id_accuracy=0

[jjn037]

I1122 20:34:16.939185 8871 sgd_solver.cpp:105] Iteration 34120, lr = 0.001
I1122 20:34:25.444725 8871 solver.cpp:219] Iteration 34140 (2.35148 iter/s, 8.50528s/20 iters), loss = 5.27471
I1122 20:34:25.444803 8871 solver.cpp:238] Train net output #0: det_accuracy = 0.953125
I1122 20:34:25.444819 8871 solver.cpp:238] Train net output #1: det_loss = 0.116758 (* 1 = 0.116758 loss)
I1122 20:34:25.444828 8871 solver.cpp:238] Train net output #2: id_accuracy = 0
I1122 20:34:25.444836 8871 solver.cpp:238] Train net output #3: id_loss = 7.04522 (* 1 = 7.04522 loss)
I1122 20:34:25.444845 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.379949 (* 1 = 0.379949 loss)
I1122 20:34:25.444854 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.0538092 (* 1 = 0.0538092 loss)
I1122 20:34:25.444862 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.0670886 (* 1 = 0.0670886 loss)
I1122 20:34:25.444875 8871 sgd_solver.cpp:105] Iteration 34140, lr = 0.001
I1122 20:34:33.161725 8871 solver.cpp:219] Iteration 34160 (2.59181 iter/s, 7.71661s/20 iters), loss = 5.08991
I1122 20:34:33.161808 8871 solver.cpp:238] Train net output #0: det_accuracy = 0.929688
I1122 20:34:33.161828 8871 solver.cpp:238] Train net output #1: det_loss = 0.10782 (* 1 = 0.10782 loss)
I1122 20:34:33.161839 8871 solver.cpp:238] Train net output #2: id_accuracy = 0
I1122 20:34:33.161851 8871 solver.cpp:238] Train net output #3: id_loss = 6.10264 (* 1 = 6.10264 loss)
I1122 20:34:33.161862 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.17604 (* 1 = 0.17604 loss)
I1122 20:34:33.161873 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.127531 (* 1 = 0.127531 loss)
I1122 20:34:33.161885 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.0870427 (* 1 = 0.0870427 loss)
I1122 20:34:33.161900 8871 sgd_solver.cpp:105] Iteration 34160, lr = 0.001
I1122 20:34:40.963131 8871 solver.cpp:219] Iteration 34180 (2.56377 iter/s, 7.801s/20 iters), loss = 5.09691
I1122 20:34:40.963213 8871 solver.cpp:238] Train net output #0: det_accuracy = 0.929688
I1122 20:34:40.963233 8871 solver.cpp:238] Train net output #1: det_loss = 0.163159 (* 1 = 0.163159 loss)
I1122 20:34:40.963245 8871 solver.cpp:238] Train net output #2: id_accuracy = 0.3
I1122 20:34:40.963258 8871 solver.cpp:238] Train net output #3: id_loss = 3.88689 (* 1 = 3.88689 loss)
I1122 20:34:40.963268 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.222411 (* 1 = 0.222411 loss)
I1122 20:34:40.963279 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.202602 (* 1 = 0.202602 loss)
I1122 20:34:40.963289 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.993391 (* 1 = 0.993391 loss)
I1122 20:34:40.963304 8871 sgd_solver.cpp:105] Iteration 34180, lr = 0.001
I1122 20:34:49.098305 8871 solver.cpp:219] Iteration 34200 (2.45858 iter/s, 8.13476s/20 iters), loss = 5.19017
I1122 20:34:49.098373 8871 solver.cpp:238] Train net output #0: det_accuracy = 0.984375
I1122 20:34:49.098388 8871 solver.cpp:238] Train net output #1: det_loss = 0.0294993 (* 1 = 0.0294993 loss)
I1122 20:34:49.098397 8871 solver.cpp:238] Train net output #2: id_accuracy = 0.2
I1122 20:34:49.098404 8871 solver.cpp:238] Train net output #3: id_loss = 4.19861 (* 1 = 4.19861 loss)
I1122 20:34:49.098413 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.0517187 (* 1 = 0.0517187 loss)
I1122 20:34:49.098422 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.0103175 (* 1 = 0.0103175 loss)
I1122 20:34:49.098429 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.00192679 (* 1 = 0.00192679 loss)
