In drsimonj/pipelearner: Tidy machine learning pipelines
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>",
fig.path = "README-"
)
pipelearner makes it easy to create machine learning pipelines in R.
Installation and background
pipelearner is currently available from github as a development package only. It can be installed by running:
# install.packages("devtools")
devtools::install_github("drsimonj/pipelearner")
pipelearner is built on top of tidyverse packages like modelr. To harness the full power of pipelearner, it will help to possess some technical knowledge of tidyverse tools such as:
%>% the pipe operator from magrittr package.- tibbles from tibble package.
map() and other iteration functions from purrr package.resample objects from modelr package.
An excellent resource to get started with these is R for Data Science, by Garrett Grolemund and Hadley Wickham.
API
Similar to the way ggplot2 elements are layered with +, you initialize and customize a pipelearner object, which is a list, with functions that can be piped into eachother with %>%. Rather than plotting, however, a pipelearner then learns.
Initialize a pipelearner object with:
pipelearner()
Customize a pipelearner with:
learn_cvpairs() to customize the cross-validation pairs.learn_curves() to customize the learning curves using incremental proportions of training data.learn_models() to add new learning models.
Learn (fit) everything and obtain a tibble of results with:
learn()
Initialization
The following initializes a pipelearner object that will use the iris data set and linear regression (lm) to learn how to predict Sepal.Length with all other available variables (Sepal.Length ~ .).
library(pipelearner)
pl <- pipelearner(iris, lm, Sepal.Length ~ .)
Print a pipelearner object to expose the list elements.
pl
Defaults to note
data is split into a single cross-validation pair of resample objects (under cv_pairs) referencing 80% of the data for training and 20% for testing.- Learning is done on the entire proportion of the training data (
train_ps == 1).
Learning
Once a pipelearner is setup, use learn() to fit all models to every combination of training proportions (train_ps) and set of training data in the cross-validation pairs (cv_pairs), and return a tibble of the results.
pl %>% learn()
Quick notes
fit contains the fitted models.params contains all model parameters including the formula.train contains a resample object referencing the data that each model was fitted to.test contains a resample object referencing test data that models were not fitted to (for later use).
Cross-validation pairs
Cross-validation pairs can be customized with learn_cvpairs(). The following implements k-fold cross validation with k = 4.
pl %>%
learn_cvpairs(crossv_kfold, k = 4) %>%
learn()
Notice the five rows where the model has been fitted to training data for each fold, represented by cv_pairs.id. The precise training data sets are also stored under train.
Learning curves
Learning curves can be customized wth learn_curves(). The following will fit the model to three proportions of the training data (.5, .75, and 1):
pl %>%
learn_curves(.5, .75, 1) %>%
learn()
Notice the three rows where the model has been fitted to the three proportions of the training data, represented by train_p. Again, train contains references to the precise data used in each case.
More models
Add more models with learn_models(). For example, the following adds a decision tree to be fitted:
pl %>%
learn_models(rpart::rpart, Sepal.Length ~ .) %>%
learn()
Notice two rows where the regression and decision tree models have been fit to the training data, represented by models.id. The different model calls also appear under model.
Things to know about learn_models():
- Unlike the other
learn_*() functions, it can be called multiple times within the pipeline.
- It is called implicitly by
pipelearner() when arguments beyond a data frame are supplied. For example, pipelearner(d, l, f, ...) is equivalent to pipelearner(d) %>% learn_models(l, f, ...).
- Its arguments can all be vectors, which will be expanded to all combinations. This makes it easy to do things like compare many models with the same formulas, compare many different formulas, or do grid-search.
For example, the following fits two models with three formulas:
pipelearner(iris) %>%
learn_models(c(lm, rpart::rpart),
c(Sepal.Length ~ Sepal.Width,
Sepal.Length ~ Sepal.Width + Petal.Length,
Sepal.Length ~ Sepal.Width + Petal.Length + Species)) %>%
learn()
The following fits a regression model and grid-searches hyperparameters of a decision tree:
pipelearner(iris) %>%
learn_models(lm, Sepal.Length ~ .) %>%
learn_models(rpart::rpart, Sepal.Length ~ .,
minsplit = c(2, 20), cp = c(0.01, 0.1)) %>%
learn()
Remember that these additional parameters (including different formulas) are contained under params.
