search

tensorflow / swift-models

Models and examples built with Swift for TensorFlow
Apache License 2.0
647 stars 147 forks source link

Convert image classification and generative examples to use X10 by default #514

Open BradLarson opened 4 years ago

BradLarson commented 4 years ago

With the release of the 0.9 toolchain containing X10, we now have a higher-performance alternative to the default eager execution mode. In the interest of having swift-models demonstrate best practices, we should have our image classification and generative models use the highest performance pathway available. For image classification and image generative models, that's pretty much always X10.

Therefore, we should make the (small) modifications needed for the following models to use X10 by default:

This is most likely a follow-on issue to issue #511, because the conversion of training loops to Epochs will conflict with updates to the training loops here. We also may need to check that X10's CPU compile times and execution speeds are acceptable for macOS.

Andrew-Pynch commented 4 years ago

@BradLarson Hello, I am the guy who wanted to help with this from this morning's design meeting. I am still very new to this codebase. Where do these models live so I can start migrating them? :smile:

BradLarson commented 4 years ago

@Andrew-Pynch - Great to hear about the interest. I've added links to each entry to describe where it lives in the repository. The only caution I'd have is that you'll want to coordinate with @xihui-wu who's working on issue #511, and not migrate any models she hasn't yet updated there. Any models not on that list, or ones that have already been updated to Epochs, should be free to update here.

Andrew-Pynch commented 4 years ago

@xihui-wu I am going to use my weekend to work on these migrations. I was hoping I would have some time this week but that is not the case :-(. Is this a time-sensitive task? I don't have a gauge for how long it will take to do this but I will get as many done as I can this weekend :-)

xihui-wu commented 4 years ago

Take your time @Andrew-Pynch #511 is up to date so you're good to work on anyone checked there.:)

BradLarson commented 4 years ago

I'll note that the following models:

will automatically be updated by PR #586, so we can hold off on this until that lands. The others could potentially be converted as a second stage using that training loop once that is available to the project.