tidymodels: error when predicting on new data with xgboost model

[anon73295571]

I trained two xgboost models. One was trained without resampling calling fit() right after model specifications were made like this...

model <- boost_tree() %>%
set_mode() %>%
set_engine() %>%
fit()%>%
predict()

.. and it works just fine, my predictions were made.

Another was trained with resampling setting up a workflow with tune::tune_grid(). Once tune_grid trained the model, i selected the best model with tune::select_best(), finalized the workflow with tune::finalize_workflow() and extracted my final model with a fit() call on the trainning data and lastly workflow::pull_workflow_fit(). However when i want to make predictions with these model on new data, the following error shows up:

Fehler in xgboost::xgb.DMatrix(data = newdata, missing = NA) : 
  'data' has class 'character' and length 1178.
  'data' accepts either a numeric matrix or a single filename.

The new data is provided to the predict() function as exactly the same tibble. Anyone knows an answer to this problem?

[julia]

If you have used one of the tune_* functions to find the best parameters for your model and then finalized your workflow with those parameters, the next step is to train or fit that workflow one more time on the whole training set. Let's walk through an example.

library(tidymodels)
#> ── Attaching packages ────────────────────────────────────── tidymodels 0.1.1 ──
#> ✓ broom     0.7.0      ✓ recipes   0.1.13
#> ✓ dials     0.0.8      ✓ rsample   0.0.7 
#> ✓ dplyr     1.0.0      ✓ tibble    3.0.3 
#> ✓ ggplot2   3.3.2      ✓ tidyr     1.1.0 
#> ✓ infer     0.5.3      ✓ tune      0.1.1 
#> ✓ modeldata 0.0.2      ✓ workflows 0.1.2 
#> ✓ parsnip   0.1.2      ✓ yardstick 0.0.7 
#> ✓ purrr     0.3.4
#> ── Conflicts ───────────────────────────────────────── tidymodels_conflicts() ──
#> x purrr::discard() masks scales::discard()
#> x dplyr::filter()  masks stats::filter()
#> x dplyr::lag()     masks stats::lag()
#> x recipes::step()  masks stats::step()

## pretend this is your training data
data("hpc_data")

xgb_spec <- boost_tree(
  trees = 1000, 
  tree_depth = tune(), 
  min_n = tune()
) %>% 
  set_engine("xgboost") %>% 
  set_mode("classification")

hpc_folds <- vfold_cv(hpc_data, strata = class)

xgb_grid <- grid_latin_hypercube(
  tree_depth(),
  min_n(),
  size = 5
)

xgb_wf <- workflow() %>%
  add_formula(class ~ .) %>%
  add_model(xgb_spec)

xgb_wf
#> ══ Workflow ════════════════════════════════════════════════════════════════════
#> Preprocessor: Formula
#> Model: boost_tree()
#> 
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> class ~ .
#> 
#> ── Model ───────────────────────────────────────────────────────────────────────
#> Boosted Tree Model Specification (classification)
#> 
#> Main Arguments:
#>   trees = 1000
#>   min_n = tune()
#>   tree_depth = tune()
#> 
#> Computational engine: xgboost

doParallel::registerDoParallel()
set.seed(123)
xgb_res <- tune_grid(
  xgb_wf,
  resamples = hpc_folds,
  grid = xgb_grid
)

xgb_res
#> # Tuning results
#> # 10-fold cross-validation using stratification 
#> # A tibble: 10 x 4
#>    splits             id     .metrics          .notes          
#>    <list>             <chr>  <list>            <list>          
#>  1 <split [3.9K/434]> Fold01 <tibble [10 × 6]> <tibble [0 × 1]>
#>  2 <split [3.9K/434]> Fold02 <tibble [10 × 6]> <tibble [0 × 1]>
#>  3 <split [3.9K/434]> Fold03 <tibble [10 × 6]> <tibble [0 × 1]>
#>  4 <split [3.9K/434]> Fold04 <tibble [10 × 6]> <tibble [0 × 1]>
#>  5 <split [3.9K/434]> Fold05 <tibble [10 × 6]> <tibble [0 × 1]>
#>  6 <split [3.9K/434]> Fold06 <tibble [10 × 6]> <tibble [0 × 1]>
#>  7 <split [3.9K/433]> Fold07 <tibble [10 × 6]> <tibble [0 × 1]>
#>  8 <split [3.9K/432]> Fold08 <tibble [10 × 6]> <tibble [0 × 1]>
#>  9 <split [3.9K/431]> Fold09 <tibble [10 × 6]> <tibble [0 × 1]>
#> 10 <split [3.9K/431]> Fold10 <tibble [10 × 6]> <tibble [0 × 1]>

Next, let's finalize this workflow and then fit() it to the training data. (The tuning process used the training data, but that was to find the best model parameters, not to train the model itself.)

trained_wf <- xgb_wf %>%
  finalize_workflow(
  select_best(xgb_res, "roc_auc")
) %>%
  fit(hpc_data)

trained_wf
#> ══ Workflow [trained] ══════════════════════════════════════════════════════════
#> Preprocessor: Formula
#> Model: boost_tree()
#> 
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> class ~ .
#> 
#> ── Model ───────────────────────────────────────────────────────────────────────
#> ##### xgb.Booster
#> raw: 2 Mb 
#> call:
#>   xgboost::xgb.train(params = list(eta = 0.3, max_depth = 3L, gamma = 0, 
#>     colsample_bytree = 1, min_child_weight = 9L, subsample = 1), 
#>     data = x, nrounds = 1000, watchlist = wlist, verbose = 0, 
#>     objective = "multi:softprob", num_class = 4L, nthread = 1)
#> params (as set within xgb.train):
#>   eta = "0.3", max_depth = "3", gamma = "0", colsample_bytree = "1", min_child_weight = "9", subsample = "1", objective = "multi:softprob", num_class = "4", nthread = "1", validate_parameters = "TRUE"
#> xgb.attributes:
#>   niter
#> callbacks:
#>   cb.evaluation.log()
#> # of features: 26 
#> niter: 1000
#> nfeatures : 26 
#> evaluation_log:
#>     iter training_merror
#>        1        0.320942
#>        2        0.301778
#> ---                     
#>      999        0.010390
#>     1000        0.010390

Now let's say we have some brand new data. You can predict() with new data on the trained workflow.

brand_new_data <- hpc_data[5, -8]
brand_new_data
#> # A tibble: 1 x 7
#>   protocol compounds input_fields iterations num_pending  hour day  
#>   <fct>        <dbl>        <dbl>      <dbl>       <dbl> <dbl> <fct>
#> 1 E              100           82         20           0  10.4 Fri

predict(trained_wf, new_data = brand_new_data)
#> # A tibble: 1 x 1
#>   .pred_class
#>   <fct>      
#> 1 VF

^{Created on 2020-07-17 by the reprex package (v0.3.0)}

The thing to remember is that if you fit the workflow then you need to predict on the workflow. If you want to predict on a model (vs. a workflow) you can tune a model specification.