def test_preprocessor_in_checkpoint(ray_start_4_cpus, tmpdir):
train_dataset = ray.data.from_pandas(train_df)
valid_dataset = ray.data.from_pandas(test_df)
class DummyPreprocessor(Preprocessor):
def __init__(self):
super().__init__()
self.is_same = True
def fit(self, dataset):
self.fitted_ = True
def _transform_pandas(self, df: "pd.DataFrame") -> "pd.DataFrame":
return df
trainer = XGBoostTrainer(
scaling_config=scale_config,
label_column="target",
params=params,
datasets={TRAIN_DATASET_KEY: train_dataset, "valid": valid_dataset},
preprocessor=DummyPreprocessor(),
)
result = trainer.fit()
# Move checkpoint to a different directory.
checkpoint_dict = result.checkpoint.to_dict()
checkpoint = Checkpoint.from_dict(checkpoint_dict)
checkpoint_path = checkpoint.to_directory(tmpdir)
resume_from = Checkpoint.from_directory(checkpoint_path)
model, preprocessor = load_checkpoint(resume_from)
assert get_num_trees(model) == 10
assert preprocessor.is_same
assert preprocessor.fitted_
def test_tune(ray_start_4_cpus):
train_dataset = ray.data.from_pandas(train_df)
valid_dataset = ray.data.from_pandas(test_df)
trainer = XGBoostTrainer(
scaling_config=scale_config,
label_column="target",
params={
**params,
**{
"max_depth": 1
}
},
datasets={
TRAIN_DATASET_KEY: train_dataset,
"valid": valid_dataset
},
)
tune.run(
trainer.as_trainable(),
config={"params": {
"max_depth": tune.randint(2, 4)
}},
num_samples=2,
)
# Make sure original Trainer is not affected.
assert trainer.params["max_depth"] == 1
def train_xgboost(num_workers: int, use_gpu: bool = False) -> Result:
train_dataset, valid_dataset, _ = prepare_data()
# Scale some random columns
columns_to_scale = ["mean radius", "mean texture"]
preprocessor = StandardScaler(columns=columns_to_scale)
# XGBoost specific params
params = {
"tree_method": "approx",
"objective": "binary:logistic",
"eval_metric": ["logloss", "error"],
}
trainer = XGBoostTrainer(
scaling_config={
"num_workers": num_workers,
"use_gpu": use_gpu,
},
label_column="target",
params=params,
datasets={
"train": train_dataset,
"valid": valid_dataset
},
preprocessor=preprocessor,
num_boost_round=100,
)
result = trainer.fit()
print(result.metrics)
return result
def test_fit(ray_start_4_cpus):
train_dataset = ray.data.from_pandas(train_df)
valid_dataset = ray.data.from_pandas(test_df)
trainer = XGBoostTrainer(
scaling_config=scale_config,
label_column="target",
params=params,
datasets={TRAIN_DATASET_KEY: train_dataset, "valid": valid_dataset},
)
trainer.fit()
def test_validation(ray_start_4_cpus):
train_dataset = ray.data.from_pandas(train_df)
valid_dataset = ray.data.from_pandas(test_df)
with pytest.raises(KeyError, match=TRAIN_DATASET_KEY):
XGBoostTrainer(
scaling_config={"num_workers": 2},
label_column="target",
params=params,
datasets={"valid": valid_dataset},
)
with pytest.raises(KeyError, match="dmatrix_params"):
XGBoostTrainer(
scaling_config={"num_workers": 2},
label_column="target",
params=params,
dmatrix_params={"data": {}},
datasets={TRAIN_DATASET_KEY: train_dataset, "valid": valid_dataset},
)
def train_model(train_dataset: ray.data.Dataset, wandb_project: str) -> Result:
"""Train a simple XGBoost model and return the result."""
trainer = XGBoostTrainer(
scaling_config={"num_workers": 2},
params={"tree_method": "auto"},
label_column="target",
datasets={"train": train_dataset},
num_boost_round=10,
run_config=RunConfig(callbacks=[
# This is the part needed to enable logging to Weights & Biases.
# It assumes you've logged in before, e.g. with `wandb login`.
WandbLoggerCallback(
project=wandb_project,
save_checkpoints=True,
)
]),
)
result = trainer.fit()
return result
def train_model(train_dataset: ray.data.Dataset, comet_project: str) -> Result:
"""Train a simple XGBoost model and return the result."""
