Performance Benchmark • embedplyr

library(embedplyr)
library(microbenchmark)
library(ggplot2)

Whereas indexing a regular matrix by rownames gets slower as the number of rows increases, embedplyr’s hash table indexing means that token embeddings can be retrieved in roughly constant time. The following test uses glove.twitter.27B.25d, which has 1.2 million rows (by no means the largest model available).

glove_embeddings <- load_embeddings("glove.twitter.27B.25d", save = FALSE)
#> Reading file from the Internet...
glove_matrix <- as.matrix(glove_embeddings)

# 50 randomly sampled tokens
tokens <- rownames(glove_embeddings)[sample(seq_len(nrow(glove_embeddings)), 50)]

bm1 <- microbenchmark(
    "base indexing (50 tokens)" = glove_matrix[tokens,],
    "embedplyr (50 tokens)" = glove_embeddings[tokens,],
    times = 500
)
autoplot(bm1)

When retrieving many token embeddings (e.g. when analyzing a set of texts), this can make a noticeable difference.