Caffe classification testing losses are always 0 and top-1 is always -nan

26 views Asked by At

I am fine-tuning a pre-trained Caffe model. For some reason, the test iteration always returns loss = 0 and top-1 as -nan. The training iteration values for loss and top-1 look reasonable to me.

I expected that top-1 is figure of accuracy for particular classification and would always range between [0-1] while loss could be float.

Here is the output over 300 iterations, I display results every 100 iterations:

I1002 08:08:24.124348   415 solver.cpp:337] Iteration 0, Testing net (#0)
I1002 08:08:24.221879   415 solver.cpp:404]     Test net output #0: loss_aeroplane_1 = 0
I1002 08:08:24.221915   415 solver.cpp:404]     Test net output #1: loss_aeroplane_2 = 0
I1002 08:08:24.221930   415 solver.cpp:404]     Test net output #2: loss_aeroplane_3 = 0
I1002 08:08:24.221944   415 solver.cpp:404]     Test net output #3: loss_aeroplane_4 = 0
I1002 08:08:24.221958   415 solver.cpp:404]     Test net output #4: loss_aeroplane_5 = 0
I1002 08:08:24.222059   415 solver.cpp:404]     Test net output #11: top-1_aeroplane_1 = -nan
I1002 08:08:24.222121   415 solver.cpp:404]     Test net output #12: top-1_aeroplane_2 = -nan
I1002 08:08:24.222144   415 solver.cpp:404]     Test net output #13: top-1_aeroplane_3 = -nan
I1002 08:08:24.222167   415 solver.cpp:404]     Test net output #14: top-1_aeroplane_4 = -nan
I1002 08:08:24.222190   415 solver.cpp:404]     Test net output #15: top-1_aeroplane_5 = -nan
I1002 08:08:27.674801   415 solver.cpp:228] Iteration 0, loss = 0
I1002 08:08:27.674849   415 solver.cpp:244]     Train net output #0: loss_aeroplane_1 = 0.015304
I1002 08:08:27.674875   415 solver.cpp:244]     Train net output #1: loss_aeroplane_2 = 5.35489
I1002 08:08:27.674890   415 solver.cpp:244]     Train net output #2: loss_aeroplane_3 = 0
I1002 08:08:27.674903   415 solver.cpp:244]     Train net output #3: loss_aeroplane_4 = 22.1282
I1002 08:08:27.674918   415 solver.cpp:244]     Train net output #4: loss_aeroplane_5 = 4.31547
I1002 08:08:27.675045   415 solver.cpp:244]     Train net output #11: top-1_aeroplane_1 = 1
I1002 08:08:27.675061   415 solver.cpp:244]     Train net output #12: top-1_aeroplane_2 = 0
I1002 08:08:27.675077   415 solver.cpp:244]     Train net output #13: top-1_aeroplane_3 = -nan
I1002 08:08:27.675091   415 solver.cpp:244]     Train net output #14: top-1_aeroplane_4 = 0
I1002 08:08:27.675112   415 solver.cpp:244]     Train net output #15: top-1_aeroplane_5 = 0.5
I1002 08:08:27.675243   415 sgd_solver.cpp:106] Iteration 0, lr = 1e-06
I1002 08:14:02.664109   415 solver.cpp:337] Iteration 100, Testing net (#0)
I1002 08:14:02.720834   415 solver.cpp:404]     Test net output #0: loss_aeroplane_1 = 0
I1002 08:14:02.720870   415 solver.cpp:404]     Test net output #1: loss_aeroplane_2 = 0
I1002 08:14:02.720908   415 solver.cpp:404]     Test net output #2: loss_aeroplane_3 = 0
I1002 08:14:02.720922   415 solver.cpp:404]     Test net output #3: loss_aeroplane_4 = 0
