Week #18
25 Feb 2019For slice 4 and slice 5, The loss doesn’t decrease. Observing the unique values in mask, I find that the number of unique values for slice 4 and slice 5 is more average, i.e. The difference in the number of different classes is small. One example for instance: one parc_1 unique color number in slice 4:
While for slice 1, the class distribution is:
So to deal with the class imbalance, we should assign weight to these class labels. I initially considered to add weight to the dataloader but since the sampler can only be a sequence, where in our case is a 2 dimensional tensor, an alternative choice which is to add weight to NLLloss() directly is chosen.
Calculate the total frequency for axial plane
For axial plane, the unique colour and their frequency are illustrated as follows, this first tensor represents the class frequency, where we can notice that 0 is the most frequent appearance class label, it represents the background.
[841437103 10250923 7549988 6202351 5479058 4958986 4507302
4243786 3792552 3194501 2646184 2244682 1963318 1780359
1661430 1592262 1547413 1525394 1493846 1440255 1356328
1244865 1128676 1056876 964887 916613 856547 815654
780511 759421 728968 671533 613634 543049 480876
436910 410222 371976 339930 299092 296347 269747
261126 362197 274102 240017 232081 224661 226577
229913 239052 260643 287046 323339 368924 429290
475533 550669 628456 701645 794142 882573 3357467
988450 989575 979536 989692 988891 1002396 996119
1010202 1002309 1002004 1000883 993450 989924 985237
976607 966696 958122 946667 935270 919341 904095
891085 871166 852290 837076 802372 778569 750211
732621 699185 662531 627733 586727 560334 542006
522943 502830 491130 455715 446945 388742 361932
415303 323668 286551 273422 261092 249175 248550
249064 255377 267272 279651 297206 316893 333156
349060 363610 415263 427623 432991 2650231 666910
1237103 465004 361275 328196 296866 274032 245754
211556 192685 161724 134744 115166 98740 80831
67831 56935 46194 38591 31462 26456 20272
18050 13141 10713 7724 6303 4786 3788
2820 1950 2069 1372 936 668 628
347 329 174 125 68 41 42
22 11 4 1 1 3 4
9 8 1]
[ 0. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13.
14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27.
28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40. 41.
42. 43. 44. 45. 46. 47. 48. 49. 50. 51. 52. 53. 54. 55.
56. 57. 58. 59. 60. 61. 62. 63. 64. 65. 66. 67. 68. 69.
70. 71. 72. 73. 74. 75. 76. 77. 78. 79. 80. 81. 82. 83.
84. 85. 86. 87. 88. 89. 90. 92. 93. 94. 95. 96. 97. 98.
99. 100. 101. 102. 103. 104. 105. 106. 107. 108. 109. 110. 111. 112.
113. 114. 115. 116. 117. 118. 119. 120. 121. 122. 123. 124. 125. 126.
127. 128. 129. 130. 131. 132. 133. 134. 135. 136. 137. 138. 139. 140.
141. 142. 143. 144. 145. 146. 147. 148. 149. 150. 151. 152. 153. 154.
155. 156. 157. 158. 159. 160. 161. 162. 163. 164. 165. 166. 167. 168.
169. 170. 171. 172. 173. 174. 175. 176. 177. 180.]
Obeserving the results, we found that there are in total 178 classes for axial plane, with class label from 0 to 177, additional 180, excluding 91. Since form the dataloader the target label shouldn’t exceed the class number, so a class mapping is added, which map 180 to 91. so the final count is :
[841437103 10250923 7549988 6202351 5479058 4958986 4507302
4243786 3792552 3194501 2646184 2244682 1963318 1780359
1661430 1592262 1547413 1525394 1493846 1440255 1356328
1244865 1128676 1056876 964887 916613 856547 815654
780511 759421 728968 671533 613634 543049 480876
436910 410222 371976 339930 299092 296347 269747
261126 362197 274102 240017 232081 224661 226577
229913 239052 260643 287046 323339 368924 429290
475533 550669 628456 701645 794142 882573 3357467
988450 989575 979536 989692 988891 1002396 996119
1010202 1002309 1002004 1000883 993450 989924 985237
976607 966696 958122 946667 935270 919341 904095
891085 871166 852290 837076 802372 778569 750211
1 732621 699185 662531 627733 586727 560334
542006 522943 502830 491130 455715 446945 388742
361932 415303 323668 286551 273422 261092 249175
248550 249064 255377 267272 279651 297206 316893
333156 349060 363610 415263 427623 432991 2650231
666910 1237103 465004 361275 328196 296866 274032
245754 211556 192685 161724 134744 115166 98740
80831 67831 56935 46194 38591 31462 26456
20272 18050 13141 10713 7724 6303 4786
3788 2820 1950 2069 1372 936 668
628 347 329 174 125 68 41
42 22 11 4 1 1 3
4 9 8]
The weight is calculated weight = 1. / count.float()
Adding the weight to NLLloss, the loss doesn’t decrease as the same for the slice 0. Since in slice 0, most pixels are background. So the loss observed before might just represent the background.
Moreover, training cross subject may lead to inference in loss decrease process. Some sudden increase of loss has been observed.
I have trained the network on slice 0, slice 5 respectively, and with 10, 20, 50 training images respectively. The loss decrease as desired. With network’s parameter num = 13,177,026