Reinstate torchaudio tutorial

[vincentqb]

Let's see if the tutorial loads correctly with current binaries. pytorch/audio#307

This partly reverts commit 9cee825 from #685.

CC @SethHWeidman @ezyang

[netlify]

Deploy preview for pytorch-tutorials-preview ready!

Built with commit 0b5aa0d

https://deploy-preview-687--pytorch-tutorials-preview.netlify.com

[SethHWeidman]

Thanks @vincentqb! I'll make sure this is merged as soon as the torchaudio issue is fixed. Glad you were able to reproduce that one.

Will hold off on merging until the new tutorials we're releasing today are merged, though.

[vincentqb]

@SethHWeidman -- I'd like to see if the latest version of torchaudio fixes the original issue. However, I see a lot of tests with "ci/circleci: pytorch_tutorial_build_manager Expected — Waiting for status to be reported" that are not running. Is that expected?

[SethHWeidman]

Hey @vincentqb, I can re-run those build scripts in Circle since I'm an admin of this repo, no prob.

However, I'm still getting the same error when I import torchaudio if I have version >=1.3.0 of torch installed. Are you able to reproduce this? If not I can walk through with you on a VC.

[vincentqb]

Hey @vincentqb, I can re-run those build scripts in Circle since I'm an admin of this repo, no prob.

However, I'm still getting the same error when I import torchaudio if I have version >=1.3.0 of torch installed. Are you able to reproduce this? If not I can walk through with you on a VC.

As mentioned in #307, I am unable to reproduce the error. I've pinged you offline :)

[SethHWeidman]

Ok, trying this - I do hope it works!! Will re-add the audio tutorial if so.

I suspect that it won't though: the reason is our tutorials pip install all the libraries in requirements.txt before running the code in the tutorials; and when I pip install torchaudio locally (or on my devserver), pip still force uninstalls torch 1.3.0 and installs torch 1.2.0 instead.

Is it possible we've fixed the issue on a development version of torchaudio, but not the version that is on PyPi? Seems to me like that version is still incompatible with torch 1.3.0.

[SethHWeidman]

We're still getting the error when importing torchaudio: https://circleci.com/gh/pytorch/tutorials/18830

[vincentqb]

I suspect that it won't though: the reason is our tutorials pip install all the libraries in requirements.txt before running the code in the tutorials; and when I pip install torchaudio locally (or on my devserver), pip still force uninstalls torch 1.3.0 and installs torch 1.2.0 instead.

Is it possible we've fixed the issue on a development version of torchaudio, but not the version that is on PyPi? Seems to me like that version is still incompatible with torch 1.3.0.

Indeed. The version on pip needs to be updated to 0.3.1, here. If this is indeed the issue, I would expect an error in the logs about incompatible versions though.

[vincentqb]

The version seems to be the problem. In linux:

conda create -n pytorch-pip python=3.7
conda activate pytorch-pip
pip install torchaudio
python -c "import torchaudio"
# success

pip install torch==1.3.0
# upgrades torch from 1.2.0 to 1.3.0 with a warning
python -c "import torchaudio"

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/private/home/vincentqb/anaconda3/envs/pytorch-pip/lib/python3.7/site-packages/torchaudio/__init__.py", line 5, in <module>
    import _torch_sox
ImportError: /private/home/vincentqb/anaconda3/envs/pytorch-pip/lib/python3.7/site-packages/_torch_sox.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZN3c1012CUDATensorIdEv

[SethHWeidman]

Great, @vincentqb, then at least we know the fix: let's get version 0.3.1 on PyPi!

[vincentqb]

The code snippet now works.

conda create -n pytorch-pip python=3.7
conda activate pytorch-pip
pip install torchaudio
python -c "import torchaudio"
# success

pip install torch==1.3.0
# already satisfied

Let's see if the tests pass.

[SethHWeidman]

@vincentqb merging!

@ezyang, to prevent this kind of thing from happening in the future, let's make sure we release versions of the domain libraries to PyPi along with our major releases.