demo docker

[mtrofin]

This is the initial step. It currently still uses the checked in gin files and it doesn't yet fix a git hash for this repo (the latter because we need at least this commit to land first). Also, pre-training (warmstart) isn't implemented yet.

see also this RFC

Issue #540

[mtrofin]

@DataCorrupted @ioghiban

[boomanaiden154]

LGTM.

I'm assuming the plan is to write a README at some point on how to run everything locally?

Do we have performance numbers yet for how well ES training works on this dataset and what sort of performance people should expect?

[boomanaiden154]

Should this be MinSizeRel to ensure we're compiling the modules with -Os or -Oz (I forget which one it is)?

[DataCorrupted]

I agree that it should be "MinSizeRel" if you are training for inlining

[petrhosek]

What's the reason for breaking each step into its own script (which is basically a one-liner)?

[mtrofin]

What's the reason for breaking each step into its own script (which is basically a one-liner)?

Would let someone skip / redo steps. If it's all in a file, there's a tendency that the file would evolve into a monolithic unit, with variables and reusable things and whatnot, and in this case, I want to avoid this.

[mtrofin]

LGTM.

I'm assuming the plan is to write a README at some point on how to run everything locally?

eventually.

Do we have performance numbers yet for how well ES training works on this dataset and what sort of performance people should expect?

No, and we don't yet have its own gin files either.