The output of MultiLabelMarginLoss() is zero

1

Hi, everyone, @SpandanMadan, @AjayTalati
A number plate recognization exercise,
samples have multiple labels, every has 7 letters, 252 labels.
the inputting image size 224224, the target vector width is 252(736),for example:
‘X’ versus 000000000000000000000001000000000000
’A’ versus 100000000000000000000000000000000000

while training ,in 2nd minibatch, the output of multilabelmarginloss() is zero, i cann’t find out the reason.

Can anyone help me?

I’ve pasted the whole source code in Multi Label Classification in pytorch

best regards sincerely

2

In first minibatch, the output of multilabelmarginloss() Loss: 246.707458
the second, it is zero.
In[10]: output_var.data
Out[10]:

2.3264e+03 6.5879e+01 -9.7149e+00 … -2.6423e+01 2.1553e+00 2.3724e+01
2.9392e+03 8.3356e+01 -1.2624e+01 … -3.3264e+01 2.6289e+00 2.9792e+01
2.1745e+03 6.1491e+01 -9.2812e+00 … -2.4580e+01 1.9765e+00 2.2134e+01
2.4772e+03 7.0001e+01 -1.0521e+01 … -2.8103e+01 2.1341e+00 2.5243e+01
2.0916e+03 5.9231e+01 -8.8647e+00 … -2.3580e+01 1.9928e+00 2.1346e+01
[torch.cuda.FloatTensor of size 5x252 (GPU 0)]

In[11]: target_var.data
Out[11]:

0     0     0  ...      0     0     0
0     0     0  ...      0     0     0
0     0     0  ...      0     0     0
0     0     0  ...      0     0     0
0     0     0  ...      0     0     0

[torch.cuda.LongTensor of size 5x252 (GPU 0)]