修改MTCNN中caffe源码，时计算loss时以对应标签是否为“-1”判别（转）

MTCNN训练不收敛原因:

地址： https://github.com/dlunion/mtcnn

我们的训练数据标签格式：

wider face:

pos/001.jpg  1 x1 y1 x2 y2 (x1 y1 x2 y2) -1 -1 -1 -1 -1 -1 -1 -1 -1 -1

part/001.jpg -1 x1 y1 x2 y2 (x1 y1 x2 y2) -1 -1 -1 -1 -1 -1 -1 -1 -1 -1

neg/001.jpg  0 -1 -1 -1 -1 (x1 y1 x2 y2) -1 -1 -1 -1 -1 -1 -1 -1 -1 -1

 

celebA:

landmark/001.jpg -1 -1 -1 -1 -1 pst1_x pst1_y pst2_x pst2_y pst3_x pst3_y pst4_x pst4_y pst5_x pst5_y

 

作者要求的训练数据标签格式：

pos/001.jpg  1 x1 y1 x2 y2 (x1 y1 x2 y2) pst1_x pst1_y pst2_x pst2_y pst3_x pst3_y pst4_x pst4_y pst5_x pst5_y

part/001.jpg -1 x1 y1 x2 y2 (x1 y1 x2 y2) pst1_x pst1_y pst2_x pst2_y pst3_x pst3_y pst4_x pst4_y pst5_x pst5_y

neg/001.jpg  0 -1 -1 -1 -1 (x1 y1 x2 y2) pst1_x pst1_y pst2_x pst2_y pst3_x pst3_y pst4_x pst4_y pst5_x pst5_y

 

在“pts_loss”层（type: "MTCNNEuclideanLoss"）中，以"label"(分类的标签)来判断是否ignore。对于我们的训练数据标签格式：

class: ignore_label=-1, 可以正常分类；

bbox regression: ignore_label=0, 有landmark中-1参加计算，导致loss无法收敛；

landmark: ignore_label=0, 有part中-1参加计算，导致loss无法收敛；

 

解决思路：

在做class，bbox regression，landmark任务时，判断便签值是否全部为—1，来作为ignore条件。

 

修改后"MTCNNEuclideanLoss.cpp"如下：

1. #include

2.  

3. #include "caffe/layers/mtcnn_euclidean_loss_layer.hpp"

4. #include "caffe/util/math_functions.hpp"

5.  

6. #include

7. using namespace std;

8.  

9. namespace caffe {

10.  

11. template

12. void MTCNNEuclideanLossLayer::Reshape(

13. const vector*>& bottom, const vector*>& top) {

14. LossLayer::Reshape(bottom, top);

15. CHECK_EQ(bottom[0]->count(1), bottom[1]->count(1))

16. << "Inputs must have the same dimension.";

17.  

18. int has_ignore_label = this->layer_param().loss_param().has_ignore_label();

19. if (has_ignore_label)

20. CHECK_EQ(bottom.size(), 3) << "has_ignore_label=true but not input label";

21.  

22. if (!has_ignore_label)

23. CHECK_EQ(bottom.size(), 2) << "has_ignore_label=false but input mismatch";

24.  

25. diff_.ReshapeLike(*bottom[0]);

26. }

27.  

28. template

29. void MTCNNEuclideanLossLayer::Forward_cpu(const vector*>& bottom,

30. const vector*>& top) {

31. int count = bottom[0]->count();

32. int has_ignore_label = this->layer_param().loss_param().has_ignore_label();

33. int ignore_label = has_ignore_label ? this->layer_param().loss_param().ignore_label() : -1;

34.  

35. if (has_ignore_label){

36. const Dtype* label = bottom[2]->cpu_data();

37. int countLabel = bottom[2]->num();

38.  

39. //label

40. Dtype* diff = diff_.mutable_cpu_data();

41. int channel = bottom[0]->channels();

42. //cout << "countLabel_forward: " << countLabel << endl;

43. //cout << "channel_forward: " << channel << endl;

44. //cout << "ignore_label_forward: " << ignore_label << endl;

45. memset(diff, 0, sizeof(Dtype)*count);

46.  

47. const Dtype* b0 = bottom[0]->cpu_data();

48. const Dtype* b1 = bottom[1]->cpu_data();

49. Dtype loss = 0;

50.  

51. // bbox regression

52. if (channel == 4)

53. {

54. for (int i = 0; i < countLabel; ++i)

55. {

56. //cout << "forware_b1_4: " << b1[i*channel + 0] << " " << b1[i*channel + 1] << " " << b1[i*channel + 2] << " " << b1[i*channel + 3] << endl;

57.  

