@ccshan when should we expect you?

Getting coffee downstairs, sorry

ah, we are in my office

@rjnw This has become n.adapt = 0, right? https://github.com/rjnw/hakaru-benchmarks/blob/3b4263c2f5facb6bacf70246fdddc8fcc9e68e29/testcode/jagssrc/NaiveBayesModel.R#L61

Yes I changed that but haven’t commited.

(I just pushed https://github.com/rjnw/hakaru-benchmarks/blob/0ab8cd764092193faab685971502fd287277d324/testcode/hksrc/NaiveBayesLikelihood.hk)

It turns out that PSI exact-inference benchmarks aren’t too hard to run. https://github.com/eth-sri/psi

for the future this has some nice potential benchmarks

@rjnw How’s it going? Sam and I think you’re working on the 6 figures (2 GMM, 2 NB, 2 LDA). Adding to the exact inference evaluation comes later.

It’s going benchmarks are running. I fixed the gmm 50, I fixed the issue that we had in the morning.

I changed time settings for Lda and accidentally broke gmm.

Also is 30seconds okay or should I add more time?

Although 30 seconds is enough, I personally think it’s better to show (if within say 100 seconds) where the Haskell backend reaches the same plateau as the LLVM backend, just more slowly.

Also, please use the name AugurV2 throughout. It’s supposedly quite different from Augur (V1), and so they consistently call it AugurV2 (one word, with A and V capitalized), not Augur.

Thanks for the update! Please keep us posted.

So I’ll be in a plane all day tomorrow — but I’ve got the icfp reviews and other comments on the paper on my laptop. So I intend to take care of those during that time. [I’ve dealt with all my other deadline-driven obligations, so this one’s at the top now.]

@rjnw Can you resend the Docker link? It seems to have expired (i.e. is more than 10000 messages ago on this slack. I had saved a link to slack…)

https://drive.google.com/file/d/10ieuJjtKazLlmBKG38n72lvLZlVNRypy/view?usp=sharing

Thanks. Downloaded (and I’ve got Docker as well).

Hopefully that’s enough for me to work offline…

There might be some issues in initializing, where it fetches the latest version from github. There is an older commit of hakaru in the docker image which should have the same issue.

Also there was a specific docker command to run this image with maple working, let me see if I can find it.

I can try running it for the first time while I wait at the gate (there’s free wifi at the airport). Hopefully that will do it.

Yes please!

[I’m going offline now, but I’ll check Slack before I go tomorrow morning… 6am, yuck.]

okay, I will post the commands here.

sudo docker import data/hakaru-docker.tar.gz testshare
sudo docker run --entrypoint "/init.sh" --mac-address 1c:1b:0d:e3:fc:22 -it testshare

It will download a lot of stuff in the beginning especially latest ghc and stuff for compiling hakaru

you can Ctrl-C it, and the older compiled files should be in hkbin folder, the hakaru code is in testcode/hksrc/. It starts in hakaru-benchmarks directory and maple binary is /hakaru-benchmarks/maple2017/bin/maple

The mac address is my personal machine as the maple license works with that mac, I hope you are not going to use it for any nefarious purposes. :sweat_smile:

I plotted with sweeps as sweep time is 45s vs 0.2s

*with number of sweeps