guoyilin / caffe Goto Github PK

在反馈函数中，似乎不是均值呢？而是不断在替换bottom里面的值，最终保留的是最后一次的结果吧？
是不是可以考虑，先caffe_set(bottom[0]->count(), 0, bottom[0]->cpu_diff());
后面再调用caffe_cpu_axpby的时候，改写为：
caffe_cpu_axpby(
channels,
Dtype(1.0 / image_count_dtype[static_cast(top0_map_Dtype[i])]),
top[0]->cpu_diff() + (i * channels),
Dtype(1.0),
bout + (static_cast(top0_map_Dtype[i])*channels)
);
因为我看到您，已经写到要计算均值的，可是为什么又注释掉呢？求解~

I0811 04:43:50.113242 6500 solver.cpp:214] Iteration 0, loss = 0
I0811 04:43:50.113284 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:43:50.113313 6500 solver.cpp:486] Iteration 0, lr = 0.0001
I0811 04:44:57.945654 6500 solver.cpp:214] Iteration 18, loss = 48.8384
I0811 04:44:57.945754 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:44:57.945766 6500 solver.cpp:486] Iteration 18, lr = 0.0001
I0811 04:46:05.653295 6500 solver.cpp:214] Iteration 36, loss = 16.5096
I0811 04:46:05.653362 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:46:05.653375 6500 solver.cpp:486] Iteration 36, lr = 0.0001
I0811 04:47:13.365404 6500 solver.cpp:214] Iteration 54, loss = -3.05176e-05
I0811 04:47:13.365469 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:47:13.365480 6500 solver.cpp:486] Iteration 54, lr = 0.0001
I0811 04:48:21.204015 6500 solver.cpp:214] Iteration 72, loss = 0.703822
I0811 04:48:21.204080 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:48:21.204092 6500 solver.cpp:486] Iteration 72, lr = 0.0001
I0811 04:49:29.063345 6500 solver.cpp:214] Iteration 90, loss = 0.234567
I0811 04:49:29.063438 6500 solver.cpp:229] Train net output #0: tripletloss = 2.11134 (* 1 = 2.11134 loss)
I0811 04:49:29.063448 6500 solver.cpp:486] Iteration 90, lr = 0.0001
I0811 04:50:36.787369 6500 solver.cpp:214] Iteration 108, loss = 1.7714
I0811 04:50:36.787441 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:50:36.787454 6500 solver.cpp:486] Iteration 108, lr = 0.0001
I0811 04:51:44.535688 6500 solver.cpp:214] Iteration 126, loss = -6.11544e-05
I0811 04:51:44.535756 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:51:44.535769 6500 solver.cpp:486] Iteration 126, lr = 0.0001
I0811 04:52:52.248569 6500 solver.cpp:214] Iteration 144, loss = 12.0253
I0811 04:52:52.248634 6500 solver.cpp:229] Train net output #0: tripletloss = 14.9139 (* 1 = 14.9139 loss)
I0811 04:52:52.248643 6500 solver.cpp:486] Iteration 144, lr = 0.0001
I0811 04:53:59.917649 6500 solver.cpp:214] Iteration 162, loss = -6.27041e-05
I0811 04:53:59.917714 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:53:59.917728 6500 solver.cpp:486] Iteration 162, lr = 0.0001
I0811 04:55:07.621796 6500 solver.cpp:214] Iteration 180, loss = 0.0672214
I0811 04:55:07.621862 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:55:07.621872 6500 solver.cpp:486] Iteration 180, lr = 0.0001
I0811 04:56:15.294675 6500 solver.cpp:214] Iteration 198, loss = 2.1794
I0811 04:56:15.294742 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:56:15.294755 6500 solver.cpp:486] Iteration 198, lr = 0.0001
I0811 04:57:22.965425 6500 solver.cpp:214] Iteration 216, loss = 0.459415
I0811 04:57:22.965522 6500 solver.cpp:229] Train net output #0: tripletloss = 4.13529 (* 1 = 4.13529 loss)
I0811 04:57:22.965535 6500 solver.cpp:486] Iteration 216, lr = 0.0001