I1002 08:14:02.720942   415 solver.cpp:404]     Test net output #4: loss_aeroplane_5 = 0
I1002 08:14:02.721050   415 solver.cpp:404]     Test net output #11: top-1_aeroplane_1 = -nan
I1002 08:14:02.721076   415 solver.cpp:404]     Test net output #12: top-1_aeroplane_2 = -nan
I1002 08:14:02.721098   415 solver.cpp:404]     Test net output #13: top-1_aeroplane_3 = -nan
I1002 08:14:02.721122   415 solver.cpp:404]     Test net output #14: top-1_aeroplane_4 = -nan
I1002 08:14:02.721133   415 solver.cpp:404]     Test net output #15: top-1_aeroplane_5 = -nan
I1002 08:14:06.076464   415 solver.cpp:228] Iteration 100, loss = 0.977386
I1002 08:14:06.076515   415 solver.cpp:244]     Train net output #0: loss_aeroplane_1 = 0
I1002 08:14:06.076532   415 solver.cpp:244]     Train net output #1: loss_aeroplane_2 = 5.73755
I1002 08:14:06.076545   415 solver.cpp:244]     Train net output #2: loss_aeroplane_3 = 0
I1002 08:14:06.076555   415 solver.cpp:244]     Train net output #3: loss_aeroplane_4 = 5.3247e-06
I1002 08:14:06.076563   415 solver.cpp:244]     Train net output #4: loss_aeroplane_5 = 0
I1002 08:14:06.076637   415 solver.cpp:244]     Train net output #11: top-1_aeroplane_1 = -nan
I1002 08:14:06.076644   415 solver.cpp:244]     Train net output #12: top-1_aeroplane_2 = 0
I1002 08:14:06.076653   415 solver.cpp:244]     Train net output #13: top-1_aeroplane_3 = -nan
I1002 08:14:06.076669   415 solver.cpp:244]     Train net output #14: top-1_aeroplane_4 = 1
I1002 08:14:06.076690   415 solver.cpp:244]     Train net output #15: top-1_aeroplane_5 = -nan
I1002 08:14:06.076812   415 sgd_solver.cpp:106] Iteration 100, lr = 1e-06
I1002 08:19:40.601531   415 solver.cpp:337] Iteration 200, Testing net (#0)
I1002 08:19:40.658157   415 solver.cpp:404]     Test net output #0: loss_aeroplane_1 = 0
I1002 08:19:40.658193   415 solver.cpp:404]     Test net output #1: loss_aeroplane_2 = 0
I1002 08:19:40.658202   415 solver.cpp:404]     Test net output #2: loss_aeroplane_3 = 0
I1002 08:19:40.658221   415 solver.cpp:404]     Test net output #3: loss_aeroplane_4 = 0
I1002 08:19:40.658231   415 solver.cpp:404]     Test net output #4: loss_aeroplane_5 = 0
I1002 08:19:40.658363   415 solver.cpp:404]     Test net output #11: top-1_aeroplane_1 = -nan
I1002 08:19:40.658376   415 solver.cpp:404]     Test net output #12: top-1_aeroplane_2 = -nan
I1002 08:19:40.658387   415 solver.cpp:404]     Test net output #13: top-1_aeroplane_3 = -nan
I1002 08:19:40.658407   415 solver.cpp:404]     Test net output #14: top-1_aeroplane_4 = -nan
I1002 08:19:40.658418   415 solver.cpp:404]     Test net output #15: top-1_aeroplane_5 = 1
I1002 08:19:44.013634   415 solver.cpp:228] Iteration 200, loss = 1.24157
I1002 08:19:44.013686   415 solver.cpp:244]     Train net output #0: loss_aeroplane_1 = 1.98091
I1002 08:19:44.013706   415 solver.cpp:244]     Train net output #1: loss_aeroplane_2 = 6.87694
I1002 08:19:44.013716   415 solver.cpp:244]     Train net output #2: loss_aeroplane_3 = 1.29873
