Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Define best_iteration only if early stopping is used. (#9403)

Browse Source 
* Define `best_iteration` only if early stopping is used.

This is the behavior specified by the document but not honored in the actual code.

- Don't set the attributes if there's no early stopping.
- Clean up the code for callbacks, and replace assertions with proper exceptions.
- Assign the attributes when early stopping `save_best` is used.
- Turn the attributes into Python properties.

---------

Co-authored-by: Philip Hyunsu Cho <chohyu01@cs.washington.edu>
This commit is contained in:
Jiaming Yuan
2023-07-24 12:43:35 +08:00
committed by GitHub
parent 01e00efc53
commit 851cba931e
10 changed files with 249 additions and 179 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tests/ci_build/lint_python.py
View File

@@ -37,6 +37,7 @@ class LintersPaths:
"demo/rmm_plugin",
"demo/json-model/json_parser.py",
"demo/guide-python/cat_in_the_dat.py",
"demo/guide-python/callbacks.py",
"demo/guide-python/categorical.py",
"demo/guide-python/feature_weights.py",
"demo/guide-python/sklearn_parallel.py",

127
tests/python/test_callback.py
View File

@@ -1,7 +1,6 @@
import json
import os
import tempfile
from contextlib import nullcontext
from typing import Union
import pytest
@@ -104,15 +103,6 @@ class TestCallbacks:
dump = booster.get_dump(dump_format='json')
assert len(dump) - booster.best_iteration == early_stopping_rounds + 1
# No early stopping, best_iteration should be set to last epoch
booster = xgb.train({'objective': 'binary:logistic',
'eval_metric': 'error'}, D_train,
evals=[(D_train, 'Train'), (D_valid, 'Valid')],
num_boost_round=10,
evals_result=evals_result,
verbose_eval=True)
assert booster.num_boosted_rounds() - 1 == booster.best_iteration
def test_early_stopping_custom_eval(self):
D_train = xgb.DMatrix(self.X_train, self.y_train)
D_valid = xgb.DMatrix(self.X_valid, self.y_valid)
@@ -204,8 +194,9 @@ class TestCallbacks:
X, y = load_breast_cancer(return_X_y=True)
n_estimators = 100
early_stopping_rounds = 5
early_stop = xgb.callback.EarlyStopping(rounds=early_stopping_rounds,
save_best=True)
early_stop = xgb.callback.EarlyStopping(
rounds=early_stopping_rounds, save_best=True
)
cls = xgb.XGBClassifier(
n_estimators=n_estimators,
eval_metric=tm.eval_error_metric_skl,
@@ -216,20 +207,27 @@ class TestCallbacks:
dump = booster.get_dump(dump_format='json')
assert len(dump) == booster.best_iteration + 1
early_stop = xgb.callback.EarlyStopping(rounds=early_stopping_rounds,
save_best=True)
early_stop = xgb.callback.EarlyStopping(
rounds=early_stopping_rounds, save_best=True
)
cls = xgb.XGBClassifier(
booster='gblinear', n_estimators=10, eval_metric=tm.eval_error_metric_skl
booster="gblinear",
n_estimators=10,
eval_metric=tm.eval_error_metric_skl,
callbacks=[early_stop],
)
with pytest.raises(ValueError):
cls.fit(X, y, eval_set=[(X, y)], callbacks=[early_stop])
cls.fit(X, y, eval_set=[(X, y)])
# No error
early_stop = xgb.callback.EarlyStopping(rounds=early_stopping_rounds,
save_best=False)
xgb.XGBClassifier(
booster='gblinear', n_estimators=10, eval_metric=tm.eval_error_metric_skl
).fit(X, y, eval_set=[(X, y)], callbacks=[early_stop])
booster="gblinear",
n_estimators=10,
eval_metric=tm.eval_error_metric_skl,
callbacks=[early_stop],
).fit(X, y, eval_set=[(X, y)])
def test_early_stopping_continuation(self):
from sklearn.datasets import load_breast_cancer
@@ -252,8 +250,11 @@ class TestCallbacks:
cls.load_model(path)
assert cls._Booster is not None
early_stopping_rounds = 3
cls.set_params(eval_metric=tm.eval_error_metric_skl)
cls.fit(X, y, eval_set=[(X, y)], early_stopping_rounds=early_stopping_rounds)
cls.set_params(
eval_metric=tm.eval_error_metric_skl,
early_stopping_rounds=early_stopping_rounds,
)
cls.fit(X, y, eval_set=[(X, y)])
