AttributeError: 'torch.FloatTensor' object has no attribute 'ndim'

[Galto2000]

Howdy folks

relatively new to Python, numpy, PyTorch (come from C++ and Matlab).

I am using CUDA 9.1, Python 3.6, Torch 0.3.0.post4, running on Ubuntu 16.04 LTS and I am starting off as follows:

python train.py --arch segnet --dataset pascal --visdom True

And I get the error below when I want to invoke visdom (i.e. when I don't include --visdom True, I don't have this issue).

  File "/home/bart/anaconda3/envs/pytorch-semseg/lib/python3.6/site-packages/visdom/__init__.py", line 438, in line
    assert Y.ndim == 1 or Y.ndim == 2, 'Y should have 1 or 2 dim'
AttributeError: 'torch.FloatTensor' object has no attribute 'ndim'

Any suggestions anyone?

Thanks in advance

Galto

[FengLoveBella]

@Galto2000 I encountered the same error, did you solve it?

[jiaxiaochong]

visual studio 2017 also meet that problem...

[scyoyo]

I also encounter this problem, change the .cpu() to .numpy() would solve that. Not sure whether it's because the code of visdom is updated from torch tensor to numpy array.

[Galto2000]

Sorry for the late response. I was not able to resolve it personally, but I did the thing scyoyo suggested and that fixed it. I also had to change some dimensions, as shown below:

    if args.visdom:
        vis = visdom.Visdom()

        loss_window = vis.line(X=torch.zeros((2,)).numpy(),
                           Y=torch.zeros((2,)).numpy(),
                           opts=dict(xlabel='minibatches',
                                     ylabel='Loss',
                                     title='Training Loss',
                                     legend=['Loss'])) 

[wisdomk]

that didn't work, any further suggestions?

[ChengYeung1222]

Use .squeeze() or .unsqueeze() builtin function to subtract or add ndims to 1 respectively. This worked in my program.