying source layer inception_3b/relu_3x3_reduce  
  36. I0319 22:27:51.348707  5080 net.cpp:705] Copying source layer inception_3b/3x3  
  37. I0319 22:27:51.350611  5080 net.cpp:705] Copying source layer inception_3b/relu_3x3  
  38. I0319 22:27:51.350620  5080 net.cpp:705] Copying source layer inception_3b/5x5_reduce  
  39. I0319 22:27:51.350699  5080 net.cpp:705] Copying source layer inception_3b/relu_5x5_reduce  
  40. I0319 22:27:51.350705  5080 net.cpp:705] Copying source layer inception_3b/5x5  
  41. I0319 22:27:51.351372  5080 net.cpp:705] Copying source layer inception_3b/relu_5x5  
  42. I0319 22:27:51.351378  5080 net.cpp:705] Copying source layer inception_3b/pool  
  43. I0319 22:27:51.351384  5080 net.cpp:705] Copying source layer inception_3b/pool_proj  
  44. I0319 22:27:51.351546  5080 net.cpp:705] Copying source layer inception_3b/relu_pool_proj  
  45. I0319 22:27:51.351552  5080 net.cpp:705] Copying source layer inception_3b/output  
  46. I0319 22:27:51.351558  5080 net.cpp:705] Copying source layer pool3/3x3_s2  
  47. I0319 22:27:51.351563  5080 net.cpp:705] Copying source layer pool3/3x3_s2_pool3/3x3_s2_0_split  
  48. I0319 22:27:51.351569  5080 net.cpp:705] Copying source layer inception_4a/1x1  
  49. I0319 22:27:51.352367  5080 net.cpp:705] Copying source layer inception_4a/relu_1x1  
  50. I0319 22:27:51.352375  5080 net.cpp:705] Copying source layer inception_4a/3x3_reduce  
  51. I0319 22:27:51.352782  5080 net.cpp:705] Copying source layer inception_4a/relu_3x3_reduce  
  52. I0319 22:27:51.352789  5080 net.cpp:705] Copying source layer inception_4a/3x3  
  53. I0319 22:27:51.354333  5080 net.cpp:705] Copying source layer inception_4a/relu_3x3  
  54. I0319 22:27:51.354341  5080 net.cpp:705] Copying source layer inception_4a/5x5_reduce  
  55. I0319 22:27:51.354420  5080 net.cpp:705] Copying source layer inception_4a/relu_5x5_reduce  
  56. I0319 22:27:51.354429  5080 net.cpp:705] Copying source layer inception_4a/5x5  
  57. I0319 22:27:51.354601  5080 net.cpp:705] Copying source layer inception_4a/relu_5x5  
  58. I0319 22:27:51.354609  5080 net.cpp:705] Copying source layer inception_4a/pool  
  59. I0319 22:27:51.354614  5080 net.cpp:705] Copying source layer inception_4a/pool_proj  
  60. I0319 22:27:51.354887  5080 net.cpp:705] Copying source layer inception_4a/relu_pool_proj  
  61. I0319 22:27:51.354894  5080 net.cpp:705] Copying source layer inception_4a/output  
  62. I0319 22:27:51.354900  5080 net.cpp:705] Copying source layer inception_4a/output_inception_4a/output_0_split  
  63. I0319 22:27:51.354910  5080 net.cpp:702] Ignoring source layer loss1/ave_pool  
  64. I0319 22:27:51.354918  5080 net.cpp:702] Ignoring source layer loss1/conv  
  65. I0319 22:27:51.354923  5080 net.cpp:702] Ignoring source layer loss1/relu_conv  
  66. I0319 22:27:51.354930  5080 net.cpp:702] Ignoring source layer loss1/fc  
  67. I0319 22:27:51.354936  5080 net.cpp:702] Ignoring source layer loss1/relu_fc  
  68. I0319 22:27:51.354943  5080 net.cpp:702] Ignoring source layer loss1/drop_fc  
  69. I0319 22:27:51.354950  5080 net.cpp:702] Ignoring source layer loss1/classifier  
  70. I0319 22:27:51.354956  5080 net.cpp:702] Ignoring source layer loss1/loss  
  71. I0319 22:27:51.354962  5080 net.cpp:705] Copying source layer inception_4b/1x1  
  72. I0319 22:27:51.355681  5080 net.cpp:705] Copying source layer inception_4b/relu_1x1  
  73. I0319 22:27:51.355690  5080 net.cpp:705] Copying source layer inception_4b/3x3_reduce  
  74. I0319 22:27:51.356190  5080 net.cpp:705] Copying source layer inception_4b/relu_3x3_reduce  
  75. I0319 22:27:51.356199  5080 net.cpp:705] Copying source layer inception_4b/3x3  
  76. I0319 22:27:51.358134  5080 net.cpp:705] Copying source layer inception_4b/relu_3x3  
  77. I0319 22:27:51.358144  5080 net.cpp:705] Copying source layer inception_4b/5x5_reduce  
  78. I0319 22:27:51.358256  5080 net.cpp:705] Copying source layer inception_4b/relu_5x5_reduce  
  79. I0319 22:27:51.358263  5080 net.cpp:705] Copying source layer inception_4b/5x5  
  80. I0319 22:27:51.358608  5080 net.cpp:705] Copying source layer inception_4b/relu_5x5  
  81. I0319 22:27:51.358616  5080 net.cpp:705] Copying source layer inception_4b/pool  