trainer = XGBoostTrainer(
scaling_config={"num_workers": 2},
params={"tree_method": "auto"},
label_column="target",
datasets={"train": train_dataset},
num_boost_round=10,
run_config=RunConfig(callbacks=[
# This is the part needed to enable logging to Comet ML.
# It assumes Comet ML can find a valid API (e.g. by setting
# the ``COMET_API_KEY`` environment variable).
CometLoggerCallback(
project_name=comet_project,
save_checkpoints=True,
)
]),
)
result = trainer.fit()
return result
def test_resume_from_checkpoint(ray_start_4_cpus, tmpdir):
train_dataset = ray.data.from_pandas(train_df)
valid_dataset = ray.data.from_pandas(test_df)
trainer = XGBoostTrainer(
scaling_config=scale_config,
label_column="target",
params=params,
num_boost_round=5,
datasets={TRAIN_DATASET_KEY: train_dataset, "valid": valid_dataset},
)
result = trainer.fit()
checkpoint = result.checkpoint
xgb_model, _ = load_checkpoint(checkpoint)
assert get_num_trees(xgb_model) == 5
# Move checkpoint to a different directory.
checkpoint_dict = result.checkpoint.to_dict()
checkpoint = Checkpoint.from_dict(checkpoint_dict)
checkpoint_path = checkpoint.to_directory(tmpdir)
resume_from = Checkpoint.from_directory(checkpoint_path)
trainer = XGBoostTrainer(
scaling_config=scale_config,
label_column="target",
params=params,
num_boost_round=5,
datasets={TRAIN_DATASET_KEY: train_dataset, "valid": valid_dataset},
resume_from_checkpoint=resume_from,
)
result = trainer.fit()
checkpoint = result.checkpoint
model, _ = load_checkpoint(checkpoint)
assert get_num_trees(model) == 10
def test_tuner_with_xgboost_trainer(self):
"""Test a successful run."""
shutil.rmtree(os.path.join(DEFAULT_RESULTS_DIR, "test_tuner"),
ignore_errors=True)
trainer = XGBoostTrainer(
label_column="target",
params={},
# TODO(xwjiang): change when dataset out-of-band ser/des is landed.
datasets={"train": gen_dataset_func_eager()},
)
# prep_v1 = StandardScaler(["worst radius", "worst area"])
# prep_v2 = StandardScaler(["worst concavity", "worst smoothness"])
param_space = {
"scaling_config": {
"num_workers": tune.grid_search([1, 2]),
},
# TODO(xwjiang): Add when https://github.com/ray-project/ray/issues/23363
# is resolved.
# "preprocessor": tune.grid_search([prep_v1, prep_v2]),
# "datasets": {
# "train": tune.choice(
# [gen_dataset_func(), gen_dataset_func(do_shuffle=True)]
# ),
# },
"params": {
"objective": "binary:logistic",
"tree_method": "approx",
"eval_metric": ["logloss", "error"],
"eta": tune.loguniform(1e-4, 1e-1),
"subsample": tune.uniform(0.5, 1.0),
"max_depth": tune.randint(1, 9),
},
}
tuner = Tuner(
trainable=trainer,
run_config=RunConfig(name="test_tuner"),
param_space=param_space,
tune_config=TuneConfig(mode="min", metric="train-error"),
)
results = tuner.fit()
assert not isinstance(results.get_best_result().checkpoint,
TrialCheckpoint)
assert len(results) == 2
def test_tuner_with_xgboost_trainer(self):
"""Test a successful run."""
