There are a number of prediction functions in XGBoost with various parameters. This
document attempts to clarify some of confusions around prediction with a focus on the
Python binding, R package is similar when strict_shape is specified (see below).
There are a number of different prediction options for the
xgboost.Booster.predict() method, ranging from pred_contribs to
pred_leaf. The output shape depends on types of prediction. Also for multi-class
classification problem, XGBoost builds one tree for each class and the trees for each
class are called a “group” of trees, so output dimension may change due to used model.
After 1.4 release, we added a new parameter called strict_shape, one can set it to
True to indicate a more restricted output is desired. Assuming you are using
xgboost.Booster, here is a list of possible returns:
When using normal prediction with strict_shape set to True:
Output is a 2-dim array with first dimension as rows and second as groups. For
regression/survival/ranking/binary classification this is equivalent to a column vector
with shape[1] == 1. But for multi-class with multi:softprob the number of
columns equals to number of classes. If strict_shape is set to False then XGBoost might
output 1 or 2 dim array.
When using output_margin to avoid transformation and strict_shape is set to True:
Similar to the previous case, output is a 2-dim array, except for that multi:softmax
has equivalent output shape of multi:softprob due to dropped transformation. If
strict shape is set to False then output can have 1 or 2 dim depending on used model.
When using preds_contribs with strict_shape set to True:
Output is a 3-dim array, with (rows, groups, columns + 1) as shape. Whether
approx_contribs is used does not change the output shape. If the strict shape
parameter is not set, it can be a 2 or 3 dimension array depending on whether
multi-class model is being used.
When using preds_interactions with strict_shape set to True:
Output is a 4-dim array, with (rows, groups, columns + 1, columns + 1) as shape.
Like the predict contribution case, whether approx_contribs is used does not change
the output shape. If strict shape is set to False, it can have 3 or 4 dims depending on
the underlying model.
When using pred_leaf with strict_shape set to True:
Output is a 4-dim array with (n_samples, n_iterations, n_classes, n_trees_in_forest)
as shape. n_trees_in_forest is specified by the numb_parallel_tree during
training. When strict shape is set to False, output is a 2-dim array with last 3 dims
concatenated into 1. Also the last dimension is dropped if it eqauls to 1. When using
apply method in scikit learn interface, this is set to False by default.
For R package, when strict_shape is specified, an array is returned, with the same
value as Python except R array is column-major while Python numpy array is row-major, so
all the dimensions are reversed. For example, for a Python predict_leaf output
obtained by having strict_shape=True has 4 dimensions: (n_samples, n_iterations,
n_classes, n_trees_in_forest), while R with strict_shape=TRUE outputs
(n_trees_in_forest, n_classes, n_iterations, n_samples).
Other than these prediction types, there’s also a parameter called iteration_range,
which is similar to model slicing. But instead of actually splitting up the model into
multiple stacks, it simply returns the prediction formed by the trees within range.
Number of trees created in each iteration eqauls to \(trees_i = num\_class \times
num\_parallel\_tree\). So if you are training a boosted random forest with size of 4, on
the 3-class classification dataset, and want to use the first 2 iterations of trees for
prediction, you need to provide iteration_range=(0, 2). Then the first \(2
\times 3 \times 4\) trees will be used in this prediction.
When a model is trained with early stopping, there is an inconsistent behavior between
native Python interface and sklearn/R interfaces. By default on R and sklearn interfaces,
the best_iteration is automatically used so prediction comes from the best model. But
with the native Python interface xgboost.Booster.predict() and
xgboost.Booster.inplace_predict() uses the full model. Users can use
best_iteration attribute with iteration_range parameter to achieve the same
behavior. Also the save_best parameter from xgboost.callback.EarlyStopping
might be useful.
There are 2 predictors in XGBoost (3 if you have the one-api plugin enabled), namely
cpu_predictor and gpu_predictor. The default option is auto so that XGBoost
can employ some heuristics for saving GPU memory during training. They might have slight
different outputs due to floating point errors.
There’s a training parameter in XGBoost called base_score, and a meta data for
DMatrix called base_margin (which can be set in fit method if you are using
scikit-learn interface). They specifies the global bias for boosted model. If the latter
is supplied then former is ignored. base_margin can be used to train XGBoost model
based on other models. See demos on boosting from predictions.
Using the native interface with DMatrix, prediction can be staged (or cached). For
example, one can first predict on the first 4 trees then run prediction on 8 trees. After
running the first prediction, result from first 4 trees are cached so when you run the
prediction with 8 trees XGBoost can reuse the result from previous prediction. The cache
expires automatically upon next prediction, train or evaluation if the cached DMatrix
object is expired (like going out of scope and being collected by garbage collector in
your language environment).
Traditionally XGBoost accepts only DMatrix for prediction, with wrappers like
scikit-learn interface the construction happens internally. We added support for in-place
predict to bypass the construction of DMatrix, which is slow and memory consuming.
The new predict function has limited features but is often sufficient for simple inference
tasks. It accepts some commonly found data types in Python like numpy.ndarray,
scipy.sparse.csr_matrix and cudf.DataFrame instead of
xgboost.DMatrix. You can call xgboost.Booster.inplace_predict() to use
it. Be aware that the output of in-place prediction depends on input data type, when
input is on GPU data output is cupy.ndarray, otherwise a numpy.ndarray
is returned.
Other than users performing encoding, XGBoost has experimental support for categorical
data using gpu_hist and gpu_predictor. No special operation needs to be done on
input test data since the information about categories is encoded into the model during
training.
After 1.4 release, all prediction functions including normal predict with various
parameters like shap value computation and inplace_predict are thread safe when
underlying booster is gbtree or dart, which means as long as tree model is used,
prediction itself should thread safe. But the safety is only guaranteed with prediction.
If one tries to train a model in one thread and provide prediction at the other using the
same model the behaviour is undefined. This happens easier than one might expect, for
instance we might accidientally call clf.set_params() inside a predict function:
def predict_fn(clf: xgb.XGBClassifier, X):
X = preprocess(X)
clf.set_params(predictor="gpu_predictor") # NOT safe!
clf.set_params(n_jobs=1) # NOT safe!
return clf.predict_proba(X, iteration_range=(0, 10))
with ThreadPoolExecutor(max_workers=10) as e:
e.submit(predict_fn, ...)