Adding new cpp functions

[Nowosad]

Dear @HajkD,

I am a regular user of philentropy as I find it very useful.
The package is usually very fast, but I have some cases when the current speed is not enough. Therefore, I added some new Rcpp functions:

1. single_distance - to calculate a distance between two vectors without the need to rbind them first
2. dist_one_to_many() - to calculate distances between one vector and many vectors (in the form of a matrix) of other values
3. dist_many_to_many() - to calculate distances between many vectors (a matrix) and many vectors (a matrix)

There are two main goals of the changes:

1. To allow easier (more customizable) selection of distances (single_distance)
2. To speed up one-to-many and many-to-many calculations

You can see an early draft of my proposed changes at #27. If you like the overall idea, then I will implement the rest of the metrics in a similar fashion.

I have also benchmarked the new functions on a simple problem of comparing many distances.

f1() uses the existing distance() function - it is slower than the rest, but it allows to select distance measure easily. f2() uses a hard-coded distance measure - it is fast, but hard to customize. f3() is slightly slower than f2(), but much faster than f1() and allows to select distance measure. f4() is much faster than the rest of the functions and it is easy to customize.

Let me know what you think about this idea.

All the best,
Jakub

[Nowosad]

I have updated the benchmark. f5() uses dist_many_many() function - it could be related to #22.

devtools::load_all("~/Software/philentropy/")
#> ℹ Loading philentropy
set.seed(2021-07-14)
m = matrix(sample(1:10, size = 1000, replace = TRUE), ncol = 10)

f1 = function(){
  r = vector(length = nrow(m))
  for (i in seq_len(nrow(m))){
    dist_sum = 0
    dist_count = 0
    for (j in seq_len(nrow(m))){
      x = rbind(m[i, ], m[j, ])
      dist_sum = dist_sum + distance(x, method = "euclidean", mute.message = TRUE)
      dist_count = dist_count + 1
    }
    r[i] = dist_sum / dist_count
  }
  r
}

f2 = function(){
  r = vector(length = nrow(m))
  for (i in seq_len(nrow(m))){
    dist_sum = 0
    dist_count = 0
    for (j in seq_len(nrow(m))){
      dist_sum = dist_sum + euclidean(m[i, ], m[j, ], FALSE)
      dist_count = dist_count + 1
    }
    r[i] = dist_sum / dist_count
  }
  r
}

f3 = function(){
  r = vector(length = nrow(m))
  for (i in seq_len(nrow(m))){
    dist_sum = 0
    dist_count = 0
    for (j in seq_len(nrow(m))){
      dist_sum = dist_sum + single_distance(m[i, ], m[j, ], "euclidean", FALSE, "")
      dist_count = dist_count + 1
    }
    r[i] = dist_sum / dist_count
  }
  r
}

f4 = function(){
  r = vector(length = nrow(m))
  for (i in seq_len(nrow(m))){
    dists = dist_one_many(m[i, ], m, "euclidean", FALSE, "")
    r[i] = mean(dists)
  }
  r
}

f5 = function(){
  dists = dist_many_many(m, m, "euclidean", FALSE, "")
  r = rowMeans(dists)
  r
}

bench::mark(f1(), f2(), f3(), f4(), f5())
#> Warning: Some expressions had a GC in every iteration; so filtering is disabled.
#> # A tibble: 5 x 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 f1()       297.39ms 306.42ms      3.26   39.27MB     8.16
#> 2 f2()        57.12ms  67.73ms     14.5    24.39MB    23.6 
#> 3 f3()        74.83ms  78.22ms     12.2    24.39MB    19.2 
#> 4 f4()         9.46ms   9.98ms     93.1     1.15MB    11.9 
#> 5 f5()         6.22ms   6.81ms    138.    121.14KB     9.98

^{Created on 2021-07-14 by the reprex package (v2.0.0)}

[HajkD]

Hi @Nowosad

I am very happy with the proposed changes and will now merge them into the master branch. Your help with improving philentropy is greatly appreciated!

Many thanks,
Hajk