1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
|
% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/scream.R
\name{scream}
\alias{scream}
\title{\if{html}{\Sexpr[stage=render,results=rd]{"\U0001f631"}} Scream.}
\usage{
scream(data, ptype, allow_novel_levels = FALSE)
}
\arguments{
\item{data}{A data frame containing the new data to check the structure
of.}
\item{ptype}{A data frame prototype to cast \code{data} to. This is commonly
a 0-row slice of the training set.}
\item{allow_novel_levels}{Should novel factor levels in \code{data} be allowed?
The safest approach is the default, which throws a warning when novel levels
are found, and coerces them to \code{NA} values. Setting this argument to \code{TRUE}
will ignore all novel levels. This argument does not apply to ordered
factors. Novel levels are not allowed in ordered factors because the
level ordering is a critical part of the type.}
}
\value{
A tibble containing the required columns after any required structural
modifications have been made.
}
\description{
\code{scream()} ensures that the structure of \code{data} is the same as
prototype, \code{ptype}. Under the hood, \code{\link[vctrs:vec_cast]{vctrs::vec_cast()}} is used, which
casts each column of \code{data} to the same type as the corresponding
column in \code{ptype}.
This casting enforces a number of important structural checks,
including but not limited to:
\itemize{
\item \emph{Data Classes} - Checks that the class of each column in \code{data} is the
same as the corresponding column in \code{ptype}.
\item \emph{Novel Levels} - Checks that the factor columns in \code{data} don't have any
\emph{new} levels when compared with the \code{ptype} columns. If there are new
levels, a warning is issued and they are coerced to \code{NA}. This check is
optional, and can be turned off with \code{allow_novel_levels = TRUE}.
\item \emph{Level Recovery} - Checks that the factor columns in \code{data} aren't
missing any factor levels when compared with the \code{ptype} columns. If
there are missing levels, then they are restored.
}
}
\details{
\code{scream()} is called by \code{\link[=forge]{forge()}} after \code{\link[=shrink]{shrink()}} but before the
actual processing is done. Generally, you don't need to call \code{scream()}
directly, as \code{forge()} will do it for you.
If \code{scream()} is used as a standalone function, it is good practice to call
\code{\link[=shrink]{shrink()}} right before it as there are no checks in \code{scream()} that ensure
that all of the required column names actually exist in \code{data}. Those
checks exist in \code{shrink()}.
}
\section{Factor Levels}{
\code{scream()} tries to be helpful by recovering missing factor levels and
warning about novel levels. The following graphic outlines how \code{scream()}
handles factor levels when coercing \emph{from} a column in \code{data} \emph{to} a
column in \code{ptype}.
\figure{factor-handling.png}
Note that ordered factor handing is much stricter than factor handling.
Ordered factors in \code{data} must have \emph{exactly} the same levels as ordered
factors in \code{ptype}.
}
\examples{
# ---------------------------------------------------------------------------
# Setup
train <- iris[1:100, ]
test <- iris[101:150, ]
# mold() is run at model fit time
# and a formula preprocessing blueprint is recorded
x <- mold(log(Sepal.Width) ~ Species, train)
# Inside the result of mold() are the prototype tibbles
# for the predictors and the outcomes
ptype_pred <- x$blueprint$ptypes$predictors
ptype_out <- x$blueprint$ptypes$outcomes
# ---------------------------------------------------------------------------
# shrink() / scream()
# Pass the test data, along with a prototype, to
# shrink() to extract the prototype columns
test_shrunk <- shrink(test, ptype_pred)
# Now pass that to scream() to perform validation checks
# If no warnings / errors are thrown, the checks were
# successful!
scream(test_shrunk, ptype_pred)
# ---------------------------------------------------------------------------
# Outcomes
# To also extract the outcomes, use the outcome prototype
test_outcome <- shrink(test, ptype_out)
scream(test_outcome, ptype_out)
# ---------------------------------------------------------------------------
# Casting
# scream() uses vctrs::vec_cast() to intelligently convert
# new data to the prototype automatically. This means
# it can automatically perform certain conversions, like
# coercing character columns to factors.
test2 <- test
test2$Species <- as.character(test2$Species)
test2_shrunk <- shrink(test2, ptype_pred)
scream(test2_shrunk, ptype_pred)
# It can also recover missing factor levels.
# For example, it is plausible that the test data only had the
# "virginica" level
test3 <- test
test3$Species <- factor(test3$Species, levels = "virginica")
test3_shrunk <- shrink(test3, ptype_pred)
test3_fixed <- scream(test3_shrunk, ptype_pred)
# scream() recovered the missing levels
levels(test3_fixed$Species)
# ---------------------------------------------------------------------------
# Novel levels
# When novel levels with any data are present in `data`, the default
# is to coerce them to `NA` values with a warning.
test4 <- test
test4$Species <- as.character(test4$Species)
test4$Species[1] <- "new_level"
test4$Species <- factor(
test4$Species,
levels = c(levels(test$Species), "new_level")
)
test4 <- shrink(test4, ptype_pred)
# Warning is thrown
test4_removed <- scream(test4, ptype_pred)
# Novel level is removed
levels(test4_removed$Species)
# No warning is thrown
test4_kept <- scream(test4, ptype_pred, allow_novel_levels = TRUE)
# Novel level is kept
levels(test4_kept$Species)
}
|
