Core: Tensor Flow not working in latest docker image HA version 0.113 (AMD64)

[jsb54321]

The problem

It looks like the docker image is having trouble building tensowflow==1.13.2 for AMD64. HA Log is showing the following error:

Unable to install package tensorflow==1.13.2: ERROR: Could not find a version that satisfies the requirement tensorflow==1.13.2 (from versions: none) ERROR: No matching distribution found for tensorflow==1.13.2

Environment

HA Version Docker 0.113 (latest)
Intel i7-2600K CPU (AMD64)
Ubuntu 18.04.4 LTS

Bashing into the docker and running pip3 install tensorflow==1.13.2 generates the same error.

Problem-relevant configuration.yaml

Traceback/Error logs

Additional information

[oriolism]

Im having the same issue in my NUCi3 running Home Assistant(Hassio).

[Mark612]

Same issue, same errors. NUC using Docker.

[MartinHjelmare]

The current situation is this:

• Python 3.8 is now the default Python version for our docker images.
• There are no wheels for Python 3.8 at version 1.13.2 of tensorflow which is the currently pinned version of tensorflow library in the tensorflow integration. Wheels for Python 3.8 start at version 2.2 of the tensorflow library.
• The library version can't be bumped immediately as there's a major version bump from 1 to 2, which contains breaking changes that need to be reviewed.
• The code owner is aware and working on resolving this as soon as possible.

[jsb54321]

Thank you for the explanation. If testing needs to be done, I'd be glad to help.

[benroscoe]

Assuming such a major jump in revision is a lot of work and might take days/weeks, it might be a good idea to flag this release as a majorly broken for those with tensorflow.
If you have tensorflow and ecobee, you're in for tons of fun reverting back to because ecobee also has that weird API refresh bug.

[MartinHjelmare]

The release notes have been updated:
https://www.home-assistant.io/blog/2020/07/22/release-113/#read-more

[hunterjm]

@pvizeli has built TensorFlow 2.2 for homeassistant wheels. I'm currently working on updating the integration. The upgrade will require some work on the user's end as well because the 1.x object detection models are not compatible with 2.x either. The new model zoo is here.

Initial changes are made, but it's taking 6 minutes to load the saved model and it's a synchronous function. Will need to do some additional work to ensure it doesn't block startup.

[pvizeli]

Is to use TF lite an option or use the REST API and we make an add-on like we did on the new ozw?

[hunterjm]

@pvizeli - I don't see any official TF Lite 3.8 builds. Either way, TF Lite running in a separate docker container (addon) is basically what the DOODS integration does.

I have TF2 working and have updated it to not block startup. Doing some cleanup then will raise a PR.

[DeadEnded]

Upgraded to 0.114, error persists - reopened my previous issue and updated details:
https://github.com/home-assistant/core/issues/38282

[Mark612]

Upgraded to 0.114, error persists - reopened my previous issue and updated details:

38282

I get the same error, running in docker.

[DeadEnded]

@Mark612 the issue is a change in the HA Docker file. If you are using Portainer like I am, you need to remove the container, and recreate it from CLI instead of having Portainer recreate it. Alternatively, change over to Docker-Compose. Once I spun it up without portainer it worked just fine.

DeadEnd

[Mark612]

That worked! Using cli and recreating everything did the trick. Thank you for the tip.