Things are slower with "dplyr" is there a faster way? - r

Question

I m just trying to calculate the relative angle between with my x,y,z data frame to the reference vector. So far, I use dplyr to group things and apply my angle function to get relative angle. However things are quite slow even for dummy data that I provide here.
set.seed(12345)
x <- replicate(1,c(replicate(1000,rnorm(50,0,0.01))))
y <- replicate(1,c(replicate(1000,rnorm(50,0,0.01))))
z <- replicate(1,c(replicate(1000,rnorm(50,0.9,0.01))))
ref_vector <- data.frame(ref_x=rep(0,100),ref_y=rep(0,100),ref_z=rep(1,100))
set <- rep(seq(1,1000),each=50)
data_rep <- data.frame(x,y,z,ref_vector,set)
>
head(data_rep)
# x y z ref_x ref_y ref_z set
# 1 0.005855288 -0.015472796 0.9059337 0 0 1 1
# 2 0.007094660 -0.013354359 0.9040137 0 0 1 1
# 3 -0.001093033 -0.014661486 0.9047502 0 0 1 1
# 4 -0.004534972 -0.002764655 0.9070553 0 0 1 1
# 5 0.006058875 -0.008339952 0.8926551 0 0 1 1
# 6 -0.018179560 -0.008412400 0.9055541 0 0 1 1
I define the angle between two vectors with this angle function,
angle <- function(x,y){
dot.prod <- x%*%y
norm.x <- norm(x,type="2")
norm.y <- norm(y,type="2")
theta <- acos(dot.prod / (norm.x * norm.y))
as.numeric(theta)
}
then lets apply this to our data_rep
library(dplyr)
system.time(df_angle <- data_rep%>%
rowwise()%>%
do(data.frame(.,angle_rad=angle(unlist(.[1:3]),unlist(.[4:6]))))%>%
group_by(set)%>%
mutate(angle=angle_rad*180/pi, mean_angle=mean(angle)))
# user system elapsed
# 64.22 0.08 64.81
# Warning message:
# Grouping rowwise data frame strips rowwise nature
As you can see, the process took around 1 min and I even did not provide all my real data set which has 350000 rows and it takes 10 min to calculate the relative angle.
I wonder is there any way to speed up this process.
Thanks!

Answer 1

Just discover linear algebra for yourself:
m1 = as.matrix(data_rep[, 1:3])
m2 = as.matrix(data_rep[, 4:6])
system.time( {
m1 = m1 / sqrt(rowSums(m1 ^ 2))
m2 = m2 / sqrt(rowSums(m2 ^ 2))
RESULT <- acos(rowSums(m1 * m2))
})
# user system elapsed
# 0.004 0.001 0.006
all.equal(df_angle$angle_rad, RESULT)
# TRUE

Answer 2

Just make a simple mutatestatement instead of your do(data.frame()) part. This improves the performance quite a bit, because you no longer have to convert each row into a data.frame
system.time(df_angle2 <- data_rep%>%
rowwise() %>%
mutate(angle_rad=angle(x = c(x,y,z),y = c(ref_x,ref_y,ref_z))) %>%
group_by(set)%>%
mutate(angle=angle_rad*180/pi, mean_angle=mean(angle)))
## user system elapsed
## 3.72 0.00 3.71
all.equal(df_angle,df_angle2)
## TRUE