def test_avoids_inferring_signature_if_not_needed(logger):
# We create an infer_model_signature that modifies the value of x
# If infer_model_signature was not invoked, x should not have been modified.
x = {"data": 0}
def modifies(_):
x["data"] = 1
resolve_input_example_and_signature(get_input_example, modifies, True,
False, logger)
assert x["data"] == 0
logger.warning.assert_not_called()
def test_happy_path_works(logger):
input_example, signature = resolve_input_example_and_signature(
get_input_example, infer_model_signature, True, True, logger)
assert input_example == "data"
assert signature == "signature"
logger.warning.assert_not_called()
def test_if_model_signature_inference_fails(logger):
error_msg = "stack overflow"
def throws(_):
raise Exception(error_msg)
input_example, signature = resolve_input_example_and_signature(
get_input_example, throws, True, True, logger)
assert input_example == "data"
assert signature is None
logger.warning.assert_called_with("Failed to infer model signature: " +
error_msg)
def test_if_getting_input_example_fails(logger):
error_msg = "NoneType has no whatever"
def throws():
raise Exception(error_msg)
input_example, signature = resolve_input_example_and_signature(
throws, infer_model_signature, True, True, logger)
assert input_example is None
assert signature is None
calls = [
call("Failed to gather input example: " + error_msg),
call("Failed to infer model signature: " +
"could not sample data to infer model signature: " + error_msg),
]
assert logger.warning.has_calls(calls)
def _log_posttraining_metadata(estimator, *args, **kwargs):
"""
Records metadata for a scikit-learn estimator after training has completed.
This is intended to be invoked within a patched scikit-learn training routine
(e.g., `fit()`, `fit_transform()`, ...) and assumes the existence of an active
MLflow run that can be referenced via the fluent Tracking API.
:param estimator: The scikit-learn estimator for which to log metadata.
:param args: The arguments passed to the scikit-learn training routine (e.g.,
`fit()`, `fit_transform()`, ...).
:param kwargs: The keyword arguments passed to the scikit-learn training routine.
"""
if hasattr(estimator, "score"):
try:
score_args = _get_args_for_score(estimator.score,
estimator.fit, args, kwargs)
training_score = estimator.score(*score_args)
except Exception as e:
msg = (
estimator.score.__qualname__ +
" failed. The 'training_score' metric will not be recorded. Scoring error: "
+ str(e))
_logger.warning(msg)
else:
try_mlflow_log(mlflow.log_metric, "training_score",
training_score)
# log common metrics and artifacts for estimators (classifier, regressor)
_log_specialized_estimator_content(estimator,
mlflow.active_run().info.run_id,
args, kwargs)
def get_input_example():
# Fetch an input example using the first several rows of the array-like
# training data supplied to the training routine (e.g., `fit()`)
fit_arg_names = _get_arg_names(estimator.fit)
X_var_name, y_var_name = fit_arg_names[:2]
input_example = _get_Xy(args, kwargs, X_var_name,
y_var_name)[0][:INPUT_EXAMPLE_SAMPLE_ROWS]
return input_example
def infer_model_signature(input_example):
if not hasattr(estimator, "predict"):
raise Exception(
"the trained model does not specify a `predict` function, "
+ "which is required in order to infer the signature")
return infer_signature(input_example,
estimator.predict(input_example))
if log_models:
# Will only resolve `input_example` and `signature` if `log_models` is `True`.
input_example, signature = resolve_input_example_and_signature(
get_input_example,
infer_model_signature,
log_input_examples,
log_model_signatures,
_logger,
)
try_mlflow_log(
log_model,
estimator,
artifact_path="model",
signature=signature,
input_example=input_example,
)
if _is_parameter_search_estimator(estimator):
if hasattr(estimator, "best_estimator_") and log_models:
try_mlflow_log(
log_model,
estimator.best_estimator_,
artifact_path="best_estimator",
signature=signature,
input_example=input_example,
)
if hasattr(estimator, "best_score_"):
try_mlflow_log(mlflow.log_metric, "best_cv_score",
estimator.best_score_)
if hasattr(estimator, "best_params_"):
best_params = {
"best_{param_name}".format(param_name=param_name):
param_value
for param_name, param_value in
estimator.best_params_.items()
}
try_mlflow_log(mlflow.log_params, best_params)
if hasattr(estimator, "cv_results_"):
try:
# Fetch environment-specific tags (e.g., user and source) to ensure that lineage
# information is consistent with the parent run
child_tags = context_registry.resolve_tags()
child_tags.update({MLFLOW_AUTOLOGGING: FLAVOR_NAME})
_create_child_runs_for_parameter_search(
cv_estimator=estimator,
parent_run=mlflow.active_run(),
child_tags=child_tags,
)
except Exception as e:
msg = (
"Encountered exception during creation of child runs for parameter search."