58. int dec = (b1[i*channel + 0] != ignore_label) && (b1[i*channel + 1] != ignore_label) && (b1[i*channel + 2] != ignore_label) && (b1[i*channel + 3] != ignore_label);

59. if ( dec==1 )

60. {

61. caffe_sub(

62. channel,

63. b0 + i * channel,

64. b1 + i * channel,

65. diff + i * channel);

66. Dtype dot = caffe_cpu_dot(channel, diff + i * channel, diff + i * channel);

67. loss += dot / Dtype(2);

68.  

69. //cout << "forware_b1_4: " << b1[i*channel + 0] << " " << b1[i*channel + 1] << " " << b1[i*channel + 2] << " " << b1[i*channel + 3] << endl;

70. }

71. }

72. }

73.  

74. // landmark

75. else if (channel == 10)

76. {

77. for (int i = 0; i < countLabel; ++i)

78. {

79. //cout << "forward_b1_10: " << b1[i*channel + 0] << " " << b1[i*channel + 1] << " " << b1[i*channel + 2] << " " << b1[i*channel + 3] << " " << b1[i*channel + 4] << " ";

80. //cout << b1[i*channel + 5] << " " << b1[i*channel + 6] << " " << b1[i*channel + 7] << " " << b1[i*channel + 8] << " " << b1[i*channel + 9] << endl;

81.  

82. int dec1 = (b1[i*channel + 0] != ignore_label) && (b1[i*channel + 1] != ignore_label) && (b1[i*channel + 2] != ignore_label) && (b1[i*channel + 3] != ignore_label) && (b1[i*channel + 4] != ignore_label);

83. int dec2 = (b1[i*channel + 5] != ignore_label) && (b1[i*channel + 6] != ignore_label) && (b1[i*channel + 7] != ignore_label) && (b1[i*channel + 8] != ignore_label) && (b1[i*channel + 9] != ignore_label);

84. if (dec1==1 && dec2==1)

85. {

86. caffe_sub(

87. channel,

88. b0 + i * channel,

89. b1 + i * channel,

90. diff + i * channel);

91. Dtype dot = caffe_cpu_dot(channel, diff + i * channel, diff + i * channel);

92. loss += dot / Dtype(2);

93.  

94. //cout << "forward_b1_10: " << b1[i*channel + 0] << " " << b1[i*channel + 1] << " " << b1[i*channel + 2] << " " << b1[i*channel + 3] << " " << b1[i*channel + 4] << " ";

95. //cout << b1[i*channel + 5] << " " << b1[i*channel + 6] << " " << b1[i*channel + 7] << " " << b1[i*channel + 8] << " " << b1[i*channel + 9] << endl;

96. }

97. }

98. }

99.  

100. // ****************org data ********************

101. //for (int i = 0; i < countLabel; ++i){

102. // if (label[i] != ignore_label){

103. // caffe_sub(

104. // channel,

105. // b0 + i * channel,

106. // b1 + i * channel,

107. // diff + i * channel);

108. // Dtype dot = caffe_cpu_dot(channel, diff + i * channel, diff + i * channel);

109. // loss += dot / Dtype(2);

110. // }

111. //}

112. // ***************** ********************

113.  

114. top[0]->mutable_cpu_data()[0] = loss;

115. }

116. else{

117. caffe_sub(

118. count,

119. bottom[0]->cpu_data(),

120. bottom[1]->cpu_data(),

121. diff_.mutable_cpu_data());

122. Dtype dot = caffe_cpu_dot(count, diff_.cpu_data(), diff_.cpu_data());

123. Dtype loss = dot / bottom[0]->num() / Dtype(2);

124. top[0]->mutable_cpu_data()[0] = loss;

125. }

126. }

127.  

128. template

129. void MTCNNEuclideanLossLayer::Backward_cpu(const vector*>& top,

130. const vector& propagate_down, const vector*>& bottom) {

131.  

132. int has_ignore_label = this->layer_param().loss_param().has_ignore_label();

133. int ignore_label = has_ignore_label ? this->layer_param().loss_param().ignore_label() : -1;

134.  