Bringing it all together
After initialization, pipelearner functions can be combined in a single pipeline. For example, the following will:
- Initialize a blank pipelearner object with the
iris data set.
- Create 50 cross-validation pairs (holding out random 20% of data by default in each)...
- to each be fitted in sample size proportions of .5 to 1 in increments of .1.
- With a regression modelling all interactions...
- and a decision tree modelling all features.
- Fit all models and return the results.
iris %>%
pipelearner() %>%
learn_cvpairs(crossv_mc, n = 50) %>%
learn_curves(seq(.5, 1, by = .1)) %>%
learn_models(lm, Sepal.Width ~ .*.) %>%
learn_models(rpart::rpart, Sepal.Width ~ .) %>%
learn()
Beyond learning
As you can see, pipelearner makes it easy to fit many models. The next step is to extract performance metrics from the tibble of results. This is where prior familiarity working with tidyverse tools becomes useful if not essential.
At present, pipelearner doesn't provide functions to extract any further information. This is because the information to be extracted can vary considerably between the models fitted to the data.
The following will demonstrate an example of visualising learning curves by extracting performance information from regression models.
r_square() is setup to extract an R-squared value. It is based on modelr::rsquare, but adjusted to handle new data sets (I've submitted an issue to incorporate into modelr).
# R-Squared scoring (because modelr rsquare doen't work right now)
response_var <- function(model) {
formula(model)[[2L]]
}
response <- function(model, data) {
eval(response_var(model), as.data.frame(data))
}
r_square <- function(model, data) {
actual <- response(model, data)
residuals <- predict(model, data) - actual
1 - (var(residuals, na.rm = TRUE) / var(actual, na.rm = TRUE))
}
Using a subset of the weather data from the nycflights13 package, fit a single regression model to 50 cross-validation pairs, holding out 15% of the data for testing in each case, in iterative training proportions. Note heavy use of tidyverse functions.
library(tidyverse)
# Create the data set
library(nycflights13)
d <- weather %>%
select(visib, humid, precip, wind_dir) %>%
drop_na() %>%
sample_n(2000)
results <- d %>%
pipelearner() %>%
learn_cvpairs(crossv_mc, n = 50, test = .15) %>%
learn_curves(seq(.1, 1, by = .1)) %>%
learn_models(lm, visib ~ .) %>%
learn()
results
New columns are added with dplyr::mutate containing the rsquared values for each set of training and test data by using purrr functions.
results <- results %>%
mutate(
rsquare_train = map2_dbl(fit, train, r_square),
rsquare_test = map2_dbl(fit, test, r_square)
)
results %>% select(cv_pairs.id, train_p, contains("rsquare"))
We can visualize these learning curves as follows:
results %>%
select(train_p, contains("rsquare")) %>%
gather(source, rsquare, contains("rsquare")) %>%
ggplot(aes(train_p, rsquare, color = source)) +
geom_jitter(width = .03, alpha = .3) +
stat_summary(geom = "line", fun.y = mean) +
stat_summary(geom = "point", fun.y = mean, size = 4) +
labs(x = "Proportion of training data used",
y = "R Squared")
The example below fits a decision tree and random forest to 20 folds of a subset of the data.
results <- d %>%
pipelearner() %>%
learn_cvpairs(crossv_kfold, k = 20) %>%
learn_models(c(rpart::rpart, randomForest::randomForest),
visib ~ .) %>%
learn()
results
Then compute R-Square statistics and visualize the results:
results %>%
mutate(rsquare_train = map2_dbl(fit, train, r_square),
rsquare_test = map2_dbl(fit, test, r_square)) %>%
select(model, contains("rsquare")) %>%
gather(source, rsquare, contains("rsquare")) %>%
ggplot(aes(model, rsquare, color = source)) +
geom_jitter(width = .05, alpha = .3) +
stat_summary(geom = "point", fun.y = mean, size = 4) +
labs(x = "Learning model",
y = "R Squared")
drsimonj/pipelearner documentation built on May 15, 2019, 2:53 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.path = "README-" )
pipelearner makes it easy to create machine learning pipelines in R.