I1002 08:19:44.013727   415 solver.cpp:244]     Train net output #3: loss_aeroplane_4 = 17.1241
I1002 08:19:44.013738   415 solver.cpp:244]     Train net output #4: loss_aeroplane_5 = 0
I1002 08:19:44.013875   415 solver.cpp:244]     Train net output #11: top-1_aeroplane_1 = 0.75
I1002 08:19:44.013890   415 solver.cpp:244]     Train net output #12: top-1_aeroplane_2 = 0.5
I1002 08:19:44.013903   415 solver.cpp:244]     Train net output #13: top-1_aeroplane_3 = 0
I1002 08:19:44.013917   415 solver.cpp:244]     Train net output #14: top-1_aeroplane_4 = 0
I1002 08:19:44.013931   415 solver.cpp:244]     Train net output #15: top-1_aeroplane_5 = 1
I1002 08:19:44.014006   415 sgd_solver.cpp:106] Iteration 200, lr = 1e-06
I1002 08:25:19.499899   415 solver.cpp:337] Iteration 300, Testing net (#0)
I1002 08:25:19.556656   415 solver.cpp:404]     Test net output #0: loss_aeroplane_1 = 0
I1002 08:25:19.556694   415 solver.cpp:404]     Test net output #1: loss_aeroplane_2 = 0
I1002 08:25:19.556702   415 solver.cpp:404]     Test net output #2: loss_aeroplane_3 = 0
I1002 08:25:19.556730   415 solver.cpp:404]     Test net output #3: loss_aeroplane_4 = 0
I1002 08:25:19.556743   415 solver.cpp:404]     Test net output #4: loss_aeroplane_5 = 0
I1002 08:25:19.556854   415 solver.cpp:404]     Test net output #11: top-1_aeroplane_1 = -nan
I1002 08:25:19.556874   415 solver.cpp:404]     Test net output #12: top-1_aeroplane_2 = -nan
I1002 08:25:19.556893   415 solver.cpp:404]     Test net output #13: top-1_aeroplane_3 = -nan
I1002 08:25:19.556913   415 solver.cpp:404]     Test net output #14: top-1_aeroplane_4 = -nan
I1002 08:25:19.556931   415 solver.cpp:404]     Test net output #15: top-1_aeroplane_5 = -nan
I1002 08:25:22.910521   415 solver.cpp:228] Iteration 300, loss = 1.06406
I1002 08:25:22.910571   415 solver.cpp:244]     Train net output #0: loss_aeroplane_1 = 0.776529
I1002 08:25:22.910589   415 solver.cpp:244]     Train net output #1: loss_aeroplane_2 = 8.75287
I1002 08:25:22.910600   415 solver.cpp:244]     Train net output #2: loss_aeroplane_3 = 4.64917e-06
I1002 08:25:22.910616   415 solver.cpp:244]     Train net output #3: loss_aeroplane_4 = 12.4562
I1002 08:25:22.910634   415 solver.cpp:244]     Train net output #4: loss_aeroplane_5 = 7.18792
I1002 08:25:22.910776   415 solver.cpp:244]     Train net output #11: top-1_aeroplane_1 = 0
I1002 08:25:22.910794   415 solver.cpp:244]     Train net output #12: top-1_aeroplane_2 = 0
I1002 08:25:22.910815   415 solver.cpp:244]     Train net output #13: top-1_aeroplane_3 = 1
I1002 08:25:22.910849   415 solver.cpp:244]     Train net output #14: top-1_aeroplane_4 = 0
I1002 08:25:22.910869   415 solver.cpp:244]     Train net output #15: top-1_aeroplane_5 = 0
I1002 08:25:22.910969   415 sgd_solver.cpp:106] Iteration 300, lr = 1e-06

Also, can someone please let me know how is Caffe calculating loss here? Is this 'total loss' (sum of losses for all classes) or what?

I1002 08:25:22.910521   415 solver.cpp:228] Iteration 300, loss = 1.06406
0

There are 0 answers