  82. I0319 22:27:51.358623  5080 net.cpp:705] Copying source layer inception_4b/pool_proj  
  83. I0319 22:27:51.358917  5080 net.cpp:705] Copying source layer inception_4b/relu_pool_proj  
  84. I0319 22:27:51.358925  5080 net.cpp:705] Copying source layer inception_4b/output  
  85. I0319 22:27:51.358932  5080 net.cpp:705] Copying source layer inception_4b/output_inception_4b/output_0_split  
  86. I0319 22:27:51.358937  5080 net.cpp:705] Copying source layer inception_4c/1x1  
  87. I0319 22:27:51.359519  5080 net.cpp:705] Copying source layer inception_4c/relu_1x1  
  88. I0319 22:27:51.359526  5080 net.cpp:705] Copying source layer inception_4c/3x3_reduce  
  89. I0319 22:27:51.360097  5080 net.cpp:705] Copying source layer inception_4c/relu_3x3_reduce  
  90. I0319 22:27:51.360105  5080 net.cpp:705] Copying source layer inception_4c/3x3  
  91. I0319 22:27:51.362634  5080 net.cpp:705] Copying source layer inception_4c/relu_3x3  
  92. I0319 22:27:51.362643  5080 net.cpp:705] Copying source layer inception_4c/5x5_reduce  
  93. I0319 22:27:51.362757  5080 net.cpp:705] Copying source layer inception_4c/relu_5x5_reduce  
  94. I0319 22:27:51.362764  5080 net.cpp:705] Copying source layer inception_4c/5x5  
  95. I0319 22:27:51.363106  5080 net.cpp:705] Copying source layer inception_4c/relu_5x5  
  96. I0319 22:27:51.363114  5080 net.cpp:705] Copying source layer inception_4c/pool  
  97. I0319 22:27:51.363121  5080 net.cpp:705] Copying source layer inception_4c/pool_proj  
  98. I0319 22:27:51.363415  5080 net.cpp:705] Copying source layer inception_4c/relu_pool_proj  
  99. I0319 22:27:51.363423  5080 net.cpp:705] Copying source layer inception_4c/output  
  100. I0319 22:27:51.363430  5080 net.cpp:705] Copying source layer inception_4c/output_inception_4c/output_0_split  
  101. I0319 22:27:51.363436  5080 net.cpp:705] Copying source layer inception_4d/1x1  
  102. I0319 22:27:51.363937  5080 net.cpp:705] Copying source layer inception_4d/relu_1x1  
  103. I0319 22:27:51.363945  5080 net.cpp:705] Copying source layer inception_4d/3x3_reduce  
  104. I0319 22:27:51.364591  5080 net.cpp:705] Copying source layer inception_4d/relu_3x3_reduce  
  105. I0319 22:27:51.364600  5080 net.cpp:705] Copying source layer inception_4d/3x3  
  106. I0319 22:27:51.367797  5080 net.cpp:705] Copying source layer inception_4d/relu_3x3  
  107. I0319 22:27:51.367806  5080 net.cpp:705] Copying source layer inception_4d/5x5_reduce  
  108. I0319 22:27:51.367959  5080 net.cpp:705] Copying source layer inception_4d/relu_5x5_reduce  
  109. I0319 22:27:51.367966  5080 net.cpp:705] Copying source layer inception_4d/5x5  
  110. I0319 22:27:51.368420  5080 net.cpp:705] Copying source layer inception_4d/relu_5x5  
  111. I0319 22:27:51.368428  5080 net.cpp:705] Copying source layer inception_4d/pool  
  112. I0319 22:27:51.368435  5080 net.cpp:705] Copying source layer inception_4d/pool_proj  
  113. I0319 22:27:51.368726  5080 net.cpp:705] Copying source layer inception_4d/relu_pool_proj  
  114. I0319 22:27:51.368733  5080 net.cpp:705] Copying source layer inception_4d/output  
  115. I0319 22:27:51.368739  5080 net.cpp:705] Copying source layer inception_4d/output_inception_4d/output_0_split  
  116. I0319 22:27:51.368748  5080 net.cpp:702] Ignoring source layer loss2/ave_pool  
  117. I0319 22:27:51.368755  5080 net.cpp:702] Ignoring source layer loss2/conv  
  118. I0319 22:27:51.368762  5080 net.cpp:702] Ignoring source layer loss2/relu_conv  
  119. I0319 22:27:51.368768  5080 net.cpp:702] Ignoring source layer loss2/fc  
  120. I0319 22:27:51.368774  5080 net.cpp:702] Ignoring source layer loss2/relu_fc  
  121. I0319 22:27:51.368780  5080 net.cpp:702] Ignoring source layer loss2/drop_fc  
  122. I0319 22:27:51.368788  5080 net.cpp:702] Ignoring source layer loss2/classifier  
  123. I0319 22:27:51.368794  5080 net.cpp:702] Ignoring source layer loss2/loss  
  124. I0319 22:27:51.368800  5080 net.cpp:705] Copying source layer inception_4e/1x1  
  125. I0319 22:27:51.369971  5080 net.cpp:705] Copying source layer inception_4e/relu_1x1  
  126. I0319 22:27:51.369981  5080 net.cpp:705] Copying source layer inception_4e/3x3_reduce  
