Does the readme mention deleting the logs?

no the make commands delete them automatically

ok

Ok @ccshan I am trying to debug that Maple problem from too-many.mpl. Which version of Maple are you using?

I’ve got 2018 installed, and the symptoms are different. [After 5 minutes of computation, it’s still going, even though the problem is “easy”].

The build ID is in the file I provided… It says Maple 2018 at startup

I’ve got 2018.2, with later IDs for both.

Not sure what you mean by “both”, but I hope I’ve answered your question.

You have.

One of the issues is that, the ‘raw’ integral (outside the product) returns a limit.

How long does it take, on your system, go get that error?

“quit” says “time=30.50" so I guess that’s the total time: about 10 seconds on the top-level call to ints and about 20 seconds on the top-level call to simplify_assuming

If you just do int(x^(Sum(piecewise(i = Hakaru:-idx(y, j), 1, 0), j = 0 .. n-1)), x = 0 .. 1)

in a maple session, does the answer have a limit in it?

Yes.

Ok, good, at least things aren’t that divergent.

Nevertheless, the issue with that is that the product will never compute and may indeed cause an enormous computation, of which the stack overflow may just be a symptom.

BTW maybe it’s relevant to recall that Loop:-ModuleLoad overrides csgn and sum…

Ah, indeed, that is a good thing to remember. Let me trigger those and try again.

with(Loop) didn’t change it for me

with(NewSLO); with(Hakaru); did, for me.

But the result stayed the same.

The fundamental problem is that ‘something’ can’t tell that (Sum(…)+1) where the Sum is over 0/1 stuff is in fact strictly positive.

That sounds like what the csgn override was supposed to fix.

Indeed - but in this case, it’s not called at all. signum is used instead.

So I’ll try to use the same trick and see if that helps.

And the results are mixed. I can get that signum to return. But the computation is still stuck inside int for me.

And I mean ‘stuck’, I have still not been able to reproduce the symptom you get!

I am currently quite befuddled by this. It doesn’t feel like I am making any progress at all.

This exactly seems to be triggering all sorts of Maple problems.

At this point, I’m kind of giving up. I’ve reported a bug to Maplesoft, let’s see what they say.

I’ve got some ideas of how I might be able to hack my way around this, in our front-end to int, but I don’t think I want to try to do it before our today deadline, as I run way way too high a risk of breaking things.

@rjnw In my VirtualBox I ran gmm-25 and nb, and it generated 3 PDF files as expected. But in the NB plots, the JAGS curve seem to have only one point on it, at time 0. Can you at least double-check that the plots you generated are like the ones currently in pipeline.pdf?

yeah, I feel the same.

Yes I updated the configuration for that recently in Main.hs in hakaru-benchmarks/runners/hk/NaiveBayesGibbs, is it possible for you to verify the change just by looking at knobs variable?

@rjnw does that mean that the image we downloaded is out of date? or that the image needs to be updated?

Just need to run git pull

I will upload a fresh image tonight to submit

I created a checklist of things to do here: https://github.com/rjnw/hakaru-benchmarks/issues/1

@ccshan and I came up with a list of things to add to the README

Thanks! I just did git pull and it will take a while to check that it produces a similar figure to the paper.

I still don’t know if our ‘artifact’ only includes the compiler back end or the whole thing…

I’m going to take a first stab at writing inside README-ARTIFACT.md to describe the hs->hk->hkr->rkt->pdf pipeline and say that hk->hkr requires Maple and if you happen to have Maple then here’s how you can run that step and here’s a small program to simplify using that Maple

(That’s on the to-do list now that Sam created.)

Ok, good.

I can report that the installation instructions for the Maple part of Hakaru at http://hakaru-dev.github.io/intro/installation/ work, as I had to perform them on a new machine earlier today.

I drafted https://github.com/rjnw/hakaru-benchmarks/commit/9f857ca4d62f4b47992fa1c97411a817f5513537 and it is ready for people to improve. In particular, it could use an explanation of where our test data comes from.

@carette our artifact include all the code you get by cloning the hakaru-benchmarks repository

I am editing

It seems that links to the original Hakaru but doesn’t “include” it?

it contains the hakaru repository as a git submodule

So does that mean it’s an artifact or not?

all the submodules are checked out when you check out the hakaru-benchmarks repository

I made some revisions, most significantly a section reorganization

It would be great to be able to run @ccshan’s simple example with some Racket/Sham code, but I think @rjnw would need to write that

It would also be great to add that example to the repository (and thus artifact).

I did the one remaining checklist item that I can do

@rjnw I think we’re all mostly waiting on you, so if there’s something we can do please let us know

I don’t have maple installed so if someone can generate the hkr file and check it in I can try to write the racket code for that

Isn’t the output also there in the example?

not hkr

what about the other issues in the checklist here? https://github.com/rjnw/hakaru-benchmarks/issues/1

the most signficant is the exact inference

I will work on the exact inference

for both clinical trial and linear regression I just print time for 20 runs on terminal is that okay?

yes, but add that to the readme

done, I created a link in the desktop folder as well so it is the first thing they see.

awesome

I just did a little more editing

let’s check with @ccshan and then upload/submit?

should we just upload the vm image or tar it as well?

I think just upload the VM image

Also update the readme on the artifact submission site

I’ll do that now

I am exporting the image it takes 45 minutes

@rjnw will the URL stay the same?

and then will upload it to drive which takes some more time

I don’t know, I can try

yes I can right click and manage versions and then upload a new one

ok

everything should be done, except the md5 and sha1 hashes of the final artifact

you can add those here: https://icfp2019aec.hotcrp.com/paper/20/edit

(and if the URL has to change for some reason)

I’ll trust @rjnw to do that, and I’m off to bed

sounds good, I will do that

Hey hey, sorry I was away, does any hkr need generated or anything else?

I am already uploading the latest version, I looked at the example and I think we don’t support the plate and bind forms in the racket implementation yet.

Ah ok so all our inference benchmarks (both approximate and exact) sample a few scalars at a time, not array

btw did you get the output from NaiveBayes with new jags config?

Still waiting for that…

new version is still uploading but I updated the sha and md5 hashes

Speaking of waiting, how’s your visa application going Rajan?