shutil.rmtree(
os.path.join(DEFAULT_RESULTS_DIR, "test_tuner"), ignore_errors=True
)
trainer = XGBoostTrainer(
label_column="target",
params={},
datasets={"train": gen_dataset_func_eager()},
)
# prep_v1 = StandardScaler(["worst radius", "worst area"])
# prep_v2 = StandardScaler(["worst concavity", "worst smoothness"])
param_space = {
"scaling_config": {
"num_workers": tune.grid_search([1, 2]),
},
# "preprocessor": tune.grid_search([prep_v1, prep_v2]),
"datasets": {
"train": tune.grid_search(
[gen_dataset_func(), gen_dataset_func(do_shuffle=True)]
),
},
"params": {
"objective": "binary:logistic",
"tree_method": "approx",
"eval_metric": ["logloss", "error"],
"eta": tune.loguniform(1e-4, 1e-1),
"subsample": tune.uniform(0.5, 1.0),
"max_depth": tune.randint(1, 9),
},
}
tuner = Tuner(
trainable=trainer,
run_config=RunConfig(name="test_tuner"),
param_space=param_space,
tune_config=TuneConfig(mode="min", metric="train-error"),
# limiting the number of trials running at one time.
# As the unit test only has access to 4 CPUs on Buildkite.
_tuner_kwargs={"max_concurrent_trials": 1},
)
results = tuner.fit()
assert not isinstance(results.get_best_result().checkpoint, TrialCheckpoint)
assert len(results) == 4
params = {
"tree_method": "approx",
"objective": "binary:logistic",
"eval_metric": ["logloss", "error"],
}
num_workers = 2
use_gpu = False # use GPUs if detected.
trainer = XGBoostTrainer(
scaling_config={
"num_workers": num_workers,
"use_gpu": use_gpu,
},
label_column="target",
params=params,
datasets={
"train": train_dataset,
"valid": valid_dataset
},
preprocessor=preprocessor,
num_boost_round=20,
)
result = trainer.fit()
print(result.metrics)
# __air_xgb_train_end__
# __air_xgb_batchpred_start__
from ray.ml.batch_predictor import BatchPredictor
from ray.ml.predictors.integrations.xgboost import XGBoostPredictor
batch_predictor = BatchPredictor.from_checkpoint(result.checkpoint,
def test_tuner_with_xgboost_trainer_driver_fail_and_resume(self):
# So that we have some global checkpointing happening.
os.environ["TUNE_GLOBAL_CHECKPOINT_S"] = "1"
shutil.rmtree(
os.path.join(DEFAULT_RESULTS_DIR, "test_tuner_driver_fail"),
ignore_errors=True,
)
trainer = XGBoostTrainer(
label_column="target",
params={},
# TODO(xwjiang): change when dataset out-of-band ser/des is landed.
datasets={"train": gen_dataset_func_eager()},
)
# prep_v1 = StandardScaler(["worst radius", "worst area"])
# prep_v2 = StandardScaler(["worst concavity", "worst smoothness"])
param_space = {
"scaling_config": {
"num_workers": tune.grid_search([1, 2]),
},
# TODO(xwjiang): Add when https://github.com/ray-project/ray/issues/23363
# is resolved.
# "preprocessor": tune.grid_search([prep_v1, prep_v2]),
# "datasets": {
# "train": tune.choice(
# [gen_dataset_func(), gen_dataset_func(do_shuffle=True)]
# ),
# },
"params": {
"objective": "binary:logistic",
"tree_method": "approx",
"eval_metric": ["logloss", "error"],
"eta": tune.loguniform(1e-4, 1e-1),
"subsample": tune.uniform(0.5, 1.0),
"max_depth": tune.randint(1, 9),
},
}
class FailureInjectionCallback(Callback):
"""Inject failure at the configured iteration number."""
def __init__(self, num_iters=10):
self.num_iters = num_iters
def on_step_end(self, iteration, trials, **kwargs):
if iteration == self.num_iters:
print(f"Failing after {self.num_iters} iters.")
raise RuntimeError
tuner = Tuner(
trainable=trainer,
run_config=RunConfig(name="test_tuner_driver_fail",
callbacks=[FailureInjectionCallback()]),
param_space=param_space,
tune_config=TuneConfig(mode="min", metric="train-error"),
)
with self.assertRaises(TuneError):
tuner.fit()
# Test resume
restore_path = os.path.join(DEFAULT_RESULTS_DIR,
"test_tuner_driver_fail")
tuner = Tuner.restore(restore_path)
# A hack before we figure out RunConfig semantics across resumes.
tuner._local_tuner._run_config.callbacks = None
results = tuner.fit()
assert len(results) == 2