booster = cls.get_booster()
assert booster.num_boosted_rounds() == \
booster.best_iteration + early_stopping_rounds + 1
@@ -280,20 +281,20 @@ class TestCallbacks:
watchlist = [(dtest, 'eval'), (dtrain, 'train')]
num_round = 4
warning_check = nullcontext()
# learning_rates as a list
# init eta with 0 to check whether learning_rates work
param = {'max_depth': 2, 'eta': 0, 'verbosity': 0,
'objective': 'binary:logistic', 'eval_metric': 'error',
'tree_method': tree_method}
evals_result = {}
with warning_check:
bst = xgb.train(param, dtrain, num_round, watchlist,
callbacks=[scheduler([
0.8, 0.7, 0.6, 0.5
])],
evals_result=evals_result)
bst = xgb.train(
param,
dtrain,
num_round,
evals=watchlist,
callbacks=[scheduler([0.8, 0.7, 0.6, 0.5])],
evals_result=evals_result,
)
eval_errors_0 = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster)
# validation error should decrease, if eta > 0
@@ -304,11 +305,15 @@ class TestCallbacks:
'objective': 'binary:logistic', 'eval_metric': 'error',
'tree_method': tree_method}
evals_result = {}
with warning_check:
bst = xgb.train(param, dtrain, num_round, watchlist,
callbacks=[scheduler(
[0.8, 0.7, 0.6, 0.5])],
evals_result=evals_result)
bst = xgb.train(
param,
dtrain,
num_round,
evals=watchlist,
callbacks=[scheduler([0.8, 0.7, 0.6, 0.5])],
evals_result=evals_result,
)
eval_errors_1 = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster)
# validation error should decrease, if learning_rate > 0
@@ -320,12 +325,14 @@ class TestCallbacks:
'eval_metric': 'error', 'tree_method': tree_method
}
evals_result = {}
with warning_check:
bst = xgb.train(param, dtrain, num_round, watchlist,
callbacks=[scheduler(
[0, 0, 0, 0]
)],
evals_result=evals_result)
bst = xgb.train(
param,
dtrain,
num_round,
evals=watchlist,
callbacks=[scheduler([0, 0, 0, 0])],
evals_result=evals_result,
)
eval_errors_2 = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster)
# validation error should not decrease, if eta/learning_rate = 0
@@ -336,12 +343,14 @@ class TestCallbacks:
return num_boost_round / (ithround + 1)
evals_result = {}
with warning_check:
bst = xgb.train(param, dtrain, num_round, watchlist,
callbacks=[
scheduler(eta_decay)
],
evals_result=evals_result)
bst = xgb.train(
param,
dtrain,
num_round,
evals=watchlist,
callbacks=[scheduler(eta_decay)],
evals_result=evals_result,
)
eval_errors_3 = list(map(float, evals_result['eval']['error']))
assert isinstance(bst, xgb.core.Booster)
@@ -351,8 +360,7 @@ class TestCallbacks:
for i in range(1, len(eval_errors_0)):
assert eval_errors_3[i] != eval_errors_2[i]
with warning_check:
xgb.cv(param, dtrain, num_round, callbacks=[scheduler(eta_decay)])
xgb.cv(param, dtrain, num_round, callbacks=[scheduler(eta_decay)])
def run_eta_decay_leaf_output(self, tree_method: str, objective: str) -> None:
# check decay has effect on leaf output.
@@ -378,7 +386,7 @@ class TestCallbacks:
param,
dtrain,
num_round,
watchlist,
evals=watchlist,
callbacks=[scheduler(eta_decay_0)],
)
@@ -391,7 +399,7 @@ class TestCallbacks:
param,
dtrain,
num_round,
watchlist,
evals=watchlist,
callbacks=[scheduler(eta_decay_1)],
)
bst_json0 = bst0.save_raw(raw_format="json")
@@ -474,3 +482,24 @@ class TestCallbacks:
callbacks=callbacks,
)
assert len(callbacks) == 1
def test_attribute_error(self) -> None:
from sklearn.datasets import load_breast_cancer
X, y = load_breast_cancer(return_X_y=True)
clf = xgb.XGBClassifier(n_estimators=8)
clf.fit(X, y, eval_set=[(X, y)])
with pytest.raises(AttributeError, match="early stopping is used"):
clf.best_iteration
with pytest.raises(AttributeError, match="early stopping is used"):
clf.best_score
booster = clf.get_booster()
with pytest.raises(AttributeError, match="early stopping is used"):
booster.best_iteration
with pytest.raises(AttributeError, match="early stopping is used"):
booster.best_score

