namespaces torch.cuda, torch.device missing from pyi

[soumith]

🐛 Bug

Auto-complete / type information for torch.cuda / torch.device are missing, as reported at #7318 (comment)

To Reproduce

Auto-complete torch.cuda.<TAB> in pycharm

[zhezh]

torch.device() torch.tensor() are still missing, I believe they are much frequently used.

[fengyanghe]

torch.randn() and torch.FloatTensor() are still missing in PyTorch 1.0.1, related to vs code and wing ide

[t-vi]

torch.FloatTensor is deprecated. Unfortunately, there doesn't seem to be a good mechanism to incorporate deprecation information into type hints, currently. At least that was the case a couple of months ago when I searched for that a bit.

[Wenlong-Shi]

torch.from_numpy()、torch.ones are still missing in 1.0.1.post2，related to Pycharm。

[fmassa]

@Wenlong-Shi I'd say torch.from_numpy is also kind of deprecated, in favor of torch.as_tensor

[Wenlong-Shi]

@fmassa torch.as_tensor is also missing in 1.0.1.post2，related to Pycharm.

[Luvata]

torch.no_grad() is also missing in in 1.0.1.post2, related to Pycharm

[670373064]

torch.cuda（）、torch.utils is also missing in in 1.0.1, related to Pycharm

[magnusjacobsen]

on 1.0.1, pylint claims torch.tensor() is not callable

[Ajk4]

On pytorch 1.0.1 on Windows I don't see torch.nn package. When ran code works just fine

[malmaud]

I'd be happy to manually add the stubs for some of these missing namespaces and submit those as a PR. Would that be acceptable, or are you looking for a more automatic codegen approach?

[t-vi]

Thank you for offering to work on these! I'm not sure whether there / what the are plans for a maintenance release, but a great first step would to fix these in master / the nightlies.

• Things related to the automatically wrapped functions would best go in the generation
  https://github.com/pytorch/pytorch/blob/master/tools/pyi/gen_pyi.py
• Things like imports and pure Python functions could go here: https://github.com/pytorch/pytorch/blob/master/torch/__init__.pyi.in

[imyhxy]

I am using torch==1.0.1.post2 and Pycharm professional 2019.1. I tested and sorted out the situation mentioned by everyone in front.

can't find reference

• torch.cat
• torch.clamp
• torch.exp
• torch.FloatTensor deprecated
• torch.from_numpy deprecated
• torch.log
• torch.LongTensor deprecated
• torch.max
• torch.min
• torch.ones
• torch.randn
• torch.reshape
• torch.sum
• torch.zeros

not callable

• torch.tensor
• torch.as_tensor

[Dessix]

The message in 1b25fdb does not appear to have intended to close this issue.
"This would close #16996, although comments on that issue reference other missing stubs so maybe it's worth keeping open as an umbrella issue." @malmaud - should this be reopened?

[malmaud]

Maybe better to just open a new issue with the things are still missing from the stubs. torch.cuda and torch.device, which is the title of this issue, have been added.

[malmaud]

I went ahead and opened new issues for nn and util.data. The other missing methods mentioned in this issue are all fixed on master but aren't on a release yet.

[zhutmost]

I have updated to the latest version (1.1.0). However, this bug is still not fixed.

[malmaud]

What specifically do you mean by 'this bug'? Like, which methods? I did a fresh install of 1.1.0 and all the new type stubs have seemingly been installed correctly.

[lianghongzhuo]

can't find reference in 1.1.0, some examples:

torch.cuda.manual_seed()
torch.utils.data.dataloader.default_collate()

[bcrchzz]

What specifically do you mean by 'this bug'? Like, which methods? I did a fresh install of 1.1.0 and all the new type stubs have seemingly been installed correctly.

@malmaud

for example, torch.arange, torch.zeros, torch.cat

[lianghongzhuo]

@bcrchzz torch.arange, torch.zeros, torch.cat works fine in 1.1.0 with the PyCharm IDE.
Did you update your pytorch?

[lianghongzhuo]

What still cannot find the reference now is torch.data in the 1.1.0 version.

[bcrchzz]

@bcrchzz torch.arange, torch.zeros, torch.cat works fine in 1.1.0 with the PyCharm IDE.
Did you update your pytorch?

@lianghongzhuo Both my pytorch and pycharm are in the latest version.

[bcrchzz]

Besides, I find more functions, torch.sqrt/zeros_like/where, which cannot be referred correctly.

[malmaud]

As @lianghongzhuo says, the type stubs for torch.arange etc have been successfully used in pycharm by others and me. If there still an issue for you, it must be on the PyCharm side - there's nothing more that can be done on the PyTorch side than shipping working type stubs.

[bcrchzz]

@malmaud Thank you a lot. I re-configured the python interpreter in the PyCharm Preference and reopened it. And the problem solved.
I use a remote python interpreter installed on a server. And I guess Pycharm may cache something locally, which makes it cannot refer the right version for lint.

Thanks again for your great jobs.

[DavidRuhe]

@bcrchzz I also use a remote interpreter. How did you reconfigure the remote interpreter? I removed it and re-added it but still have the reference problems.

Also: did you install pytorch from root, conda or pip?

[Genius1237]

I get errors for torch.tensor,torch.float, torch.ones, torch.zeros and torch.zeros on 1.1.0 installed from pypi on VS Code.

[DavidRuhe]

For me torch.log, torch.stack, torch.exp, torch.randn, torch.empty and torch.FloatTensor do not work yet on 1.1.0 installed from pypi.

[malmaud]

I think there's no doubt that the stubs of the torch.* methods are being installed with 1.1.0. If your editor isn't making use of them, the issue must be with your editor and its interaction with your Python environment - there isn't anything more that can be on the PyTorch side.

[jakkes]

Using torch v1.1.0 and VSCode v1.33.1 (latest) with python 3.6.7, pretty much all torch.* methods do not show up (including all posted by the @imyhxy further up), nor is torch.tensor callable. It works fine in pycharm (apart from the callability of torch.tensor).

[Whisht]

I found that the stubs of torch.* works well in pycharm python console, but still cannot find reference in the editor. Even I reinstall pycharm as well as remote python interpreter. So confusing.