Roadmap to stabilise `experimental_training` feature?

[Daniel McKenna]

What are the blockers for `experimental_training` feature being stabilised? It would be good if we could remove the `Vec<u8>` and just replace it with the `MetaGraphDef`, that would also open things up to implement some of the python utils like graph freezing and exporting modified graphs. I'm happy to contribute and try and push this forwards.

Also as a side note would it be possible to use github issues for tracking things like bugs, feature requests and tasks planned before each release? And keep the google group for people just looking for support in getting going? Just because keeping everything within github does help keep the workflow nice :)

[Adam Crume]

There weren't specific blockers for stabilizing features guarded by `experimental_training`; I just wanted to make sure we've kicked the tires a bit before marking it as stable, since I was concerned that we might need to make some backwards-incompatible changes to it (and I think I have).  I was planning to remove the need for the `experimental_training` feature before the next release.  If you want to submit a PR to remove `#[cfg(feature = "experimental_training")]` everywhere, I'll be happy to review it.  I don't want to directly remove the SavedModelBundle::meta_graph_def field (if that's what you're referring to), because that could break users, but we could mark it as deprecated.

We are using GitHub issues for bugs and feature requests, although perhaps not as consistently as we should.  Are you suggesting using something like labels or milestones to track which specific issues are expected to be in each release?  For this to work well, we'd need an actual planning process, and there currently isn't one.  One difficulty is that we can't dedicate a certain amount of time to work on each release, because to the best of my knowledge, contributors (including myself) are working on this on a best-effort basis.  I'd like to get a release out soon, but we can try planning the next one.  I'll come up with an RFC for a hopefully lightweight process.

On Mon, Aug 3, 2020 at 3:44 AM Daniel McKenna <danielm...@gmail.com> wrote:

What are the blockers for `experimental_training` feature being stabilised? It would be good if we could remove the `Vec<u8>` and just replace it with the `MetaGraphDef`, that would also open things up to implement some of the python utils like graph freezing and exporting modified graphs. I'm happy to contribute and try and push this forwards.

Also as a side note would it be possible to use github issues for tracking things like bugs, feature requests and tasks planned before each release? And keep the google group for people just looking for support in getting going? Just because keeping everything within github does help keep the workflow nice :)

--
You received this message because you are subscribed to the Google Groups "Rust for TensorFlow" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rust+uns...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/rust/8c5cf261-bae1-41cf-9a3e-f4c0eb55d707o%40tensorflow.org.

[Daniel McKenna]

yeah I meant the meta_graph_def field. Marking it as deprecated it sounds like the right move :)

Yeah labels and milestones would be good, something to point contributors towards blockers for releases. Alternatively, could go for a more "agile" release early and release often. I'm just partially thinking about the gap between 0.14.0 and 0.15.0 and the fact there were some useful features in the master branch but not released while more work was done.

And yeah I'll happily do the PR I was looking for some opportunities I could contribute, I use this crate a lot and was looking to contribute some PRs to address features we miss from the python version etc 

[Daniel McKenna]

Also I've just submitted a PR :) https://github.com/tensorflow/rust/pull/267 I'll look to what other things I can do with the experimental graphdef such as replicating graph freezing