4
tests/python/test_predict.py
View File

@@ -173,7 +173,7 @@ class TestInplacePredict:
np.testing.assert_allclose(predt_from_dmatrix, predt_from_array)
with pytest.raises(ValueError):
booster.predict(test, iteration_range=(0, booster.best_iteration + 2))
booster.predict(test, iteration_range=(0, booster.num_boosted_rounds() + 2))
default = booster.predict(test)
@@ -181,7 +181,7 @@ class TestInplacePredict:
np.testing.assert_allclose(range_full, default)
range_full = booster.predict(
test, iteration_range=(0, booster.best_iteration + 1)
test, iteration_range=(0, booster.num_boosted_rounds())
)
np.testing.assert_allclose(range_full, default)

13
tests/python/test_training_continuation.py
View File

@@ -100,8 +100,8 @@ class TestTrainingContinuation:
res2 = mean_squared_error(
y_2class,
gbdt_04.predict(
dtrain_2class, iteration_range=(0, gbdt_04.best_iteration + 1)
)
dtrain_2class, iteration_range=(0, gbdt_04.num_boosted_rounds())
),
)
assert res1 == res2
@@ -112,7 +112,7 @@ class TestTrainingContinuation:
res2 = mean_squared_error(
y_2class,
gbdt_04.predict(
dtrain_2class, iteration_range=(0, gbdt_04.best_iteration + 1)
dtrain_2class, iteration_range=(0, gbdt_04.num_boosted_rounds())
)
)
assert res1 == res2
@@ -126,7 +126,7 @@ class TestTrainingContinuation:
res1 = gbdt_05.predict(dtrain_5class)
res2 = gbdt_05.predict(
dtrain_5class, iteration_range=(0, gbdt_05.best_iteration + 1)
dtrain_5class, iteration_range=(0, gbdt_05.num_boosted_rounds())
)
np.testing.assert_almost_equal(res1, res2)
@@ -138,15 +138,16 @@ class TestTrainingContinuation:
@pytest.mark.skipif(**tm.no_sklearn())
def test_training_continuation_updaters_json(self):
# Picked up from R tests.
updaters = 'grow_colmaker,prune,refresh'
updaters = "grow_colmaker,prune,refresh"
params = self.generate_parameters()
for p in params:
p['updater'] = updaters
p["updater"] = updaters
self.run_training_continuation(params[0], params[1], params[2])
@pytest.mark.skipif(**tm.no_sklearn())
def test_changed_parameter(self):
from sklearn.datasets import load_breast_cancer
X, y = load_breast_cancer(return_X_y=True)
clf = xgb.XGBClassifier(n_estimators=2)
clf.fit(X, y, eval_set=[(X, y)], eval_metric="logloss")