Can't expand to check the data of a Variable when debugging pytorch code

[rardz]

for the pytorch Variable object listed in the debug panel, expand to see its data part shows no tensor which suppose to have in a pytorch Variable, and among the items in data, there is another nested data part which have endless nested data part in it.

• VSCode Version: Code 1.19.3 (7c4205b5c6e52a53b81c69d2b2dc8a627abaa0ba, 2018-01-25T10:32:23.601Z)
• OS Version: Linux x64 4.13.1-gentoo
• Extensions:

Extension	Author (truncated)	Version
python	ms-	2018.2.1
vscode-ai	ms-	0.1.7
cpptools	ms-	0.16.0
rust	rus	0.4.0

---

Steps to Reproduce:

Reproduces without extensions: Yes/No

[DonJayamanne]

Please could you try debugging using the experimental debugger.
Try changing the type from python to pythonExperimental in launch.json.

[rardz]

tried as your advice, but changing to pythonExperimental do not help.
the tensor is still absent in the data, while the data part is still expanding endless nesting.

[DonJayamanne]

• Please could you provide a sample code to replicate this issue (please include the steps as well)
• Please provide the version of program used
• Are you using anaconda etc

[rardz]

• any pytorch code, as long as it includes tensor, could replicate this issue. for example:

import torch
from torch.autograd import Variable

x = torch.randn(2,2)
y = torch.Tensor([1])
z = Variable(torch.ones(2, 3))

print(x, y, z)

in the code above, I can't track the tensor data in x, y, z when debugging with vscode.

• by the version of program used do you mean the vscode or the python extension? for vscode, I tried 1.19.3 and 1.21.1, for the python extension, I tried 2018.2.1 and the recently updated 2018.3.1, all come the same issue.

• yes, it is anaconda.
  but weeks ago, vscode worked well with my os environment, showing the tensor information correctly when debugging. It could hardly be my environment problem since pycharm could debug very well, while vscode fails.

[malmaud]

Can confirm this. The fact that data expands infinitely isn't surprising since data is an accessor that eventually returns self, but it is confusing/annoying that the repr of a torch.autograd.variable.Variable is informative about the Tensor contents of the object while its representation in the VSCode variable display isn't. In PyTorch master and soon-to-be PyTorch v0.4, all tensors are of Variable type and so suffer from this problem, exacerbating this issue even more.

A workaround at the moment is to add eg batch.numpy() to the watch list, since numpy tensors are displayed meaningfully in the VSCode debugger.

[brettcannon]

Flagging as macOS/Linux since PyTorch isn't on Windows.

And can someone provide the repr() and str() of a Variable object so we know what you're expecting and which representation VS Code is showing you?

[malmaud]

Thanks for the quick response! FWIW, PyTorch can be installed easily on Windows via a 3rd-party Conda package.

I've attached a screenshot which both shows a minimal code fragment along with the expected and VS Code representation.

[brettcannon]

@malmaud can you show what is literally returned by repr(hi.y) and str(hi.y)? Since you're using IPython I have a sneaking suspicion it's using Jupyter hooks and that's where the discrepancy is coming from.

[malmaud]

No such luck, I'm afraid:

maltop:~ malmaud$ python
Python 3.6.4 |Anaconda, Inc.| (default, Jan 16 2018, 12:04:33)
[GCC 4.2.1 Compatible Clang 4.0.1 (tags/RELEASE_401/final)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> from torch.autograd import Variable
>>> x=Variable(torch.Tensor([1,2]))
>>> str(x)
'\n 1\n 2\n[torch.FloatTensor of size (2,)]\n'
>>> repr(x)
'\n 1\n 2\n[torch.FloatTensor of size (2,)]\n'
>>> x.__repr__()
'\n 1\n 2\n[torch.FloatTensor of size (2,)]\n'
>>> x.__str__()
'\n 1\n 2\n[torch.FloatTensor of size (2,)]\n'
>>> x

 1
 2
[torch.FloatTensor of size (2,)]

[brettcannon]

Wow, I have never seen a repr with newlines before. 😨

I think part of the problem is how would we represent a multi-line repr in a single line to show in the variables list? If we strip them out then you would have 1 2 [torch.FloatTensor of size (2,)]. Is that useful enough?

[brettcannon]

What do you think @DonJayamanne ?

[malmaud]

I would personally certainly prefer stripping out the new lines to the current behavior.