Zero-dimensional tensor concatenation problem

[leobxpan]

Hi there,

Thank you for the code!

While training the ResNet50 model using the market1501 dataset, I got the following Runtime error:

Traceback (most recent call last):
  File "examples/triplet_loss.py", line 232, in <module>
    main(parser.parse_args())
  File "examples/triplet_loss.py", line 151, in main
    trainer.train(epoch, train_loader, optimizer)
  File "/home/bxpan/.local/lib/python3.5/site-packages/open_reid-0.2.0-py3.5.egg/reid/trainers.py", line 33, in train
  File "/home/bxpan/.local/lib/python3.5/site-packages/open_reid-0.2.0-py3.5.egg/reid/trainers.py", line 83, in _forward
  File "/home/bxpan/.local/lib/python3.5/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/bxpan/.local/lib/python3.5/site-packages/open_reid-0.2.0-py3.5.egg/reid/loss/triplet.py", line 26, in forward
RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated

The problem turned out to happen at this specific line of code in triplet.py: dist_ap = torch.cat(dist_ap) I have printed out dist_ap, which is a python list full of zero-dimensional (its printed-out size is: torch.Size([])) tensors (I used a batch-size of 64 so the list has a length of 64):

[tensor(0.2895, device='cuda:0'), tensor(0.3334, device='cuda:0'), tensor(0.3334, device='cuda:0'), tensor(0.3175, device='cuda:0'), tensor(0.3078, device='cuda:0'), tensor(0.3078, device='cuda:0'), tensor(0.3045, device='cuda:0'), tensor(0.3045, device='cuda:0'), tensor(0.2636, device='cuda:0'), tensor(0.2630, device='cuda:0'), tensor(0.2497, device='cuda:0'), tensor(0.2636, device='cuda:0'), tensor(0.2967, device='cuda:0'), tensor(0.2657, device='cuda:0'), tensor(0.2967, device='cuda:0'), tensor(0.2936, device='cuda:0'), tensor(0.3517, device='cuda:0'), tensor(0.2939, device='cuda:0'), tensor(0.3517, device='cuda:0'), tensor(0.3185, device='cuda:0'), tensor(0.3318, device='cuda:0'), tensor(0.3357, device='cuda:0'), tensor(0.3260, device='cuda:0'), tensor(0.3357, device='cuda:0'), tensor(0.2928, device='cuda:0'), tensor(0.2906, device='cuda:0'), tensor(0.2928, device='cuda:0'), tensor(0.2906, device='cuda:0'), tensor(0.1992, device='cuda:0'), tensor(0.2086, device='cuda:0'), tensor(0.2086, device='cuda:0'), tensor(0.2040, device='cuda:0'), tensor(0.2742, device='cuda:0'), tensor(0.2836, device='cuda:0'), tensor(0.3117, device='cuda:0'), tensor(0.3117, device='cuda:0'), tensor(0.2838, device='cuda:0'), tensor(0.2686, device='cuda:0'), tensor(0.2435, device='cuda:0'), tensor(0.2838, device='cuda:0'), tensor(0.3124, device='cuda:0'), tensor(0.3268, device='cuda:0'), tensor(0.3304, device='cuda:0'), tensor(0.3304, device='cuda:0'), tensor(0.2591, device='cuda:0'), tensor(0.2671, device='cuda:0'), tensor(0.2825, device='cuda:0'), tensor(0.2825, device='cuda:0'), tensor(0.3309, device='cuda:0'), tensor(0.2836, device='cuda:0'), tensor(0.3126, device='cuda:0'), tensor(0.3309, device='cuda:0'), tensor(0.3232, device='cuda:0'), tensor(0.3493, device='cuda:0'), tensor(0.3493, device='cuda:0'), tensor(0.3379, device='cuda:0'), tensor(0.3044, device='cuda:0'), tensor(0.3173, device='cuda:0'), tensor(0.3173, device='cuda:0'), tensor(0.3009, device='cuda:0'), tensor(0.2941, device='cuda:0'), tensor(0.3048, device='cuda:0'), tensor(0.3048, device='cuda:0'), tensor(0.2704, device='cuda:0')]

The values of the tensors seem to be of no problem, but the concatenation fails. Any idea about what the problem is?

