Vector indexing vs. hashing (envir)

[trinker]

Initial testing indicates that a vector lookup may be a better choice than hashing for items such as syllable lookup:

syn <- list2df(as.list(env.syl))[, 1]
names(syn)  <- list2df(as.list(env.syl))[, 2]

dat <- bag_o_words(pres_debate_raw2012$dialogue)

vect=syn[dat]
hash=hash_look(dat, env.syl)

Unit: milliseconds
 expr      min       lq   median       uq      max neval
 vect 35.46802 35.54943 35.66862 35.74349 36.67534   100
 hash 45.01694 45.22314 45.39621 45.70317 49.95538   10

[Dasonk]

It might depend on the size of the lookup table. You might want to vary the lookup table size when doing your benchmarking.

[trinker]

I'll do some more investigating on varying sized vectors. I may use an ifelse and handle based on these findings.

[trinker]

Good call Dason. Here's a larger vector:

syn <- list2df(as.list(env.syl))[, 1]
names(syn)  <- list2df(as.list(env.syl))[, 2]

dat <- bag_o_words(raj$dialogue)

vect=syn[dat]
hash=hash_look(dat, env.syl) 

Unit: milliseconds
 expr      min       lq   median       uq      max neval
 vect 230.3721 230.6975 231.3035 232.3678 251.2343    20
 hash 137.1972 143.8174 145.0490 150.2753 197.6331    20

Here are the lengths for the 2 vectors:

sapply(list(
    bag_o_words(pres_debate_raw2012$dialogue), 
    bag_o_words(raj$dialogue)), length)
[1]  8309 23783

So 8,309 vs 23,783 or 4 times the size of the first one.

[trinker]

Ok so using some replication I get that there's a difference in performance after about 15000 words:

Code

library (pacman)
p_load(embodied, matlab, dplyr)
dat <- bag_o_words(raj$dialogue)

n <- c(seq(7000, 20000, by=1000), length(dat))
n <- rep(n, each=2)

pars <- c(replicate(length(n)/2, sample(c(TRUE, FALSE), 2)))

FUN <- function() {
    comps <- lapply(1:length(n), function(i) {
        cat(paste("\n=============\n", i, " in ", length(n), "\n=============\n\n"))
        tic(gcFirst=FALSE) 

        if (pars[i]) {
            syn[dat[1:c(n[i])]]
        } else {
            hash_look(dat[1:c(n[i])], env.syl) 
        }

        as.numeric(gsub("[a-z]|\\s+", "", capture.output(toc(echo=TRUE) )))
    })

    data.frame(Approach = c("hash", "vector")[as.numeric(pars) + 1], 
        n = n, time = unlist(comps))
}

out <- lapply(1:100, function(i) FUN())

dat2 <- do.call(rbind, out)

dat3 <- dat2 %.%
    group_by(Approach, n) %.%
    summarise(ave=mean(time)) %.%
    arrange(n, Approach)
dat3

ggplot(dat2, aes(x=n, y=time, group=Approach, colour=Approach)) + 
     geom_point(size=2, alpha=.2, shape=19) + 
     geom_jitter(position = position_jitter(width=200, height = 0)) +
     geom_line(data=dat3, aes(x=n, y=ave, group=Approach, 
        colour=Approach), size=1.5) +
    ggtitle("Vector vs. Hash Approaches to lookup") +
    xlab("Number of Words") + ylab("Time (sec)")

ggsave("output.png", height=5, width=7)