" Child runs may be missing. Exception: {}".format(
str(e)))
_logger.warning(msg)
try:
cv_results_df = pd.DataFrame.from_dict(
estimator.cv_results_)
_log_parameter_search_results_as_artifact(
cv_results_df,
mlflow.active_run().info.run_id)
except Exception as e:
msg = (
"Failed to log parameter search results as an artifact."
" Exception: {}".format(str(e)))
_logger.warning(msg)
def train(original, *args, **kwargs):
def record_eval_results(eval_results, metrics_logger):
"""
Create a callback function that records evaluation results.
"""
@exception_safe_function
def callback(env):
res = {}
for data_name, eval_name, value, _ in env.evaluation_result_list:
key = data_name + "-" + eval_name
res[key] = value
metrics_logger.record_metrics(res, env.iteration)
eval_results.append(res)
return callback
def log_feature_importance_plot(features, importance, importance_type):
"""
Log feature importance plot.
"""
import matplotlib.pyplot as plt
indices = np.argsort(importance)
features = np.array(features)[indices]
importance = importance[indices]
num_features = len(features)
# If num_features > 10, increase the figure height to prevent the plot
# from being too dense.
w, h = [6.4, 4.8] # matplotlib's default figure size
h = h + 0.1 * num_features if num_features > 10 else h
fig, ax = plt.subplots(figsize=(w, h))
yloc = np.arange(num_features)
ax.barh(yloc, importance, align="center", height=0.5)
ax.set_yticks(yloc)
ax.set_yticklabels(features)
ax.set_xlabel("Importance")
ax.set_title("Feature Importance ({})".format(importance_type))
fig.tight_layout()
tmpdir = tempfile.mkdtemp()
try:
# pylint: disable=undefined-loop-variable
filepath = os.path.join(
tmpdir, "feature_importance_{}.png".format(imp_type))
fig.savefig(filepath)
try_mlflow_log(mlflow.log_artifact, filepath)
finally:
plt.close(fig)
shutil.rmtree(tmpdir)
# logging booster params separately via mlflow.log_params to extract key/value pairs
# and make it easier to compare them across runs.
params = args[0] if len(args) > 0 else kwargs["params"]
try_mlflow_log(mlflow.log_params, params)
unlogged_params = [
"params",
"train_set",
"valid_sets",
"valid_names",
"fobj",
"feval",
"init_model",
"evals_result",
"learning_rates",
"callbacks",
]
log_fn_args_as_params(original, args, kwargs, unlogged_params)
all_arg_names = inspect.getargspec(original)[0] # pylint: disable=W1505
num_pos_args = len(args)
# adding a callback that records evaluation results.
eval_results = []
callbacks_index = all_arg_names.index("callbacks")
run_id = mlflow.active_run().info.run_id
with batch_metrics_logger(run_id) as metrics_logger:
callback = record_eval_results(eval_results, metrics_logger)
if num_pos_args >= callbacks_index + 1:
tmp_list = list(args)
tmp_list[callbacks_index] += [callback]
args = tuple(tmp_list)
elif "callbacks" in kwargs and kwargs["callbacks"] is not None:
kwargs["callbacks"] += [callback]
else:
kwargs["callbacks"] = [callback]
# training model
model = original(*args, **kwargs)
# If early_stopping_rounds is present, logging metrics at the best iteration
# as extra metrics with the max step + 1.
early_stopping_index = all_arg_names.index("early_stopping_rounds")
early_stopping = (num_pos_args >= early_stopping_index + 1
or "early_stopping_rounds" in kwargs)
if early_stopping:
extra_step = len(eval_results)
metrics_logger.record_metrics(
{"stopped_iteration": extra_step})
# best_iteration is set even if training does not stop early.
metrics_logger.record_metrics(
{"best_iteration": model.best_iteration})
# iteration starts from 1 in LightGBM.
results = eval_results[model.best_iteration - 1]
metrics_logger.record_metrics(results, step=extra_step)
