0.3.2 point release.

[vincentqb]

Create 0.3.2 point release.

1. create release branch with updated version number (e.g. 0.3.2 point release. #354)
2. run audio tutorial as test
3. bump master version (e.g. move master to 0.5.0 #414)
4. make sure the circle ci upload jobs are running on the pull request (e.g. Update version number to v0.4.0 #410)
5. generate wheels in 0.3.1 point release #308
6. generate conda binaries in 0.3.1 point release #308
7. update readme.
8. update docs on gh-pages branch using the 0.3.2 branch. (see here for 0.6)
9. push release notes
10. update fbcode internally, see here

Permissions needed for point release:

• Admin on github repo
• pytorch-builders
• Write to Amazon S3 pytorch bucket
• Maintainer on pypi
• Maintainer on conda pytorch channel
• Need CircleCI secrets?

Fixes #338.

cc @seemethere

[vincentqb]

For reference, updated .circleci/.config.yml.in, ran .circleci/regenerate.py, and added the changes.

[vincentqb]

@ezyang -- I do not have permissions to upload to S3, is that the case? I see the following errors in the build.

nightly_binary_linux_wheel_py3.5 - Success
nightly_binary_linux_wheel_py3.5_upload - Unauthorized

And I do not see torchaudio-0.3.2-* in the amazon bucket.

[ezyang]

I've added you to pytorch-builders which should give you permissions to do so

[vincentqb]

I've added you to pytorch-builders which should give you permissions to do so

@ezyang -- The tests now pass :) However, when I publish the wheels, I get an access denied error, see below. Note that the file does exist.

❯ cd builder/wheel
❯ ./publish_wheels.sh torchaudio-0.3.2

fatal error: An error occurred (404) when calling the HeadObject operation: Key "whl/nightly/torchaudio-0.3.2-cp27-cp27m-linux_x86_64.whl" does not exist

❯ aws s3 cp s3://pytorch/whl/nightly/torchaudio-0.3.2-cp27-cp27m-linux_x86_64.whl s3://pytorch/whl/torchaudio-0.3.2-cp27-cp27m-manylinux1_x86_64.whl --acl public-read

copy failed: s3://pytorch/whl/nightly/torchaudio-0.3.2-cp27-cp27m-linux_x86_64.whl to s3://pytorch/whl/torchaudio-0.3.2-cp27-cp27m-manylinux1_x86_64.whl An error occurred (AccessDenied) when calling the CopyObject operation: Access Denied

as done here.

[ezyang]

You need write access to the pytorch S3 bucket. @soumith can hook you up.

[seemethere]

@vincentqb You should now have access to push to the pytorch S3 bucket

[vincentqb]

@ezyang I've generated the wheels and pushed them on pypi. For conda, builder_conda/publish_conda.sh doesn't find the files (for torchaudio and torchtext, ok with torchvision) on conda (with conda search) and so the script isn't successful. Since the script was successful for 0.3.1, I'm wondering if you ran another script when preparing the binaries for conda. Which one would prepare/upload the tar.gz to conda?

[ezyang]

It looks like you waited too long and the release builds got GC'ed, at least, that's what this seems to suggest: https://anaconda.org/pytorch-nightly/torchaudio/files?page=3

[vincentqb]

It looks like you waited too long and the release builds got GC'ed, at least, that's what this seems to suggest: https://anaconda.org/pytorch-nightly/torchaudio/files?page=3

Good point, reran CircleCI. Upload is currently failing though.

[ezyang]

We may be out of space on the org again. anaconda/anaconda-client#501 describes some situations which cause failure

[vincentqb]

We may be out of space on the org again. Anaconda-Platform/anaconda-client#501 describes some situations which cause failure

@ezyang

• Logged in through ssh on one of the upload circleci.
• Applied patch

@@ -244,6 +244,9 @@
                 raise
             else:
                 logger.info('Distribution already exists. Skipping upload.\n')
+        except Exception as e:
+            logger.info('Uncaught exception: %s' %  (str(e),))
+            raise
 
         if upload_info:
             logger.info("Upload complete\n")

from comment to

/opt/conda/pkgs/anaconda-client-1.7.2-py27_0/lib/python2.7/site-packages/binstar_client/commands/upload.py

• Reran upload command from .circleci/config.yml:

(base) root@70309b619fef:~# anaconda upload ~/workspace/torchaudio-0.3.2-py37.tar.bz2 -u pytorch-nightly --label main --no-progress --force
Using Anaconda API: https://api.anaconda.org
Using "pytorch-nightly" as upload username
Processing '/root/workspace/torchaudio-0.3.2-py37.tar.bz2'
Detecting file type...
File type is "conda"
Extracting conda package attributes for upload
Creating package "torchaudio"
Creating release "0.3.2"
Uploading file "pytorch-nightly/torchaudio/0.3.2/linux-64/torchaudio-0.3.2-py37.tar.bz2"
[WARNING] Distribution "linux-64/torchaudio-0.3.2-py37.tar.bz2" already exists. Removing.
Upload complete

conda package located at:
https://anaconda.org/pytorch-nightly/torchaudio

• File is listed on link in output :)

[vincentqb]

Triggering upload worked this time. Successfully pushed to conda.

[vincentqb]

Binaries are now available. Self-approving PR and merging.