Closed virajbshah closed 2 months ago
A few thoughts on this:
tf.estimator
API is used only by the ranking loss, which was an experiment that didn't really work well (the MAPE/MSPE losses work a lot better). If it's still the case, I'd be for removing the ranking loss and getting rid of the tensorflow-ranking
dependency which will also simplify our dependency structure and deployment process.WDYT? And would you like to take a look at this?
Removing tensorlow-ranking
should be an easy choice if the ranking loss is not very effective, especially since it'll make installation/set-up a little simpler since there won't be a separate process for MacOS.
I can look into it, but I think it will stay low priority since it isn't blocking anything (other than TF 2.16).
Turns out removing tfr
was very straightforward. I have left the PR (#128) with the corresponding changes as a draft in case we think any further discussion about whether it should be kept around or not is in order.
It seems reasonable enough to me, and not being able to upgrade to TF 2.16 I think is pretty good motivation to just remove the functionality. I'd say open the PR for review and ping reviewers.
Fixed by PR #128
Python 3.12 supports only TensorFlow 2.16 - which drops the
tf.estimator
API, used by the current version oftensorflow-ranking
used in Gematria. Untiltensorflow-ranking
gains compatibility with TensorFlow 2.16 (and hence Python 3.12), we'll have to stick with Python 3.11 and TensorFlow 2.15.