Question

我在R中有一个使用deSolve包中的odesolver的微分方程模型。但是，目前模型运行速度非常慢。我认为这可能与我提供给odesolver编写得不好的函数有关，但无法弄清楚究竟是什么减慢了它以及我如何加速它。有没有人有任何想法？

我做了一个与我的工作方式类似的例子：

library(data.table)
library(deSolve)

matrix_1 <- matrix(runif(100),10,10)
matrix_1[which(matrix_1 > 0.5)] <- 1
matrix_1[which(matrix_1 < 0.5)] <- 0

matrix_2 <- matrix(runif(100),10,10)
matrix_2[which(matrix_2 > 0.5)] <- 1
matrix_2[which(matrix_2 < 0.5)] <- 0

group_ID <- rep(c(1,2), 5)
N <- runif(10, 0, 100000)

Nchange <- function(t, N, parameters) {
  with(as.list(c(N, parameters)), {
    N_per_1 <- matrix_1 * N_per_connection  
    N_per_1[is.na(N_per_1)] <- 0
    total_N_2 <- as.vector(N_per_1)
    if (nrow(as.matrix(N_per_1)) > 1) {
      total_N_2 <- colSums(N_per_1[drop = FALSE])
    }
    N_per_1_cost <- N_per_1
    for (i in possible_competition) {
      column <- as.vector(N_per_1[, i])
      if (sum(column) > 0) {
        active_groups <- unique(group_ID[column > 0])
        if (length(active_groups) > 1){
          group_ID_dets <- data.table("group_ID" = group_ID, "column"= column, "n_IDS" = 1:length(group_ID))
          group_ID_dets$portions <- ave(group_ID_dets$column, group_ID_dets$group_ID, FUN = function(x) x / sum(x))
          group_ID_dets[is.na(group_ID_dets)] <- 0
          totals <- as.vector(unlist(tapply(group_ID_dets$column, group_ID_dets$group_ID, function(x) sum(x))))
          totals[is.na(totals)] <- 0
          totals <- totals*2 - sum(totals)
          totals[totals < 0] <- 0
          group_ID_totals <- data.table("group_ID" = unique(group_ID), "totals" = as.vector(totals))
          group_ID_dets$totals <- group_ID_totals$totals[match(group_ID_dets$group_ID, group_ID_totals$group_ID)]
          N_per_1[, i] <- group_ID_dets$totals * group_ID_dets$portions
        }
      }
    }

    res_per_1 <- N_per_1 * 0.1

    N_per_2 <- matrix_2 * N_per_connection
    N_per_2[is.na(N_per_2)] <- 0 
    res_per_2 <- N_per_2 * 0.1

    dN <- rowSums(res_per_1)  - rowSums(N_per_1_cost * 0.00003) + rowSums(res_per_2)  - 
      rowSums(N_per_2 * 0.00003) - N*0.03

    list(c(dN))
  })
}  # function describing differential equations
N_per_connection <- N/(rowSums(matrix_1) + rowSums(matrix_2))
possible_competition <- which(colSums(matrix_1 != 0)>1)
times <- seq(0, 100, by = 1) 
out <- ode(y = N, times = times, func = Nchange, parms = NULL)

Answer 1

识别瓶颈的一个好方法是使用profiler，profvis包提供了深入研究结果的好方法。将代码包装在p <- profvis({YourCodeInHere})中，然后使用print(p)查看结果，可获得以下见解：

占用时间最多的行是（按时间降序排列）：

group_ID_totals <- data.table("group_ID" = unique(group_ID), "totals" = as.vector(totals))

group_ID_dets$portions <- ave(group_ID_dets$column, group_ID_dets$group_ID, FUN = function(x) x / sum(x))

group_ID_dets <- data.table("group_ID" = group_ID, "column"= column, "n_IDS" = 1:length(group_ID))

totals <- as.vector(unlist(tapply(group_ID_dets$column, group_ID_dets$group_ID, function(x) sum(x))))

group_ID_dets$totals <- group_ID_totals$totals[match(group_ID_dets$group_ID, group_ID_totals$group_ID)]

我不熟悉您的ODE的详细信息，但您应该专注于优化这些任务。我认为更大的问题是你在循环中运行这些命令。通常，您会听到R中的循环速度很慢，但在答案here中可以找到关于此问题的更细致的讨论。这里的一些提示可能会帮助您重构代码/循环。祝你好运！

在R中加速odesolver的问题

1 个答案: