Question

我尝试并行化ape::dist_topo()，这是一个计算无根树之间距离的函数。

通常函数的工作方式如下（代表：4个随机树，每个树有5片叶子）：

library(tidyverse)
# devtools::install_github("hadley/multidplyr")
library(multidplyr)
library(ape)
set.seed(3)

trees <- 
  map(rep(5, 4), rtree) %>% 
  do.call(c.phylo, .) %>% # To transform my list of phylo objects in a multiPhylo object
  unroot.multiPhylo()

dist.topo(trees)
#      tree1 tree2 tree3
# tree2     4            
# tree3     4     2      
# tree4     4     4     2

我创建了一个函数来计算data.frame中的距离2乘2（为了按行分割成簇）：

dist.topo2 <- function(multiphylo){
  expand.grid(multiphylo, multiphylo) %>% 
    as.tibble() %>% 
    mutate(dist = map2(Var1, Var2, dist.topo)) %>% 
    pull(dist) %>% 
    matrix(., nrow = sqrt(length(.))) %>% 
    as.dist()
}

dist.topo2(trees)
#   1 2 3
# 2 4    
# 3 4 2  
# 4 4 4 2

正如所料，结果是相同的（无论名称如何）。

然后我在我的管道中添加了multidplyr::partition()和multidplyr::collect()函数：

dist.topo3 <- function(multiphylo){
  expand.grid(multiphylo, multiphylo) %>% 
    as.tibble() %>% 
    partition() %>%
    mutate(dist = purrr::map2(Var1, Var2, ape::dist.topo)) %>% 
    collect() %>%
    pull(dist) %>% 
    matrix(., nrow = sqrt(length(.))) %>% 
    as.dist()
}

dist.topo3(trees)
#   1 2 3
# 2 4    
# 3 0 4  
# 4 2 4 4
# Warning messages:
# 1: In bind_rows_(x, .id) :
#   Vectorizing 'multiPhylo' elements may not preserve their attributes
# 2: In bind_rows_(x, .id) :
#   Vectorizing 'multiPhylo' elements may not preserve their attributes
# 3: In bind_rows_(x, .id) :
#   Vectorizing 'multiPhylo' elements may not preserve their attributes
# 4: In bind_rows_(x, .id) :
#   Vectorizing 'multiPhylo' elements may not preserve their attributes
# 5: In bind_rows_(x, .id) :
#   Vectorizing 'multiPhylo' elements may not preserve their attributes
# 6: In bind_rows_(x, .id) :
#   Vectorizing 'multiPhylo' elements may not preserve their attributes

正如您所看到的，距离是不同的，而操作并没有改变。

我该如何解决？也许这是不可能的（See here）

由于

注意：我知道这个解决方案可能不是最优的（特别是因为它计算每个距离两次），但它不是重点。

Answer 1

我宁愿使用包{furrr}而不是{multidplyr}：

library(furrr)
plan(multiprocess)

dist.topo4 <- function(multiphylo) {

  dists <- expand.grid(multiphylo, multiphylo) %>% 
    setNames(c("x", "y")) %>%
    future_pmap_dbl(ape::dist.topo)

  n <- length(multiphylo)
  dim(dists) <- c(n, n)

  as.dist(dists)
}

结果

> dist.topo4(trees)
  1 2 3
2 4    
3 4 2  
4 4 4 2

Answer 2

问题在于partition会随机对data.frame进行分片，而collect会随机取消data.frame。如果您将行号添加为列并在收集后进行排列，则会解决问题

dist.topo3 <- function(multiphylo){
  expand.grid(multiphylo, multiphylo) %>% 
    as.tibble() %>% 
    mutate(rn = row_number()) %>%
    partition(rn) %>%
    mutate(dist = purrr::map2(Var1, Var2, ape::dist.topo)) %>% 
    collect() %>%
    arrange(rn) %>%
    pull(dist) %>% 
    matrix(., nrow = sqrt(length(.))) %>% 
    as.dist()
}
dist.topo3(trees)
#   1 2 3
# 2 4    
# 3 4 2  
# 4 4 4 2

使用multidplyr进行矢量化不会呈现正确的输出

2 个答案:

结果