tensorflow should be a main dependency (instead of a dev dependency)

[metasyn]

I am trying to use scikeras in a poetry based project. I ran

poetry add scikeras

which worked. At runtime, I got an error saying that tensorflow 2.7.0 was needed. This should've been caught at the time I added the package during poetry. The only reason, from my understanding that it wasn't was because tensorflow is a dev dependency, rather than a real dependency, for poetry.

[adriangb]

This is a known issue, and stems from a limitation of how TensorFlow packages itself. The TLDR is that there are other versions of tensor flow like tensorflow-cpu, so if we specified tensorflow as a hard dependency users who want to use tensorflow-cpu wouldn't be able to install SciKeras. For more background, see #221.

[metasyn]

Ah, I see.

#221 (comment)

I think option 1 here is a bit cleaner, but makes sense to go either way to solve the problem. No worries if you just want to close this, I should've read the readme closer. Thanks for the cool package :)

[adriangb]

Yeah I guess we can make TensorFlow an optional dependency and the update everywhere in the documentation to use scikeras[tensorflow] so that at least users copy pasting from the readme won't get caught off guard that easily. A PR for something like this would be gladly accepted :)

[metasyn]

I see. What do you think of adding both?

poetry add --optional tensorflow==2.7.0
poetry add --optional tensorflow-cpu==2.7.0

?