Think about rowwise grouping

[hadley]

rowwise <- function(df) {
  n <- nrow(df)

  attr(df, "indices") <- as.list(seq_len(n) - 1) 
  attr(df, "drop") <- FALSE
  attr(df, "group_sizes") <- rep(1L, n)
  attr(df, "biggest_group_size") <- 1L
  attr(df, "labels") <- data.frame(row = seq_len(n))
  attr(df, "vars") <- list(quote(row)) # list(substitute(bootstrap(m)))
  class(df) <- c("grouped_df", "tbl_df", "tbl", "data.frame")

  df
}

# Not useful for vectorised functions. But can be useful for
# non vectorised
summarise(rowwise(mtcars), vs + am)

[hadley]

@romainfrancois any thoughts on implementing this? We could either abuse the existing grouping mechanism (as in the code example) or create a new class.

I think the main use case will be in conjunction with do(), so that you can work with data frames whose columns are lists containing (e.g.) linear models.

[romainfrancois]

Abusing the existing grouping mechanism is cheap, most functions will work (perhaps we would need to be careful about vars.

Conceptually, it might be interesting to structure the data differently so that internally we know there is only one index per group, but that might be quite an investment of time to implement versions that take advantage of that knowledge.

Perhaps at least we could do:

class(df) <- c("rowwise_df", "grouped_df", "tbl_df", "tbl", "data.frame")

so that we can adapt do.

[hadley]

I think we also need to specialise summarise for that case (so we can conceptually use [[ instead of [ internally), but probably everything else is the same. i.e. it should be possible to do:

models <- mtcars %.% group_by(cyl) %.% do(mod = lm(mpg ~ vs, data = .)
models %.% summarise(rsq = summary(mod)$r.squared)

currently you need

models %.% summarise(rsq = summary(mod[[1]])$r.squared)

[romainfrancois]

mod can only be expanded to mod[[1]] when we know we only have one row per group. So we need a way to communicate this information from rowwise and do.

do here would create a data frame with 3 rows right ? one for each value of cyl?

[hadley]

Right. We could either add a flag or just use the additional class, as you suggested.

[romainfrancois]

I've put some code in so that:

> models <- mtcars %.% group_by(cyl) %.% do(mod = lm(mpg ~ vs, data = .) )
>
> rowwise(models) %>% mutate( rsq = summary(mod)$r.squared )
Source: local data frame [3 x 3]
Groups: cyl

  cyl     mod         rsq
1   4 <S3:lm> 0.002381953
2   6 <S3:lm> 0.281055142
3   8 <S3:lm> 0.000000000
> rowwise(models) %>% summarise( rsq = summary(mod)$r.squared )
Source: local data frame [3 x 1]

          rsq
1 0.002381953
2 0.281055142
3 0.000000000

With:

> rowwise
function(data){
  structure( data, class = c("rowwise_df", "tbl_df", "data.frame") )
}
<environment: namespace:dplyr>

[romainfrancois]

rowwise should probably respect a bit more the groupings. Not sure what to do with the other verbs. At the moment, verb( rowwise(data) ) does verb( ungroup(data) ), which might not be ideal.

But most of the internal stuff knows about rowwise:

> mtcars %>% rowwise() %>% summarise( a = n() )
Source: local data frame [32 x 1]

   a
1  1
2  1
3  1
4  1
5  1
6  1
7  1
8  1
9  1
10 1
11 1
12 1
13 1
14 1
15 1
16 1
17 1
18 1
19 1
20 1
21 1
22 1
23 1
24 1
25 1
26 1
27 1
28 1
29 1
30 1
31 1
32 1

[hadley]

Thanks - this looks great! I'll fill in the missing pieces and then we should be good to submit to cran :)