delesley
commented
6 years ago Tensorflow/fold is alive, and is still being used internally within Google. However, the project is currently on the back-burner. What happened is that several of the team members who originally developed tensorflow/fold left to work on other projects, leaving only one person (me) to maintain it. My time is extremely limited, since I am also working on several other projects, and I have simply not had the cycles to respond personally to issues as they arise.
As far as I'm concerned, the Loom and Blocks libraries are essentially feature complete. I'll try to cut a new release sometime soon to bring it up to date with the latest tensorflow, but don't expect any major new updates. The teams within Google that are using Loom seem happy with the current version.
Moving forward, there are a number of performance issues with Loom that cannot be resolved with the current architecture. It works well for relatively simple dynamic computation graphs like tree-LSTMs, but performance rapidly degrades when you start doing things that are more complicated. I have been working on ways to get around those limitations, but it basically involves a complete rewrite of how dynamic batching interacts with tensorflow, and thus will not be compatible with the current Loom API. The Loom and Blocks libraries should remain usable in their current form -- they're open source, and they don't rely on tensorflow internals, so there's no reason for them to stop working. However, within the next year, I would expect to see better ways to write these sorts of models.
I hope that clarifies things,
-DeLesley
On Wed, Aug 16, 2017 at 2:51 AM, Egor Bulychev notifications@github.com wrote:
Hi,
I noticed that none of the issues have been closed during last several weeks. So I have the question - is tensorflow/fold alive? Several last closed issues were closed by people who opened them.
Best regards!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/fold/issues/82, or mute the thread https://github.com/notifications/unsubscribe-auth/AGGbTeTuwin2sxTKdsF6EOA8huT6_UXIks5sYruZgaJpZM4O4rNf .
-- DeLesley Hutchins | Software Engineer | delesley@google.com | 505-206-0315
EgorBu
Hi,
I noticed that none of the issues have been closed during last several weeks. So I have the question - is tensorflow/fold alive? Several last closed issues were closed by people who opened them.
Best regards!