Investigate using vctrs for coercion

[DavisVaughan]

Coercion of new_data columns into their correct type (maybe characters can be coerced to factors automatically)

[DavisVaughan]

suppressPackageStartupMessages({
  library(hardhat)
  library(vctrs)
  library(gapminder)
  library(rsample)
  library(dplyr)
})
#> Warning: package 'rsample' was built under R version 3.5.2
#> Warning: package 'dplyr' was built under R version 3.5.2

options(rlang__backtrace_on_error = "none")

split <- initial_split(gapminder)
gap_train <- training(split)
gap_test <- testing(split)

# 0 row slice (this is `info`!)
gap_train_0 <- vec_slice(gap_train, 0L)

# all of the df - df casts are due to df_col_cast()
# which is called from vec_cast()

# ///////////////////////

# continent is character not factor
gap_test2 <- mutate(gap_test, continent = as.character(continent))

# silently fixes that
common2 <- vec_cast(gap_test2, gap_train_0)

# recovered levels and type
levels(common2$continent)
#> [1] "Africa"   "Americas" "Asia"     "Europe"   "Oceania"

# Takeaway) This is good

# ///////////////////////

# continent is character not factor, AND has too many implicit levels internally
gap_test2b <- mutate(gap_test, continent = as.character(continent))
gap_test2b$continent[1] <- "jupyter"

# noisily fixes that
common2b <- vec_cast(gap_test2b, gap_train_0)
#> Warning: Lossy cast from <character> to <factor<69262>>
#> Locations: 1

# recovered levels and type
levels(common2b$continent)
#> [1] "Africa"   "Americas" "Asia"     "Europe"   "Oceania"

# Takeaway) This is good, but a better warning would be
# good. This is a TOO MANY LEVELS problem.

# ///////////////////////

# continent is character not factor, AND doesn't have enough levels internally
gap_test2c <- mutate(gap_test, continent = as.character(continent))
gap_test2c$continent <- gsub("Asia", NA_character_, gap_test2c$continent)

# silently fixes that
common2c <- vec_cast(gap_test2c, gap_train_0)

# recovered levels and type
levels(common2c$continent)
#> [1] "Africa"   "Americas" "Asia"     "Europe"   "Oceania"

# Takeaway) This is good

# ///////////////////////

# continent is numeric
gap_test3 <- mutate(gap_test, continent = 1)

# not the best error message
common3 <- vec_cast(gap_test3, gap_train_0)
#> Error: Can't cast <double> to <factor<69262>>

# Takeaway) This is good, but want a better error message
# would really like this better message to be vctrs' job

# ///////////////////////

# too many columns
gap_test4 <- mutate(gap_test, x = 4)

# this actually throws a decent error message,
# but I think we would still rather have our own
vec_cast(gap_test4, gap_train)
#> Warning: Lossy cast from <tbl_df<
#>   country  : factor<bf6dc>
#>   continent: factor<69262>
#>   year     : integer
#>   lifeExp  : double
#>   pop      : integer
#>   gdpPercap: double
#>   x        : double
#> >> to <tbl_df<
#>   country  : factor<bf6dc>
#>   continent: factor<69262>
#>   year     : integer
#>   lifeExp  : double
#>   pop      : integer
#>   gdpPercap: double
#> >>
#> Locations: 
#> Dropped variables: `x`
#> # A tibble: 426 x 6
#>    country     continent  year lifeExp      pop gdpPercap
#>    <fct>       <fct>     <int>   <dbl>    <int>     <dbl>
#>  1 Afghanistan Asia       1952    28.8  8425333      779.
#>  2 Afghanistan Asia       1987    40.8 13867957      852.
#>  3 Albania     Europe     1987    72    3075321     3739.
#>  4 Albania     Europe     1992    71.6  3326498     2497.
#>  5 Albania     Europe     2002    75.7  3508512     4604.
#>  6 Algeria     Africa     1957    45.7 10270856     3014.
#>  7 Algeria     Africa     1977    58.0 17152804     4910.
#>  8 Algeria     Africa     1987    65.8 23254956     5681.
#>  9 Algeria     Africa     2007    72.3 33333216     6223.
#> 10 Angola      Africa     1962    34    4826015     4269.
#> # … with 416 more rows

# Takeaway) This is good, but probably want a custom error message
# because vctrs "types" its warnings, I think I can do this

# ///////////////////////

# not enough columns
gap_test5 <- select(gap_test, -pop)

# silently adds column of NA values
vec_cast(gap_test5, gap_train_0)
#> # A tibble: 426 x 6
#>    country     continent  year lifeExp   pop gdpPercap
#>    <fct>       <fct>     <int>   <dbl> <int>     <dbl>
#>  1 Afghanistan Asia       1952    28.8    NA      779.
#>  2 Afghanistan Asia       1987    40.8    NA      852.
#>  3 Albania     Europe     1987    72      NA     3739.
#>  4 Albania     Europe     1992    71.6    NA     2497.
#>  5 Albania     Europe     2002    75.7    NA     4604.
#>  6 Algeria     Africa     1957    45.7    NA     3014.
#>  7 Algeria     Africa     1977    58.0    NA     4910.
#>  8 Algeria     Africa     1987    65.8    NA     5681.
#>  9 Algeria     Africa     2007    72.3    NA     6223.
#> 10 Angola      Africa     1962    34      NA     4269.
#> # … with 416 more rows

# Takeaway) We should let shrink() be noisy here and error

# ///////////////////////

# Too many levels, but the factor's actual values never use that level
gap_test6 <- mutate(gap_test, continent = factor(continent, c(levels(continent), "extra_level")))

# silently fixes that
common6 <- vec_cast(gap_test6, gap_train_0)

levels(common6$continent)
#> [1] "Africa"   "Americas" "Asia"     "Europe"   "Oceania"

# Takeaway) Silence is okay here as the values aren't actually affected

# ///////////////////////

# Too many levels (in test), AND the factor's actual values use that level
# (here we drop Africa from the train data to demonstrate)
gap_train_0_6b <- mutate(gap_train_0, continent = factor(continent, levels(continent)[-1]))

# noisily drops the level and coerces problematic
# positions to NA
common6b <- vec_cast(gap_test, gap_train_0_6b)
#> Warning: Lossy cast from <factor<69262>> to <factor<e5252>>
#> Locations: 6, 7, 8, 9, 10, 11, 12, 13, 33, 34, 35, 42, 43, 47, 48, 49, 5...

# no africa
levels(common6b$continent)
#> [1] "Americas" "Asia"     "Europe"   "Oceania"

# Takeaway) Noisy is good, but I think I want a different warning.
# Again, capture the typed warning. This could check if `x` is a factor
# then you'd know the lossy cast is specific to having too many factor levels

# ///////////////////////

# not enough levels (in test)
gap_test7 <- mutate(gap_test, continent = factor(continent, levels(continent)[-1]))

# silently fixes that
common7 <- vec_cast(gap_test7, gap_train_0)

levels(common7$continent)
#> [1] "Africa"   "Americas" "Asia"     "Europe"   "Oceania"

^{Created on 2019-02-25 by the reprex package (v0.2.1.9000)}

[DavisVaughan]

Closed by #71

[github-actions]

This issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.