我尝试创建stop_gradient
/ tf.gradients
的变体,该变体使用tf.assign
参数跨越任意一组列,使用{{1 } / pmax
。下面的函数的第一个版本对过滤器值进行硬编码,但适用于pmin
:
filter_value
返回:
.SD
下面的函数的第二个版本需要第二个.SDcols
参数,但是我无法使用.SD
,而是必须将各个列向量作为列表让事情发挥作用:
testFuncV1 <- function(...) {
cols <- list(...)
num_cols <- length(cols)
num_records <- length(cols[[1]])
max_records <- c()
for (record_num in 1:num_records) {
v <- c()
for (l in cols) {
v <- c(v, l[[record_num]])
}
filt_v <- Filter(function(x) { x <= 1 }, v)
if (length(filt_v) == 0) {
max_records <- c(max_records, NA)
} else {
max_records <- c(max_records, max(filt_v))
}
}
max_records
}
test_dt_v1 <- data.table(a = c(1,3,5), b = c(2,3,-1), c = c(-3, 5, 2))
test_dt_v1[, max_with_filter := do.call(testFuncV1, .SD), .SDcols = c('a', 'b', 'c')]
也会返回:
a b c max_with_filter
1: 1 2 -3 1
2: 3 3 5 NA
3: 5 -1 2 -1
理想情况下,我可以使用filter
找出适合.SD
的方法,或者使用testFuncV2 <- function(cols, filter) {
num_cols <- length(cols)
num_records <- length(cols[[1]])
max_records <- c()
for (record_num in 1:num_records) {
v <- c()
for (l in cols) {
v <- c(v, l[[record_num]])
}
filt_v <- Filter(function(x) { x <= filter }, v)
if (length(filt_v) == 0) {
max_records <- c(max_records, NA)
} else {
max_records <- c(max_records, max(filt_v))
}
}
max_records
}
test_dt_v2 <- data.table(a = c(1,3,5), b = c(2,3,-1), c = c(-3, 5, 2))
test_dt_v2[, max_with_filter := do.call(testFuncV2, list(list(test_dt_v2$a, test_dt_v2$b, test_dt_v2$c), 1))]
的替代方法(我也尝试过)周围,无济于事。提前谢谢!
答案 0 :(得分:1)
以下是使用apply(MARGIN=1, ...)
func <- function(x, threshold) {
if (any(x <= threshold)) return(max(x[x <= threshold]))
NA
}
test_dt_v1[, max_with_filter := apply(.SD, 1, func, threshold=1),
.SDcols=c("a","b","c")]
使用do.call
和pmax
的另一个选项,首先将值高于1转换为NA(想法来自rowwise maximum for R)
test_dt_v1[, max_with_filter := do.call(pmax, c(`is.na<-`(.SD, .SD>1), na.rm=T))]