135. if (has_ignore_label){

136. const Dtype* b1 = bottom[1]->cpu_data();

137. const Dtype* label = bottom[2]->cpu_data();

138. int countLabel = bottom[2]->num();

139. int channels = bottom[0]->channels();

140. //cout << "countLabel_backword: " << countLabel << endl;

141. //cout << "channels_backword: " << channels << endl;

142. //cout << "ignore_label_backword: " << ignore_label << endl;

143. for (int i = 0; i < 2; ++i) {

144. if (propagate_down[i]) {

145. memset(bottom[i]->mutable_cpu_diff(), 0, sizeof(Dtype)*bottom[i]->count());

146.  

147. const Dtype sign = (i == 0) ? 1 : -1;

148. const Dtype alpha = sign * top[0]->cpu_diff()[0] / bottom[i]->num();

149.  

150. // bbox regression

151. if (channels == 4)

152. {

153. for (int j = 0; j < countLabel; ++j)

154. {

155. int dec = (b1[j*channels + 0] != ignore_label) && (b1[j*channels + 1] != ignore_label) && (b1[j*channels + 2] != ignore_label) && (b1[j*channels + 3] != ignore_label);

156. if (dec==1)

157. {

158. caffe_cpu_axpby(

159. channels, // count

160. alpha, // alpha

161. diff_.cpu_data() + channels * j, // a

162. Dtype(0), // beta

163. bottom[i]->mutable_cpu_diff() + channels * j); // b

164. }

165. }

166. }

167.  

168. // landmark

169. else if (channels == 10)

170. {

171. for (int j = 0; j < countLabel; ++j)

172. {

173. int dec1 = (b1[j*channels + 0] != ignore_label) && (b1[j*channels + 1] != ignore_label) && (b1[j*channels + 2] != ignore_label) && (b1[j*channels + 3] != ignore_label) && (b1[j*channels + 4] != ignore_label);

174. int dec2 = (b1[j*channels + 5] != ignore_label) && (b1[j*channels + 6] != ignore_label) && (b1[j*channels + 7] != ignore_label) && (b1[j*channels + 8] != ignore_label) && (b1[j*channels + 9] != ignore_label);

175. if (dec1 == 1 && dec2 == 1)

176. {

177. caffe_cpu_axpby(

178. channels, // count

179. alpha, // alpha

180. diff_.cpu_data() + channels * j, // a

181. Dtype(0), // beta

182. bottom[i]->mutable_cpu_diff() + channels * j); // b

183. }

184. }

185. }

186.  

187. // ***********************org data********************

188. //for (int j = 0; j < countLabel; ++j){

189. // if (label[j] != ignore_label){

190. // caffe_cpu_axpby(

191. // channels, // count

192. // alpha, // alpha

193. // diff_.cpu_data() + channels * j, // a

194. // Dtype(0), // beta

195. // bottom[i]->mutable_cpu_diff() + channels * j); // b

196. // }

197. //}

198.  
199.  

200. }

201. }

202. }

203. else{

204. for (int i = 0; i < 2; ++i) {

205. if (propagate_down[i]) {

206. const Dtype sign = (i == 0) ? 1 : -1;

207. const Dtype alpha = sign * top[0]->cpu_diff()[0] / bottom[i]->num();

208. caffe_cpu_axpby(

209. bottom[i]->count(), // count

210. alpha, // alpha

211. diff_.cpu_data(), // a

212. Dtype(0), // beta

213. bottom[i]->mutable_cpu_diff()); // b

214. }

215. }

216. }

217. }

218.  

219. #ifdef CPU_ONLY

220. STUB_GPU(MTCNNEuclideanLossLayer);

221. #endif

222.  

223. INSTANTIATE_CLASS(MTCNNEuclideanLossLayer);

224. REGISTER_LAYER_CLASS(MTCNNEuclideanLoss);

225.  

226. } // namespace caffe

227.  

相应的"MTCNNEuclideanLoss.cu"如下：

1. #include

2.  

3. #include "caffe/layers/mtcnn_euclidean_loss_layer.hpp"

4. #include "caffe/util/math_functions.hpp"

5.  

6. #include

7. using namespace std;

8.  

9. namespace caffe {

10.  

11. template

12. void MTCNNEuclideanLossLayer::Reshape(

13. const vector*>& bottom, const vector*>& top) {

14. LossLayer::Reshape(bottom, top);

15. CHECK_EQ(bottom[0]->count(1), bottom[1]->count(1))

16. << "Inputs must have the same dimension.";

17.  

18. int has_ignore_label = this->layer_param().loss_param().has_ignore_label();

19. if (has_ignore_label)

20. CHECK_EQ(bottom.size(), 3) << "has_ignore_label=true but not input label";

21.  