Installation and background
pipelearner is currently available from github as a development package only. It can be installed by running:
# install.packages("devtools") devtools::install_github("drsimonj/pipelearner")
pipelearner is built on top of tidyverse packages like modelr. To harness the full power of pipelearner, it will help to possess some technical knowledge of tidyverse tools such as:
%>%the pipe operator from magrittr package.- tibbles from tibble package.
map()and other iteration functions from purrr package.resampleobjects from modelr package.
An excellent resource to get started with these is R for Data Science, by Garrett Grolemund and Hadley Wickham.
API
Similar to the way ggplot2 elements are layered with +, you initialize and customize a pipelearner object, which is a list, with functions that can be piped into eachother with %>%. Rather than plotting, however, a pipelearner then learns.
Initialize a pipelearner object with:
pipelearner()
Customize a pipelearner with:
learn_cvpairs()to customize the cross-validation pairs.learn_curves()to customize the learning curves using incremental proportions of training data.learn_models()to add new learning models.
Learn (fit) everything and obtain a tibble of results with:
learn()
Initialization
The following initializes a pipelearner object that will use the iris data set and linear regression (lm) to learn how to predict Sepal.Length with all other available variables (Sepal.Length ~ .).
library(pipelearner) pl <- pipelearner(iris, lm, Sepal.Length ~ .)
Print a pipelearner object to expose the list elements.
pl
Defaults to note
datais split into a single cross-validation pair of resample objects (undercv_pairs) referencing 80% of the data for training and 20% for testing.- Learning is done on the entire proportion of the training data (
train_ps == 1).
Learning
Once a pipelearner is setup, use learn() to fit all models to every combination of training proportions (train_ps) and set of training data in the cross-validation pairs (cv_pairs), and return a tibble of the results.
pl %>% learn()
Quick notes
fitcontains the fitted models.paramscontains all model parameters including the formula.traincontains a resample object referencing the data that each model was fitted to.testcontains a resample object referencing test data that models were not fitted to (for later use).
Cross-validation pairs
Cross-validation pairs can be customized with learn_cvpairs(). The following implements k-fold cross validation with k = 4.
pl %>% learn_cvpairs(crossv_kfold, k = 4) %>% learn()
Notice the five rows where the model has been fitted to training data for each fold, represented by cv_pairs.id. The precise training data sets are also stored under train.
Learning curves
Learning curves can be customized wth learn_curves(). The following will fit the model to three proportions of the training data (.5, .75, and 1):
pl %>% learn_curves(.5, .75, 1) %>% learn()
Notice the three rows where the model has been fitted to the three proportions of the training data, represented by train_p. Again, train contains references to the precise data used in each case.
More models
Add more models with learn_models(). For example, the following adds a decision tree to be fitted:
pl %>% learn_models(rpart::rpart, Sepal.Length ~ .) %>% learn()
Notice two rows where the regression and decision tree models have been fit to the training data, represented by models.id. The different model calls also appear under model.
Things to know about learn_models():
- Unlike the other
learn_*()functions, it can be called multiple times within the pipeline. - It is called implicitly by
pipelearner()when arguments beyond a data frame are supplied. For example,pipelearner(d, l, f, ...)is equivalent topipelearner(d) %>% learn_models(l, f, ...). - Its arguments can all be vectors, which will be expanded to all combinations. This makes it easy to do things like compare many models with the same formulas, compare many different formulas, or do grid-search.
For example, the following fits two models with three formulas:
pipelearner(iris) %>% learn_models(c(lm, rpart::rpart), c(Sepal.Length ~ Sepal.Width, Sepal.Length ~ Sepal.Width + Petal.Length, Sepal.Length ~ Sepal.Width + Petal.Length + Species)) %>% learn()
The following fits a regression model and grid-searches hyperparameters of a decision tree:
pipelearner(iris) %>% learn_models(lm, Sepal.Length ~ .) %>% learn_models(rpart::rpart, Sepal.Length ~ ., minsplit = c(2, 20), cp = c(0.01, 0.1)) %>% learn()
Remember that these additional parameters (including different formulas) are contained under params.
