Remove tensorflow as dependency #860
Conversation
There was a problem hiding this comment.
Thanks for the PR! Version checking is a good idea.
I am not sure whether we should let the users install tensorflow themselves. The best way would be if we don't install tensorflow only when tensorflow or tensorflow-gpu is already installed. Not sure if this can be done with pip.
This PR can still be merged since kerastuner is still forcing the installation of tensorflow.
So this PR won't make a difference for now except for the version checking.
|
@gabrieldemarmiesse Would you be willing to add a similar change / error message for keras-tuner? I just released a minor kerastuner version (1.0.1) w/o this dependency but haven't updated GH yet, and a check like this would be nice |
|
Sure thing. It's easy enough. I'll do it tonight/tomorrow. |
|
Great, thanks! No rush |
* Removed tensorflow as a dependency. * Moved the check in a function. * Added packaging as a dependency. * Formatted the message better. * Install tensorflow in the CI. * Fixed the version number in the CI * Update setup.py * Update .travis.yml Co-authored-by: Haifeng Jin <[email protected]>
Which issue(s) does this Pull Request fix?
resolves #846
Details of the Pull Request
Tensorflow comes now into two packages: tensorflow and tensorflow-cpu. By including either one or the other in the setup.py's requirements, we force the install on the user's machine.
Since we don't know which one the user wants, we should refuse to install any. To prevent the use of outdated tensorflow packages, we can add a check when users use autokeras for the first time.
This is similar to keras-team/keras-tuner#177 but with an additional check at import time for the tensorflow version.