22. if (!has_ignore_label)

23. CHECK_EQ(bottom.size(), 2) << "has_ignore_label=false but input mismatch";

24.  

25. diff_.ReshapeLike(*bottom[0]);

26. }

27.  

28. template

29. void MTCNNEuclideanLossLayer::Forward_gpu(const vector*>& bottom,

30. const vector*>& top) {

31. int count = bottom[0]->count();

32. int has_ignore_label = this->layer_param().loss_param().has_ignore_label();

33. int ignore_label = has_ignore_label ? this->layer_param().loss_param().ignore_label() : -1;

34.  

35. if (has_ignore_label){

36. //label

37. const Dtype* label = bottom[2]->cpu_data();

38. Dtype* diff = diff_.mutable_gpu_data();

39. int countLabel = bottom[2]->num();

40. int channel = bottom[0]->channels();

41. //cout << "ignore_label_forward: " << ignore_label << endl; //

42. caffe_gpu_memset(sizeof(Dtype)*count, 0, diff);

43.  

44. const Dtype* b0 = bottom[0]->gpu_data();

45. const Dtype* b1 = bottom[1]->gpu_data();

46. const Dtype* b1_cpu = bottom[1]->cpu_data();

47. Dtype loss = 0;

48.  

49. //cout << "channel_forward " << channel << endl;

50. // bbox regression

51. if (channel == 4)

52. {

53. for (int i = 0; i < countLabel; ++i)

54. {

55. //cout << "forware_b1_4: " << b1_cpu[i*channel + 0] << " " << b1_cpu[i*channel + 1] << " " << b1_cpu[i*channel + 2] << " " << b1_cpu[i*channel + 3] << endl;

56.  

57. int dec = (b1_cpu[i*channel + 0] != ignore_label) && (b1_cpu[i*channel + 1] != ignore_label) && (b1_cpu[i*channel + 2] != ignore_label) && (b1_cpu[i*channel + 3] != ignore_label);

58. if (dec == 1)

59. {

60. caffe_gpu_sub(

61. channel,

62. b0 + i * channel,

63. b1 + i * channel,

64. diff + i * channel);

65. Dtype dot;

66. caffe_gpu_dot(channel, diff + i * channel, diff + i * channel, &dot);

67. loss += dot / Dtype(2);

68. }

69. }

70. }

71.  

72. // landmark

73. else if (channel == 10)

74. {

75. for (int i = 0; i < countLabel; ++i)

76. {

77. int dec1 = (b1_cpu[i*channel + 0] != ignore_label) && (b1_cpu[i*channel + 1] != ignore_label) && (b1_cpu[i*channel + 2] != ignore_label) && (b1_cpu[i*channel + 3] != ignore_label) && (b1_cpu[i*channel + 4] != ignore_label);

78. int dec2 = (b1_cpu[i*channel + 5] != ignore_label) && (b1_cpu[i*channel + 6] != ignore_label) && (b1_cpu[i*channel + 7] != ignore_label) && (b1_cpu[i*channel + 8] != ignore_label) && (b1_cpu[i*channel + 9] != ignore_label);

79. if (dec1 == 1 && dec2 == 1)

80. {

81. caffe_gpu_sub(

82. channel,

83. b0 + i * channel,

84. b1 + i * channel,

85. diff + i * channel);

86. Dtype dot;

87. caffe_gpu_dot(channel, diff + i * channel, diff + i * channel, &dot);

88. loss += dot / Dtype(2);

89. }

90. }

91. }

92.  
93.  

94. // ***********************org data ********************

95. //for (int i = 0; i < countLabel; ++i){

96. // if (label[i] != ignore_label){

97. // caffe_gpu_sub(

98. // channel,

99. // b0 + i * channel,

100. // b1 + i * channel,

101. // diff + i * channel);

102. // Dtype dot;

103. // caffe_gpu_dot(channel, diff + i * channel, diff + i * channel, &dot);

104. // loss += dot / Dtype(2);

105. // }

106. //}

107. // **************************** **********************

108.  

109. top[0]->mutable_cpu_data()[0] = loss;

110. }

111. else{

112. int count = bottom[0]->count();

113. caffe_gpu_sub(

114. count,

115. bottom[0]->gpu_data(),

116. bottom[1]->gpu_data(),

117. diff_.mutable_gpu_data());

118. Dtype dot;

119. caffe_gpu_dot(count, diff_.gpu_data(), diff_.gpu_data(), &dot);

120. Dtype loss = dot / bottom[0]->num() / Dtype(2);

121. top[0]->mutable_cpu_data()[0] = loss;

122. }

123. }

124.  