I1122 20:34:49.098440 8871 sgd_solver.cpp:105] Iteration 34200, lr = 0.001
I1122 20:34:57.153692 8871 solver.cpp:219] Iteration 34220 (2.48293 iter/s, 8.05499s/20 iters), loss = 5.04621
I1122 20:34:57.153774 8871 solver.cpp:238] Train net output #0: det_accuracy = 0.976562
I1122 20:34:57.153792 8871 solver.cpp:238] Train net output #1: det_loss = 0.0611082 (* 1 = 0.0611082 loss)
I1122 20:34:57.153805 8871 solver.cpp:238] Train net output #2: id_accuracy = -nan
I1122 20:34:57.153815 8871 solver.cpp:238] Train net output #3: id_loss = 0 (* 1 = 0 loss)
I1122 20:34:57.153827 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.120212 (* 1 = 0.120212 loss)
I1122 20:34:57.153837 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.0470721 (* 1 = 0.0470721 loss)
I1122 20:34:57.153849 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.00547891 (* 1 = 0.00547891 loss)
I1122 20:34:57.153864 8871 sgd_solver.cpp:105] Iteration 34220, lr = 0.001
^[[B^[[B^[[B^[[B^[[B^[[B^[[B^[[B^[[B^[[B^[[B^[[BI1122 20:35:04.813046 8871 solver.cpp:219] Iteration 34240 (2.61132 iter/s, 7.65896s/20 iters), loss = 4.89673
I1122 20:35:04.813129 8871 solver.cpp:238] Train net output #0: det_accuracy = 1
I1122 20:35:04.813149 8871 solver.cpp:238] Train net output #1: det_loss = 0.022339 (* 1 = 0.022339 loss)
I1122 20:35:04.813160 8871 solver.cpp:238] Train net output #2: id_accuracy = -nan
I1122 20:35:04.813174 8871 solver.cpp:238] Train net output #3: id_loss = 0 (* 1 = 0 loss)
I1122 20:35:04.813184 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.126798 (* 1 = 0.126798 loss)
I1122 20:35:04.813195 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.02272 (* 1 = 0.02272 loss)
I1122 20:35:04.813206 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.0226682 (* 1 = 0.0226682 loss)
I1122 20:35:04.813221 8871 sgd_solver.cpp:105] Iteration 34240, lr = 0.001
I1122 20:35:13.222926 8871 solver.cpp:219] Iteration 34260 (2.37827 iter/s, 8.40946s/20 iters), loss = 5.04766
I1122 20:35:13.222995 8871 solver.cpp:238] Train net output #0: det_accuracy = 0.875
I1122 20:35:13.223009 8871 solver.cpp:238] Train net output #1: det_loss = 0.255105 (* 1 = 0.255105 loss)
I1122 20:35:13.223017 8871 solver.cpp:238] Train net output #2: id_accuracy = 0
I1122 20:35:13.223027 8871 solver.cpp:238] Train net output #3: id_loss = 4.59072 (* 1 = 4.59072 loss)
I1122 20:35:13.223036 8871 solver.cpp:238] Train net output #4: loss_bbox = 0.393796 (* 1 = 0.393796 loss)
I1122 20:35:13.223043 8871 solver.cpp:238] Train net output #5: rpn_cls_loss = 0.0156542 (* 1 = 0.0156542 loss)
I1122 20:35:13.223052 8871 solver.cpp:238] Train net output #6: rpn_loss_bbox = 0.00982659 (* 1 = 0.00982659 loss)

I added your labled_matching_layer.* and labled_matching_layer.* into https://github.com/soeaver/py-RFCN-priv/tree/master/caffe-priv, but the others like lib/ tools/ experiment/ use yours, and this code doesn't have mpi and my train.prototxt have not
mem_param {
optimize_train: true
optimize_test: true
}

I have successed to make caffe_priv after add the two layer, but when I try to train my model, the id_accuracy seems to get some problems

[jjn037]

And I wonder that if
mem_param {
optimize_train: true
optimize_test: true
}
is important for oim part, and if I have added another files for oim part

[Cysu]

The mem_param is for memory optimization, which reduces the memory consumption by half. As long as you can train the net with your caffe, you can ignore them.

I am not sure what the data you are using, but you may tweak the scale (reciprocal of the temperature) befor the softmax layer. It might help ease the training.