The repository still has one with y-axis 0–100, what should I do with legend for 45–85%?

Yeah, well, can you move the legend up?

(to the middle vertically)

Given that the AugurV2 curve doesn’t move, I prefer this version over the one in the git repository even if the legend stays in the lower-right corner translucent. Similarly for NaiveBayesGibbs-Likelihood. But this issue is minor.

It’s more important to have PSI plots. But I know you have to go soon!

I am still here for a couple hours, I will see if I can start the psi plots while I get ready.

Can psi take command line arguments?

I don’t see a way…

(and that might mean generating psi code with different n on the fly, I understand)

I think so, what is the range you want for clinicalTrial?

It seems that [10,20..100] would stress it enough?

(I mean, it seems that it gets pretty bogged down at n=100 already)

How many times should I run it? and is it okay if I just give you the raw data?

@rjnw we can definitely turn raw data into plots

Sure, why don’t you start with one run and give me the raw data, then if you have more time then do more runs and plot them with error bars? I say this because I can throw together a plot in tikz but I don’t have a quick way at hand to plot the error bars the way you do them nicely.

Okay, it’s running now let’s see how long it takes for one run.

I also pushed new naivebayes plots with suggested axis and legend position

looks good to me

@rjnw Table 1

needs LDA startup times for LLVM-backend and AugurV2, no?

Oh yeah forgot, doing it now.

update on psi clinical trial, it increases 2x every +10, with 787s for 80 and running currently for 90. I am not running for 100. I think I can run it two more times while I pack.

what about psi linear regression?

I’m not sure you need to run psi clinical trial more than once. I’d rather you run psi linear regression once, for n=10,20,..,90

clinicaltrial ../../testcode/psisrc/ct10.psi 1.61s user 0.02s system 99% cpu 1.632 total ../../testcode/psisrc/ct20.psi $bin $i 5.82s user 0.04s system 99% cpu 5.867 total ../../testcode/psisrc/ct30.psi $bin $i 15.44s user 0.11s system 99% cpu 15.571 total ../../testcode/psisrc/ct40.psi $bin $i 50.39s user 0.30s system 99% cpu 50.741 total ../../testcode/psisrc/ct50.psi $bin $i 101.33s user 0.34s system 99% cpu 1:41.79 total ../../testcode/psisrc/ct60.psi $bin $i 224.49s user 0.78s system 99% cpu 3:45.58 total ../../testcode/psisrc/ct80.psi $bin $i 787.87s user 2.01s system 99% cpu 13:10.96 total ../../testcode/psisrc/ct90.psi $bin $i 1434.59s user 3.01s system 99% cpu 23:59.44 total

Cool, what about 70?

linear regression psi lr ../../testcode/psisrc/lr10.psi $bin $i 0.41s user 0.00s system 99% cpu 0.413 total ../../testcode/psisrc/lr20.psi $bin $i 0.63s user 0.02s system 99% cpu 0.653 total ../../testcode/psisrc/lr30.psi $bin $i 0.91s user 0.01s system 99% cpu 0.923 total ../../testcode/psisrc/lr40.psi $bin $i 1.24s user 0.02s system 99% cpu 1.258 total ../../testcode/psisrc/lr50.psi $bin $i 1.59s user 0.03s system 99% cpu 1.618 total ../../testcode/psisrc/lr60.psi $bin $i 1.87s user 0.02s system 99% cpu 1.891 total ../../testcode/psisrc/lr70.psi $bin $i 2.20s user 0.04s system 99% cpu 2.237 total ../../testcode/psisrc/lr80.psi $bin $i 2.49s user 0.05s system 99% cpu 2.544 total ../../testcode/psisrc/lr90.psi $bin $i 2.87s user 0.03s system 99% cpu 2.905 total

missed 70 for ct, running now.

table 1 per model time is the time for hk-maple right?

yes

Plotting psi times…

Sorry, I ran 70 but forgot to add time. Doing it again

~/w/h/r/psi > time ../../other/psi/psi ../../testcode/psisrc/ct70.psi
452.29user 1.32system 7:34.23elapsed 99%CPU (0avgtext+0avgdata 3279712maxresident)k

I pushed the lda numbers for table1. Is there anything else, I am going to log out soon.

Thanks! I take it you’re not up for running MALLET LDA :wink:

I forgot how I did it last time.

It would be totally new because it’s LDA, not NB.

http://mallet.cs.umass.edu/topics.php

importing data in mallet is completely different, for 20newsgroup we just gave it raw data. For KOS I don’t know how to do that.

Also 20newsgroup their website itself had examples.

Hey, sorry I didn’t notice, it seems LR is much faster than CT with PSI?

Yeah

It doesn’t scale that bad

Do you think you can run larger n, like 100,200 up to 1000?

running now.

Thanks and bon voyage…

Thank you!

Oh would you please report your PSI and D versions?

(for PSI, commit hash)

I think I did already

it’s in other/psi along with augur

psi lr ../../testcode/psisrc/lr100.psi $bin $i 3.29s user 0.03s system 99% cpu 3.323 total ../../testcode/psisrc/lr10.psi $bin $i 0.40s user 0.01s system 99% cpu 0.409 total ../../testcode/psisrc/lr200.psi $bin $i 8.80s user 0.13s system 99% cpu 8.938 total ../../testcode/psisrc/lr20.psi $bin $i 0.63s user 0.00s system 99% cpu 0.635 total ../../testcode/psisrc/lr300.psi $bin $i 20.02s user 0.20s system 99% cpu 20.244 total ../../testcode/psisrc/lr30.psi $bin $i 0.93s user 0.01s system 99% cpu 0.938 total ../../testcode/psisrc/lr400.psi $bin $i 41.24s user 0.39s system 99% cpu 41.676 total ../../testcode/psisrc/lr40.psi $bin $i 1.26s user 0.01s system 99% cpu 1.273 total ../../testcode/psisrc/lr500.psi $bin $i 79.87s user 0.51s system 99% cpu 1:20.66 total ../../testcode/psisrc/lr50.psi $bin $i 1.57s user 0.02s system 99% cpu 1.592 total ../../testcode/psisrc/lr600.psi $bin $i 136.46s user 0.85s system 99% cpu 2:17.45 total ../../testcode/psisrc/lr60.psi $bin $i 1.79s user 0.05s system 99% cpu 1.841 total ../../testcode/psisrc/lr700.psi $bin $i 220.28s user 1.04s system 99% cpu 3:41.78 total ../../testcode/psisrc/lr70.psi $bin $i 2.14s user 0.05s system 99% cpu 2.186 total ../../testcode/psisrc/lr800.psi $bin $i 347.42s user 2.00s system 99% cpu 5:49.86 total ../../testcode/psisrc/lr80.psi $bin $i 2.43s user 0.04s system 99% cpu 2.485 total ../../testcode/psisrc/lr900.psi $bin $i 498.65s user 2.39s system 99% cpu 8:21.68 total ../../testcode/psisrc/lr90.psi $bin $i 2.80s user 0.06s system 99% cpu 2.860 total

logging off my machine, I still have slack on my phone if something comes up.

Cool, thanks!

@ccshan when did you become an economist?

https://hsm.stackexchange.com/questions/140/why-is-price-on-the-vertical-axis-and-quantity-on-the-horizontal-axis

I’m happy to switch it back. I thought it’s PLDI-ish advice to plot things so that “up is better”.

well, I think when I’ve said things like that I’ve mostly meant that you want the graph to go a particular way to indicate success

I think the dependent variable should always be on the y axis