125. template

126. void MTCNNEuclideanLossLayer::Backward_gpu(const vector*>& top,

127. const vector& propagate_down, const vector*>& bottom) {

128.  

129. int has_ignore_label = this->layer_param().loss_param().has_ignore_label();

130. int ignore_label = has_ignore_label ? this->layer_param().loss_param().ignore_label() : -1;

131.  

132. if (has_ignore_label){

133. const Dtype* b1 = bottom[1]->cpu_data();

134. const Dtype* label = bottom[2]->cpu_data();

135. int countLabel = bottom[2]->num();

136. int channels = bottom[0]->channels();

137. for (int i = 0; i < 2; ++i) {

138. if (propagate_down[i]) {

139. caffe_gpu_memset(sizeof(Dtype)*bottom[i]->count(), 0, bottom[i]->mutable_gpu_diff());

140.  

141. const Dtype sign = (i == 0) ? 1 : -1;

142.  

143. // bbox regression

144. if (channels == 4)

145. {

146. for (int j = 0; j < countLabel; ++j)

147. {

148. const Dtype alpha = sign * top[0]->cpu_diff()[0];

149. int dec = (b1[j*channels + 0] != ignore_label) && (b1[j*channels + 1] != ignore_label) && (b1[j*channels + 2] != ignore_label) && (b1[j*channels + 3] != ignore_label);

150. if (dec == 1)

151. {

152. caffe_gpu_axpby(

153. channels, // count

154. alpha, // alpha

155. diff_.gpu_data() + channels * j, // a

156. Dtype(0), // beta

157. bottom[i]->mutable_gpu_diff() + channels * j); // b

158. }

159. }

160. }

161.  

162. // landmark

163. else if (channels == 10)

164. {

165. for (int j = 0; j < countLabel; ++j)

166. {

167. const Dtype alpha = sign * top[0]->cpu_diff()[0];

168. int dec1 = (b1[j*channels + 0] != ignore_label) && (b1[j*channels + 1] != ignore_label) && (b1[j*channels + 2] != ignore_label) && (b1[j*channels + 3] != ignore_label) && (b1[j*channels + 4] != ignore_label);

169. int dec2 = (b1[j*channels + 5] != ignore_label) && (b1[j*channels + 6] != ignore_label) && (b1[j*channels + 7] != ignore_label) && (b1[j*channels + 8] != ignore_label) && (b1[j*channels + 9] != ignore_label);

170. if (dec1 == 1 && dec2 == 1)

171. {

172. caffe_gpu_axpby(

173. channels, // count

174. alpha, // alpha

175. diff_.gpu_data() + channels * j, // a

176. Dtype(0), // beta

177. bottom[i]->mutable_gpu_diff() + channels * j); // b

178. }

179. }

180. }

181.  

182. // ******************* org data**********************

183. //for (int j = 0; j < countLabel; ++j){

184. // const Dtype alpha = sign * top[0]->cpu_diff()[0];

185. // if (label[j] != ignore_label){

186. // caffe_gpu_axpby(

187. // channels, // count

188. // alpha, // alpha

189. // diff_.gpu_data() + channels * j, // a

190. // Dtype(0), // beta

191. // bottom[i]->mutable_gpu_diff() + channels * j); // b

192. // }

193. //}

194.  
195.  
196.  

197. }

198. }

199. }

200. else{

201. for (int i = 0; i < 2; ++i) {

202. if (propagate_down[i]) {

203. const Dtype sign = (i == 0) ? 1 : -1;

204. const Dtype alpha = sign * top[0]->cpu_diff()[0] / bottom[i]->num();

205. caffe_gpu_axpby(

206. bottom[i]->count(), // count

207. alpha, // alpha

208. diff_.gpu_data(), // a

209. Dtype(0), // beta

210. bottom[i]->mutable_gpu_diff()); // b

211. }

212. }

213. }

214. }

215.  

216. INSTANTIATE_LAYER_GPU_FUNCS(MTCNNEuclideanLossLayer);

217.  

218. } // namespace caffe

219.  

小计，完成mtcnn_educlidean_loss_layer.cu的修改，发现：const Dtype* b1 = bottom[1]->gpu_data();，gpu读取的数据是不能打印和取出来的，改成cpu模式：const Dtype* b1_cpu = bottom[1]->cpu_data()就行了；