  127. I0319 22:27:51.370717  5080 net.cpp:705] Copying source layer inception_4e/relu_3x3_reduce  
  128. I0319 22:27:51.370725  5080 net.cpp:705] Copying source layer inception_4e/3x3  
  129. I0319 22:27:51.374668  5080 net.cpp:705] Copying source layer inception_4e/relu_3x3  
  130. I0319 22:27:51.374678  5080 net.cpp:705] Copying source layer inception_4e/5x5_reduce  
  131. I0319 22:27:51.374831  5080 net.cpp:705] Copying source layer inception_4e/relu_5x5_reduce  
  132. I0319 22:27:51.374840  5080 net.cpp:705] Copying source layer inception_4e/5x5  
  133. I0319 22:27:51.375728  5080 net.cpp:705] Copying source layer inception_4e/relu_5x5  
  134. I0319 22:27:51.375737  5080 net.cpp:705] Copying source layer inception_4e/pool  
  135. I0319 22:27:51.375744  5080 net.cpp:705] Copying source layer inception_4e/pool_proj  
  136. I0319 22:27:51.376332  5080 net.cpp:705] Copying source layer inception_4e/relu_pool_proj  
  137. I0319 22:27:51.376340  5080 net.cpp:705] Copying source layer inception_4e/output  
  138. I0319 22:27:51.376346  5080 net.cpp:705] Copying source layer pool4/3x3_s2  
  139. I0319 22:27:51.376353  5080 net.cpp:705] Copying source layer pool4/3x3_s2_pool4/3x3_s2_0_split  
  140. I0319 22:27:51.376359  5080 net.cpp:705] Copying source layer inception_5a/1x1  
  141. I0319 22:27:51.378186  5080 net.cpp:705] Copying source layer inception_5a/relu_1x1  
  142. I0319 22:27:51.378196  5080 net.cpp:705] Copying source layer inception_5a/3x3_reduce  
  143. I0319 22:27:51.379343  5080 net.cpp:705] Copying source layer inception_5a/relu_3x3_reduce  
  144. I0319 22:27:51.379353  5080 net.cpp:705] Copying source layer inception_5a/3x3  
  145. I0319 22:27:51.383293  5080 net.cpp:705] Copying source layer inception_5a/relu_3x3  
  146. I0319 22:27:51.383303  5080 net.cpp:705] Copying source layer inception_5a/5x5_reduce  
  147. I0319 22:27:51.383548  5080 net.cpp:705] Copying source layer inception_5a/relu_5x5_reduce  
  148. I0319 22:27:51.383558  5080 net.cpp:705] Copying source layer inception_5a/5x5  
  149. I0319 22:27:51.384449  5080 net.cpp:705] Copying source layer inception_5a/relu_5x5  
  150. I0319 22:27:51.384459  5080 net.cpp:705] Copying source layer inception_5a/pool  
  151. I0319 22:27:51.384465  5080 net.cpp:705] Copying source layer inception_5a/pool_proj  
  152. I0319 22:27:51.385397  5080 net.cpp:705] Copying source layer inception_5a/relu_pool_proj  
  153. I0319 22:27:51.385406  5080 net.cpp:705] Copying source layer inception_5a/output  
  154. I0319 22:27:51.385414  5080 net.cpp:705] Copying source layer inception_5a/output_inception_5a/output_0_split  
  155. I0319 22:27:51.385421  5080 net.cpp:705] Copying source layer inception_5b/1x1  
  156. I0319 22:27:51.388157  5080 net.cpp:705] Copying source layer inception_5b/relu_1x1  
  157. I0319 22:27:51.388166  5080 net.cpp:705] Copying source layer inception_5b/3x3_reduce  
  158. I0319 22:27:51.389539  5080 net.cpp:705] Copying source layer inception_5b/relu_3x3_reduce  
  159. I0319 22:27:51.389549  5080 net.cpp:705] Copying source layer inception_5b/3x3  
  160. I0319 22:27:51.395212  5080 net.cpp:705] Copying source layer inception_5b/relu_3x3  
  161. I0319 22:27:51.395225  5080 net.cpp:705] Copying source layer inception_5b/5x5_reduce  
  162. I0319 22:27:51.395584  5080 net.cpp:705] Copying source layer inception_5b/relu_5x5_reduce  
  163. I0319 22:27:51.395594  5080 net.cpp:705] Copying source layer inception_5b/5x5  
  164. I0319 22:27:51.396921  5080 net.cpp:705] Copying source layer inception_5b/relu_5x5  
  165. I0319 22:27:51.396931  5080 net.cpp:705] Copying source layer inception_5b/pool  
  166. I0319 22:27:51.396939  5080 net.cpp:705] Copying source layer inception_5b/pool_proj  
  167. I0319 22:27:51.397862  5080 net.cpp:705] Copying source layer inception_5b/relu_pool_proj  
  168. I0319 22:27:51.397871  5080 net.cpp:705] Copying source layer inception_5b/output  
  169. I0319 22:27:51.397879  5080 net.cpp:705] Copying source layer pool5/7x7_s1  
  170. I0319 22:27:51.397886  5080 net.cpp:705] Copying source layer pool5/drop_7x7_s1  
  171. I0319 22:27:51.397893  5080 net.cpp:705] Copying source layer loss3/classifier  
  172. I0319 22:27:51.406652  5080 net.cpp:702] Ignoring source layer loss3/loss3  

