drsimonj / pipelearner Goto Github PK

I have another issue. This looks to be triggered by the contains("rsquare")) part.

results %>%

• add_rsquare() %>% 
  

• select(cv_pairs.id, contains("rsquare")) %>% 
  

• gather(source, rsquare, contains("rsquare")) %>%
  

• mutate(source = gsub("rsquare_", "", source)) %>% 
  

• ggplot(aes(cv_pairs.id, rsquare, color = source)) +
  

• geom_point() +
  

• labs(x = "Fold", y = "R Squared")
  

Error in .p(.x[[i]], ...) : argument ".y" is missing, with no default

traceback()
20: .p(.x[[i]], ...)
19: isTRUE(.p(.x[[i]], ...))
18: some(.x, identical, .y)
17: contains("rsquare")
16: eval(expr, envir, enclos)
15: eval(x$expr, data, x$env)
14: FUN(X[[i]], ...)
13: lapply(x, lazy_eval, data = data)
12: lazyeval::lazy_eval(args, names_list)
11: select_vars_(names(.data), dots)
10: select_.data.frame(.data, .dots = lazyeval::lazy_dots(...))
9: select_(.data, .dots = lazyeval::lazy_dots(...))
8: select(., cv_pairs.id, contains("rsquare"))
7: function_list[i]
6: freduce(value, _function_list)
5: _fseq(_lhs)
4: eval(expr, envir, enclos)
3: eval(quote(_fseq(_lhs)), env, env)
2: withVisible(eval(quote(_fseq (_lhs)), env, env))
1: results %>% add_rsquare() %>% select(cv_pairs.id, contains("rsquare")) %>%
gather(source, rsquare, contains("rsquare")) %>% mutate(source = gsub("rsquare_",
"", source)) %>% ggplot(aes(cv_pairs.id, rsquare, color = source))

What do you think is causing this?`

Currently, predict.pipelearner assumes that the output of predict on each fit will return a single vector of values. However, this isn't always the case. For example, default settings on predict.rpart when a classification tree is run will return a data frame of predicted probabilities.

Current version of purrr (0.2.3) results in following error being thrown:

`cross_d()` is deprecated; please use `cross_df()` instead. 

As I was going through your example I ran into this issue?
'> results <- d %>%

• pipelearner(lm, visib ~ .) %>% 
  

• learn_cvpairs(k = 10) %>% 
  

• learn()
  

Error: 'as_tibble' is not an exported object from 'namespace:tibble'

traceback()
12: stop(gettextf("'%s' is not an exported object from 'namespace:%s'",
name, getNamespaceName(ns)), call. = FALSE, domain = NA)
11: getExportedValue(pkg, name)
10: tibble::as_tibble
9: pipelearner.data.frame(., lm, visib ~ .)
8: pipelearner(., lm, visib ~ .)
7: function_list[i]
6: freduce(value, _function_list)
5: _fseq(_lhs)
4: eval(expr, envir, enclos)
3: eval(quote(_fseq(_lhs)), env, env)
2: withVisible(eval(quote(_fseq (_lhs)), env, env))
1: d %>% pipelearner(lm, visib ~ .) %>% learn_cvpairs(k = 10) %>%
learn()`

It turned out that I needed a new version of tibble. I thought I got that when I installed tidyverse but I guess I was incorrect.

Currently, predict.pipelearner() applies default predict() to each fit. However, sometimes this needs to be adjusted. For example, changing predict to produce probabilities or classes.

In general, a user will want to predict values and score/evaluated their fit after learning all models via learn(). The exact functions to do this are many. However, pipeable functions could be written that takes the tibble coming from learn() as well as a function that will take the relevant columns (e.g., test, target, and fit), and output the predicted values. It will then be the responsibility of the user to create a function that accepts these arguments.

e.g.,...

pl %>% learn() %>%
pl_predict("test_hat", FUN = function(test, target, fit) {
   # etc... to produce vectors of fitted values
}) %>%
pl_score("test_rsqr", FUN = function(test, target, test_hat) {
   # etc...
})

The functions in resamplr also use modelr::resample objects and includes all cross validation methods from scikit-learn.

It looks like learn_cvpairs can be written the same way as learn_models with arbitrary cross validation functions as long as the cross validation function returns a df with train, test, and .id columns (which is the case with resamplr).

I ran into an issue using gam from the mcgv package with pipelearner. To illustrate, lm can be used like this:

iris %>%
  lm(Sepal.Length ~ Sepal.Width, .)

and so can be piped into pipelearner:

iris %>%
  pipelearner(lm, Sepal.Length ~ Sepal.Width) %>%
  learn()

However, gam requires "data = " explicitly. The analogous expression to lm fails:

iris %>%
  gam(Sepal.Length ~ s(Sepal.Width), .)
Error in eval(expr, envir, enclos) : object 'Sepal.Length' not found

iris %>%
  pipelearner(gam, Sepal.Length ~ s(Sepal.Width)) %>%
  learn()
Error in eval(expr, envir, enclos) : object 'Sepal.Length' not found

This works:

iris %>%
  gam(Sepal.Length ~ s(Sepal.Width), data = .)

Is there a syntax which will allow pipelearner to run gam, or does pipelearner require modification?