I0811 04:58:31.153684 6500 solver.cpp:214] Iteration 234, loss = 0.0995185
I0811 04:58:31.153771 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:58:31.153781 6500 solver.cpp:486] Iteration 234, lr = 0.0001
I0811 04:59:39.472342 6500 solver.cpp:214] Iteration 252, loss = 1.246
I0811 04:59:39.472421 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 04:59:39.472436 6500 solver.cpp:486] Iteration 252, lr = 0.0001
I0811 05:00:47.818704 6500 solver.cpp:214] Iteration 270, loss = 2.32974
I0811 05:00:47.818778 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:00:47.818792 6500 solver.cpp:486] Iteration 270, lr = 0.0001
I0811 05:01:56.114686 6500 solver.cpp:214] Iteration 288, loss = 0.0435698
I0811 05:01:56.114778 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:01:56.114791 6500 solver.cpp:486] Iteration 288, lr = 0.0001
I0811 05:03:04.489054 6500 solver.cpp:214] Iteration 306, loss = 2.23856
I0811 05:03:04.489162 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:03:04.489187 6500 solver.cpp:486] Iteration 306, lr = 0.0001
I0811 05:04:12.155246 6500 solver.cpp:214] Iteration 324, loss = 101.816
I0811 05:04:12.155308 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:04:12.155318 6500 solver.cpp:486] Iteration 324, lr = 0.0001
I0811 05:05:19.840358 6500 solver.cpp:214] Iteration 342, loss = 0.825645
I0811 05:05:19.840417 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:05:19.840427 6500 solver.cpp:486] Iteration 342, lr = 0.0001
I0811 05:06:27.506650 6500 solver.cpp:214] Iteration 360, loss = 0.746133
I0811 05:06:27.506710 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:06:27.506718 6500 solver.cpp:486] Iteration 360, lr = 0.0001
I0811 05:07:35.187403 6500 solver.cpp:214] Iteration 378, loss = -5.87702e-05
I0811 05:07:35.187466 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:07:35.187479 6500 solver.cpp:486] Iteration 378, lr = 0.0001
I0811 05:08:43.087249 6500 solver.cpp:214] Iteration 396, loss = 0.414626
I0811 05:08:43.087309 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:08:43.087319 6500 solver.cpp:486] Iteration 396, lr = 0.0001
I0811 05:09:50.796950 6500 solver.cpp:214] Iteration 414, loss = 2.79512
I0811 05:09:50.797013 6500 solver.cpp:229] Train net output #0: tripletloss = 10.4305 (* 1 = 10.4305 loss)
I0811 05:09:50.797021 6500 solver.cpp:486] Iteration 414, lr = 0.0001
I0811 05:10:58.498670 6500 solver.cpp:214] Iteration 432, loss = 3.24082
I0811 05:10:58.498751 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:10:58.498765 6500 solver.cpp:486] Iteration 432, lr = 0.0001
I0811 05:12:06.169255 6500 solver.cpp:214] Iteration 450, loss = 1.53988
I0811 05:12:06.169314 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:12:06.169327 6500 solver.cpp:486] Iteration 450, lr = 0.0001
I0811 05:13:13.850697 6500 solver.cpp:214] Iteration 468, loss = 1.0264
I0811 05:13:13.850757 6500 solver.cpp:229] Train net output #0: tripletloss = 9.23815 (* 1 = 9.23815 loss)
I0811 05:13:13.850766 6500 solver.cpp:486] Iteration 468, lr = 0.0001
I0811 05:14:21.507980 6500 solver.cpp:214] Iteration 486, loss = 1.96649
I0811 05:14:21.508039 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:14:21.508051 6500 solver.cpp:486] Iteration 486, lr = 0.0001
I0811 05:15:29.194263 6500 solver.cpp:214] Iteration 504, loss = -5.94854e-05
