Rbind list of vectors with differing lengths

问题

I am new to R and I am trying to build a frequency/severity simulation. Everything is working fine except that it takes about 10min to do 10000 simulations for each of 700 locations. For the simulation of one individual location, I got a list of vectors with varying lengths and I would like to efficiently rbind these vectors, filling in NAs for all non-existing values. I would like R to return a data.frame to me. So far, I used rbind.fill.matrix after converting the vectors in the list to matrices of 1 row. However, I am hoping that I could use something like bind_rows (dplyr) or rbindfill but I don't know how to transform the vectors into something that I could use for these functions. Thank you in advance for your help!

set.seed(1223)

library(data.table)

numsim = 10

rN.D <- function(numsim) rpois(numsim, 4) 
rX.D <- function(numsim) rnorm(numsim, mean = 5, sd = 4)

freqs <- rN.D(numsim)
obs <- lapply(freqs, function(x) rX.D(x))
#obs is the list that I would like to rbind (efficiently!) and have a data.frame returned to me

回答1:

We can append NAs at the end to make the length same for each of the list elements and then do the rbind

out <- do.call(rbind, lapply(obs, `length<-`, max(lengths(obs))))
as.data.frame(out) # if we need a data.frame as output

---

or using tidyverse

library(tidyverse)
obs %>%
   set_names(seq_along(.)) %>% 
   stack %>% 
   group_by(ind) %>% 
   mutate(Col = paste0("Col", row_number())) %>% 
   spread(Col, values)

回答2:

Everything is working fine except that it takes [too long] to do [numsim] simulations

If your real application uses rnorm or similar, you can make a single call to it:

set.seed(1223)
numsim = 3e5
freqs = rN.D(numsim)
maxlen = max(freqs)
m = matrix(, maxlen, numsim)
m[row(m) <= freqs[col(m)]] <- rX.D(sum(freqs))

res = as.data.table(t(m))

I am filling the data the "wrong way" (with each simulation on a column instead of a row) and then transposing since R fills matrix values using "column-major" order.

---

If you need to use lapply, here's a benchmark for the final step:

set.seed(1223)

library(dplyr); library(tidyr); library(purrr)
library(data.table)

numsim = 3e5

rN.D <- function(numsim) rpois(numsim, 4) 
rX.D <- function(numsim) rnorm(numsim, mean = 5, sd = 4)

freqs <- rN.D(numsim)
obs <- lapply(freqs, function(x) rX.D(x))

system.time({
tidyres = obs %>%
   set_names(seq_along(.)) %>% 
   stack %>% 
   group_by(ind) %>% 
   mutate(Col = paste0("Col", row_number())) %>% 
   spread(Col, values)
})
#    user  system elapsed 
#   16.56    0.31   16.88     

system.time({
    out <- do.call(rbind, lapply(obs, `length<-`, max(lengths(obs))))
    bres = as.data.frame(out)
})
#    user  system elapsed 
#    0.50    0.05    0.55 

system.time(
    dtres <- setDT(transpose(obs))
)
#    user  system elapsed 
#    0.03    0.01    0.05 

The last approach is fastest compared to the other two (both from @akrun's answer).

Comment. I would recommend using only data.table or tidyverse. Mixing and matching will get messy very quickly. When I was setting this example up, I saw that purrr has it's own transpose function, so if you loaded packages in a different order, code like this can give different results without warning.

来源：https://stackoverflow.com/questions/51486317/rbind-list-of-vectors-with-differing-lengths