Thank you very much.

Boxiao

[frhf]

Hi,

I got the same problem recently. I think it is connected to a newer version of pytorch.

What worked for me is replacing torch.cat with torch.stack, but I am not entirely sure if this solution is unproblematic.

Regards Frank

[leobxpan]

Hi Frank,

Sorry for the late response. I've tried your solution and it works. I'm not sure if torch.cat() can concatenate 1-dimensional tensors in previous versions of PyTorch. If it can, then this might be related to a version change. Better invite the author to check this @Cysu

Boxiao

[lijianhackthon]

Yeah, I come across the same issue and my Pytorch version is 0.4. Hope the author @Cysu can look into this.

[diaoenmao]

Actually, this issue is fixed for me at 0.4.1

[insikk]

@dem123456789 It works fine with pytorch 0.3.0. I saw this error on pytorch 0.4.0, so I upgraded it to 0.4.1. Problem still exists.

[mhyousefi]

I'm having a similar problem, where I cant concatenate the elements in a list of zero-dimensional tensors:

def basic_fun(x_cloned):
    res = []
    for i in range(len(x)):
        res.append(x_cloned[i] * x_cloned[i])
    print(res)
    return torch.cat(res)


def get_grad(inp, grad_var):
    A = basic_fun(inp)
    A.backward()
    return grad_var.grad


x = Variable(torch.FloatTensor([1, 2]), requires_grad=True)
x_cloned = x.clone()
print(get_grad(x_cloned, x))

Here are my terminal logs:

[tensor(1., grad_fn=<ThMulBackward>), tensor(4., grad_fn=<ThMulBackward>)]
Traceback (most recent call last):
  File "<path>/playground.py", line 23, in <module>
    print(get_grad(x_cloned, x))
  File "<path>/playground.py", line 16, in get_grad
    A = basic_fun(inp)
  File "<path>/playground.py", line 12, in basic_fun
    return torch.cat(res)
RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated

[mhyousefi]

Alright so apparently, I need to do torch.stack(res, dim=0) It produces: tensor([1., 4.], grad_fn=<StackBackward>)

[Florian1990]

I'm having a similar problem, where I cant concatenate the elements in a list of zero-dimensional tensors:

def basic_fun(x_cloned):
    res = []
    for i in range(len(x)):
        res.append(x_cloned[i] * x_cloned[i])
    print(res)
    return torch.cat(res)

By slicing items of one-dimensional tensors you get zero-dimensional tensors that cannot be concatenated. To force getting one-dimensional tensors you can slice x_cloned[i, None].

Side note: I am not sure what you are doing in production, but element-wise multiplication in pytorch is easily done using the * operator:

def basic_fun(x_cloned):
    return x_cloned * x_cloned

[GR4HAM]

Another option is to use unsqueeze to turn a 0-dim tensor into a 1-dim tensor: res.append((x_cloned[i] * x_cloned[i]).unsqueeze(0))

[themis0888]

Well, it seems that last versions supported that operation, but from 0.4, you should unsqueeze the tensors which are the elements of the list. You can do that with

for i in range(len(lst)): list[i] = torch.unsqueeze(list[i], dim = -1)

So the list should look like this. [tensor([0.2895], device='cuda:0'), tensor([0.3895], device='cuda:0')... ]

[wujunyi627]

I can run the code in pytorch in 0.4.1 version. you need to change the triplet.py(reid/loss triplet.py)

[xinyi97]

Hi,

I got the same problem recently. I think it is connected to a newer version of pytorch.

What worked for me is replacing torch.cat with torch.stack, but I am not entirely sure if this solution is unproblematic.

Regards Frank

It's useful