I0811 05:15:29.194319 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:15:29.194329 6500 solver.cpp:486] Iteration 504, lr = 0.0001
I0811 05:16:36.892102 6500 solver.cpp:214] Iteration 522, loss = 1.02462
I0811 05:16:36.892195 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:16:36.892205 6500 solver.cpp:486] Iteration 522, lr = 0.0001
I0811 05:17:44.602576 6500 solver.cpp:214] Iteration 540, loss = -5.96487e-05
I0811 05:17:44.602639 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:17:44.602649 6500 solver.cpp:486] Iteration 540, lr = 0.0001
I0811 05:18:52.329092 6500 solver.cpp:214] Iteration 558, loss = 0.168909
I0811 05:18:52.329149 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:18:52.329159 6500 solver.cpp:486] Iteration 558, lr = 0.0001
I0811 05:20:00.048117 6500 solver.cpp:214] Iteration 576, loss = -6.00815e-05
I0811 05:20:00.048174 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:20:00.048185 6500 solver.cpp:486] Iteration 576, lr = 0.0001
I0811 05:21:07.706097 6500 solver.cpp:214] Iteration 594, loss = -6.01411e-05
I0811 05:21:07.706153 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:21:07.706162 6500 solver.cpp:486] Iteration 594, lr = 0.0001
I0811 05:22:15.423184 6500 solver.cpp:214] Iteration 612, loss = 0.130687
I0811 05:22:15.423243 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:22:15.423254 6500 solver.cpp:486] Iteration 612, lr = 0.0001
I0811 05:23:23.084491 6500 solver.cpp:214] Iteration 630, loss = -6.03199e-05
I0811 05:23:23.084550 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:23:23.084563 6500 solver.cpp:486] Iteration 630, lr = 0.0001
I0811 05:24:30.802199 6500 solver.cpp:214] Iteration 648, loss = -6.03199e-05
I0811 05:24:30.802258 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:24:30.802266 6500 solver.cpp:486] Iteration 648, lr = 0.0001
I0811 05:25:38.485126 6500 solver.cpp:214] Iteration 666, loss = 0.0311326
I0811 05:25:38.485186 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:25:38.485196 6500 solver.cpp:486] Iteration 666, lr = 0.0001
I0811 05:26:46.154042 6500 solver.cpp:214] Iteration 684, loss = 0.476817
I0811 05:26:46.154126 6500 solver.cpp:229] Train net output #0: tripletloss = 0.774595 (* 1 = 0.774595 loss)
I0811 05:26:46.154137 6500 solver.cpp:486] Iteration 684, lr = 0.0001
I0811 05:27:53.814285 6500 solver.cpp:214] Iteration 702, loss = 0.216171
I0811 05:27:53.814348 6500 solver.cpp:229] Train net output #0: tripletloss = 1.01635 (* 1 = 1.01635 loss)
I0811 05:27:53.814358 6500 solver.cpp:486] Iteration 702, lr = 0.0001
I0811 05:29:01.489145 6500 solver.cpp:214] Iteration 720, loss = 1.53251
I0811 05:29:01.489208 6500 solver.cpp:229] Train net output #0: tripletloss = 0 (* 1 = 0 loss)
I0811 05:29:01.489217 6500 solver.cpp:486] Iteration 720, lr = 0.0001
I0811 05:30:09.188515 6500 solver.cpp:214] Iteration 738, loss = 1.87498
I0811 05:30:09.188573 6500 solver.cpp:229] Train net output #0: tripletloss = 0.0788164 (* 1 = 0.0788164 loss)

Hello! I see your cifar-10 using the triplet loss replacing the softmax. If we want to test the triplet the accuracy, what can I do?

Existing in caffe general way should I do??
Or, Do you have a special way for metric learning(triplet loss)?

thank you