Discriminative loss error: IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)

[sharmalakshay93]

Hi,

When running python3 lanenet/train.py --dataset ./data/training_data_example, I'm seeing the following exception:

Traceback (most recent call last):
  File "lanenet/train.py", line 156, in <module>
    main()
  File "lanenet/train.py", line 144, in main
    train_iou = train(train_loader, model, optimizer, epoch)
  File "lanenet/train.py", line 68, in train
    total_loss, binary_loss, instance_loss, out, train_iou = compute_loss(net_output, binary_label, instance_label)
  File "/usr/local/lib/python3.6/dist-packages/lanenet-0.1.0-py3.6.egg/lanenet/model/model.py", line 75, in compute_loss
  File "/home/lashar/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/lanenet-0.1.0-py3.6.egg/lanenet/model/loss.py", line 33, in forward
  File "/usr/local/lib/python3.6/dist-packages/lanenet-0.1.0-py3.6.egg/lanenet/model/loss.py", line 71, in _discriminative_loss
  File "/home/lashar/.local/lib/python3.6/site-packages/torch/functional.py", line 1100, in norm
    return _VF.frobenius_norm(input, _dim, keepdim=keepdim)
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)

Any clues about how I can fix this? Not sure if I'm doing something incorrectly.

Thanks!

[sharmalakshay93]

PS: I just saw issues/12, and its related commit. Changing it back to dim=0 makes it work. However, since the aforementioned issue says that the correct value is dim=1, not sure if this works as intended? Would appreciate it if clarification could be provided!

Thanks!

[Flying-Osterich]

I'm having the same problem. Reverting back to dim=0 resolves the issue.

[DongwhanLee]

PS: I just saw issues/12, and its related commit. Changing it back to dim=0 makes it work. However, since the aforementioned issue says that the correct value is dim=1, not sure if this works as intended? Would appreciate it if clarification could be provided!

Thanks!

Hi, I am running your code, and I encountered the same problem. I tried both dim=0 and dim=1, but it doesn't work.

Have been there any updates after your last commit? Please let me know how to fix it and run the train! Thank you.

[mengnutonomy]

I also tried both dim=0 and dim=1, both not working. @klintan would like to know any updates on this? thanks.

[mummy2358]

Managed to fix it. It should be dim=1. The norm should be calculated along the "embedding" axis. Problem comes from

embedding_i = embedding_b[seg_mask_i]

which break the dims to get a single dimensional vector.

Changing it to: embedding_I = embedding_b * seg_mask_i

works for me

[mengnutonomy]

@mummy2358 Thank you. Will give it a try.

[klintan]

@mummy2358 awesome thanks! Feel free to add a PR for the fix :)

[mengnutonomy]

@mummy2358 I do not know why, on my side, after changing to embedding_i = embedding_b * seg_mask_i I still have the same problem. Not sure if it's computer related, as I tried on another pc, it works. And on that pc, revert to dim=0 and w/o aforementioned change, also works.

[william3863-prog]

May i knw where exactly the posotion of dim?

[diazoangga]

@mummy2358 I do not know why, on my side, after changing to embedding_i = embedding_b * seg_mask_i I still have the same problem. Not sure if it's computer related, as I tried on another pc, it works. And on that pc, revert to dim=0 and w/o aforementioned change, also works.

hey @mengnutonomy, it happened to me also. I don't know if this does anything, but it seems you have to run 'python setup.py install' everytime you change your coding. Hope it'll work :)

[Answergeng]

@mummy2358 I do not know why, on my side, after changing to embedding_i = embedding_b * seg_mask_i I still have the same problem. Not sure if it's computer related, as I tried on another pc, it works. And on that pc, revert to dim=0 and w/o aforementioned change, also works.

hey @mengnutonomy, it happened to me also. I don't know if this does anything, but it seems you have to run 'python setup.py install' everytime you change your coding. Hope it'll work :)

thx a lot ! rerun works for me

[xudh1991]

            if len(embedding_i.shape)==1:  
                embedding_i=torch.unsqueeze(embedding_i,0)