a doubt about "Line 97: priors = priors[pos_idx].view(-1, 4)" in multibox_loss.py

[AllenMao]

hi, i have a doubt, such as the title. the shape of priors is (num_priors, 4), and pos_idx is (num, num_priors, 4). this code how to run success? thank u!

[huangzsdy]

@AllenMao I am confused with that problem,too. It works well because both of 'priors' and 'pos_idx' are Variable. If they are tensor, it will fail to run success. Besides, I also find that this code only works under <0.4.0 version. I tried with pytorch-0.3.1, it works well. @bailvwangzi what's your opinion?

[lthngan]

@huangzsdy: I got the same problem not because they are tensor. It is because of dimensions are not appropriate. For example when running on PASCAL priors: [8732, 4] pos_idx: [32, 8732, 4] How can priors[pos_idx] work no matter Variable or Tensor? If you said it works with Variable, can I like this? va_pos_idx = Variable(pos_idx) va_priors = Variable(priors) va_priors = va_priors[va_pos_idx].view(-1, 4) priors = torch.Tensor(va_priors)

Thank you

[AllenMao]

@huangzsdy i used pytoch-0.3.1.

[huangzsdy]

@lthngan My environment is Python 3.6.2 :: Anaconda custom (64-bit), pytorch0.3.1.post2. It works well. When the program run to this code, the type of priors or pos_idx is Variable, not tensor. If you repeat these operations in python shell, it will be clear. like this: #>>> torch.version '0.3.1.post2' #>>> priors = torch.randn(8732, 4) #>>> pos_idx = priors > 0 #>>> pos_idx = pos_idx.unsqueeze(0).expand_as(torch.randn(32,8732,4)) #>>> priors[pos_idx] Traceback (most recent call last): File "", line 1, in priors[pos_idx] RuntimeError: inconsistent tensor size, expected src [8732 x 4] and mask [32 x 8732 x 4] to have the same number of elements, but go t 34928 and 1117696 elements respectively at /opt/conda/conda-bld/pytorch_1518243271935/work/torch/lib/TH/generic/THTensorMath.c:197

but if they are Variable, it works well #>>> from torch.autograd import Variable #>>> priors = Variable(priors) #>>> pos_idx = priors > 0 #>>> pos_idx = pos_idx.unsqueeze(0).expand_as(torch.randn(32,8732,4)) #>>> priors[pos_idx] Variable containing: 8.5339e-01 5.7070e-01 4.0415e-01 ⋮
8.4663e-01 2.7078e-01 2.4516e+00 [torch.FloatTensor of size 561664] #>>> type(priors) <class 'torch.autograd.variable.Variable'> #>>> type(pos_idx) <class 'torch.autograd.variable.Variable'>

However, It failed when i used pytorch0.4.0(python3.6.2). It came out this error,'IndexError: too many indices for tensor of dimension 2'. If you want to run through this code with pytorch0.4.0 , you can add 'priors = priors.expand_as(pos_idx)' before this code.

The above is just what I found through the experiment. I am confused with this problem,too. If you figure it out, pls let me know...thx.

[huangzsdy]

@AllenMao What's your python version ? What 's your error message ?

[AllenMao]

@huangzsdy the same as you. My environment is python3.5.5 and pytorch0.3.1 priors = priors.expand_as(pos_idx) print(priors[pos_idx].view(-1, 4)) and priors = Variable(priors) print(priors[pos_idx].view(-1, 4)), their results is same on pytoch-0.3.1.

if use 'priors = priors.expand_as(pos_idx)' or pytorch-0.4.0, it came out error in later. so i use pytoch-0.3.1 and run well.

[geyuxi]

I think I've found a solution, and it works. My python version is 3.6.9 and pytorch version is 1.1.0. Firstly, as @huangzsdy said, add priors = priors.expand_as(pos_idx) before priors = priors[pos_idx].view(-1, 4) in multibox_loss.py Secondly, change loss_c[pos] = 0 to loss_c[pos.view(-1, 1)] = 0. Thirdly, change loss_l /= N loss_l_repul /= N loss_c /= N to loss_l /= N.float() loss_l_repul /= N.float() loss_c /= N.float(). And finally, change all loss.data[0] or loss_*.data[0] to loss.data or loss_*.data in train.py. Besides, if your loss value is Nan, you should reduce the value of the learning-rate.