 

材料集合：

http://deeplearning.net/2014/09/19/googles-entry-to-imagenet-2014-challenge/

 

[1] Imagenet 2014 LSVRC results, http://karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/,Last retrieved on: 19-09-2014.

[2] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich, Going Deeper with Convolutions, Arxiv Link: http://arxiv.org/abs/1409.4842.

[3] GoogLeNet presentation, http://image-net.org/challenges/LSVRC/2014/slides/GoogLeNet.pptx, Last retrieved on: 19-09.2014..

[4] What I learned from competing against a convnet on imagenet, http://karpathy.github.io/2014/09/02/what-i-learned-from-competing-against-a-convnet-on-imagenet/, Last retrieved on: 19-09-2014.

[5] Girshick, Ross, et al. “Rich feature hierarchies for accurate object detection and semantic segmentation.” arXiv preprint arXiv:1311.2524 (2013).

 

 

 

版權聲明：本文為博主原創文章，未經博主允許不得轉載。


免責聲明！

本站轉載的文章為個人學習借鑒使用，本站對版權不負任何法律責任。如果侵犯了您的隱私權益，請聯系本站郵箱yoyou2525@163.com刪除。



猜您在找 tensorflow學習筆記——GoogLeNet GoogLeNet學習心得 GoJS學習筆記 (轉) NuGet學習筆記（轉） extjs學習筆記（轉） MiniUI學習筆記一【轉】 【轉】PCI學習筆記 TortoiseHg 學習筆記 （轉） GLSL學習筆記 [轉] FPGA學習筆記（轉）
 
粵ICP備18138465號   © 2018-2025 CODEPRJ.COM