Bringing it all together
After initialization, pipelearner functions can be combined in a single pipeline. For example, the following will:
- Initialize a blank pipelearner object with the
irisdata set. - Create 50 cross-validation pairs (holding out random 20% of data by default in each)...
- to each be fitted in sample size proportions of .5 to 1 in increments of .1.
- With a regression modelling all interactions...
- and a decision tree modelling all features.
- Fit all models and return the results.
iris %>% pipelearner() %>% learn_cvpairs(crossv_mc, n = 50) %>% learn_curves(seq(.5, 1, by = .1)) %>% learn_models(lm, Sepal.Width ~ .*.) %>% learn_models(rpart::rpart, Sepal.Width ~ .) %>% learn()
Beyond learning
As you can see, pipelearner makes it easy to fit many models. The next step is to extract performance metrics from the tibble of results. This is where prior familiarity working with tidyverse tools becomes useful if not essential.
At present, pipelearner doesn't provide functions to extract any further information. This is because the information to be extracted can vary considerably between the models fitted to the data.
The following will demonstrate an example of visualising learning curves by extracting performance information from regression models.
r_square() is setup to extract an R-squared value. It is based on modelr::rsquare, but adjusted to handle new data sets (I've submitted an issue to incorporate into modelr).
# R-Squared scoring (because modelr rsquare doen't work right now) response_var <- function(model) { formula(model)[[2L]] } response <- function(model, data) { eval(response_var(model), as.data.frame(data)) } r_square <- function(model, data) { actual <- response(model, data) residuals <- predict(model, data) - actual 1 - (var(residuals, na.rm = TRUE) / var(actual, na.rm = TRUE)) }
Using a subset of the weather data from the nycflights13 package, fit a single regression model to 50 cross-validation pairs, holding out 15% of the data for testing in each case, in iterative training proportions. Note heavy use of tidyverse functions.
library(tidyverse) # Create the data set library(nycflights13) d <- weather %>% select(visib, humid, precip, wind_dir) %>% drop_na() %>% sample_n(2000) results <- d %>% pipelearner() %>% learn_cvpairs(crossv_mc, n = 50, test = .15) %>% learn_curves(seq(.1, 1, by = .1)) %>% learn_models(lm, visib ~ .) %>% learn() results
New columns are added with dplyr::mutate containing the rsquared values for each set of training and test data by using purrr functions.
results <- results %>% mutate( rsquare_train = map2_dbl(fit, train, r_square), rsquare_test = map2_dbl(fit, test, r_square) ) results %>% select(cv_pairs.id, train_p, contains("rsquare"))
We can visualize these learning curves as follows:
results %>% select(train_p, contains("rsquare")) %>% gather(source, rsquare, contains("rsquare")) %>% ggplot(aes(train_p, rsquare, color = source)) + geom_jitter(width = .03, alpha = .3) + stat_summary(geom = "line", fun.y = mean) + stat_summary(geom = "point", fun.y = mean, size = 4) + labs(x = "Proportion of training data used", y = "R Squared")
The example below fits a decision tree and random forest to 20 folds of a subset of the data.
results <- d %>% pipelearner() %>% learn_cvpairs(crossv_kfold, k = 20) %>% learn_models(c(rpart::rpart, randomForest::randomForest), visib ~ .) %>% learn() results
Then compute R-Square statistics and visualize the results:
results %>% mutate(rsquare_train = map2_dbl(fit, train, r_square), rsquare_test = map2_dbl(fit, test, r_square)) %>% select(model, contains("rsquare")) %>% gather(source, rsquare, contains("rsquare")) %>% ggplot(aes(model, rsquare, color = source)) + geom_jitter(width = .05, alpha = .3) + stat_summary(geom = "point", fun.y = mean, size = 4) + labs(x = "Learning model", y = "R Squared")
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.