# logging feature importance as artifacts.
for imp_type in ["split", "gain"]:
features = model.feature_name()
importance = model.feature_importance(importance_type=imp_type)
try:
log_feature_importance_plot(features, importance, imp_type)
except Exception:
_logger.exception(
"Failed to log feature importance plot. LightGBM autologging "
"will ignore the failure and continue. Exception: ")
imp = {ft: imp for ft, imp in zip(features, importance.tolist())}
tmpdir = tempfile.mkdtemp()
try:
filepath = os.path.join(
tmpdir, "feature_importance_{}.json".format(imp_type))
with open(filepath, "w") as f:
json.dump(imp, f, indent=2)
try_mlflow_log(mlflow.log_artifact, filepath)
finally:
shutil.rmtree(tmpdir)
# train_set must exist as the original train function already ran successfully
train_set = args[1] if len(args) > 1 else kwargs.get("train_set")
# it is possible that the dataset was constructed before the patched
# constructor was applied, so we cannot assume the input_example_info exists
input_example_info = getattr(train_set, "input_example_info", None)
def get_input_example():
if input_example_info is None:
raise Exception(ENSURE_AUTOLOGGING_ENABLED_TEXT)
if input_example_info.error_msg is not None:
raise Exception(input_example_info.error_msg)
return input_example_info.input_example
def infer_model_signature(input_example):
model_output = model.predict(input_example)
model_signature = infer_signature(input_example, model_output)
return model_signature
# Whether to automatically log the trained model based on boolean flag.
if log_models:
# Will only resolve `input_example` and `signature` if `log_models` is `True`.
input_example, signature = resolve_input_example_and_signature(
get_input_example,
infer_model_signature,
log_input_examples,
log_model_signatures,
_logger,
)
try_mlflow_log(
log_model,
model,
artifact_path="model",
signature=signature,
input_example=input_example,
)
return model
def train(original, *args, **kwargs):
def record_eval_results(eval_results, metrics_logger):
"""
Create a callback function that records evaluation results.
"""
if LooseVersion(xgboost.__version__) >= LooseVersion("1.3.0"):
# In xgboost >= 1.3.0, user-defined callbacks should inherit
# `xgboost.callback.TrainingCallback`:
# https://xgboost.readthedocs.io/en/latest/python/callbacks.html#defining-your-own-callback # noqa
class Callback(
xgboost.callback.TrainingCallback,
metaclass=ExceptionSafeAbstractClass,
):
def after_iteration(self, model, epoch, evals_log):
"""
Run after each iteration. Return True when training should stop.
"""
# `evals_log` is a nested dict (type: Dict[str, Dict[str, List[float]]])
# that looks like this:
# {
# "train": {
# "auc": [0.5, 0.6, 0.7, ...],
# ...
# },
# ...
# }
evaluation_result_dict = {}
for data_name, metric_dict in evals_log.items():
for metric_name, metric_values_on_each_iter in metric_dict.items(
):
key = "{}-{}".format(data_name, metric_name)
# The last element in `metric_values_on_each_iter` corresponds to
# the meric on the current iteration
evaluation_result_dict[
key] = metric_values_on_each_iter[-1]
metrics_logger.record_metrics(evaluation_result_dict,
epoch)
eval_results.append(evaluation_result_dict)
# Return `False` to indicate training should not stop
return False
return Callback()
else:
@exception_safe_function
def callback(env):
metrics_logger.record_metrics(
dict(env.evaluation_result_list), env.iteration)
eval_results.append(dict(env.evaluation_result_list))
return callback
def log_feature_importance_plot(features, importance, importance_type):
"""
Log feature importance plot.
"""
import matplotlib.pyplot as plt
features = np.array(features)
importance = np.array(importance)
indices = np.argsort(importance)
features = features[indices]
importance = importance[indices]
num_features = len(features)
# If num_features > 10, increase the figure height to prevent the plot
# from being too dense.
w, h = [6.4, 4.8] # matplotlib's default figure size
h = h + 0.1 * num_features if num_features > 10 else h
fig, ax = plt.subplots(figsize=(w, h))
yloc = np.arange(num_features)
ax.barh(yloc, importance, align="center", height=0.5)
ax.set_yticks(yloc)
ax.set_yticklabels(features)
ax.set_xlabel("Importance")
ax.set_title("Feature Importance ({})".format(importance_type))
fig.tight_layout()
tmpdir = tempfile.mkdtemp()
try:
# pylint: disable=undefined-loop-variable
filepath = os.path.join(
tmpdir, "feature_importance_{}.png".format(imp_type))
fig.savefig(filepath)
try_mlflow_log(mlflow.log_artifact, filepath)
finally:
plt.close(fig)
shutil.rmtree(tmpdir)
# logging booster params separately via mlflow.log_params to extract key/value pairs
# and make it easier to compare them across runs.
params = args[0] if len(args) > 0 else kwargs["params"]
try_mlflow_log(mlflow.log_params, params)
unlogged_params = [
"params",
"dtrain",
"evals",
"obj",
"feval",
"evals_result",
"xgb_model",
"callbacks",
"learning_rates",
]
log_fn_args_as_params(original, args, kwargs, unlogged_params)
all_arg_names = inspect.getargspec(original)[0] # pylint: disable=W1505
num_pos_args = len(args)
# adding a callback that records evaluation results.
eval_results = []
callbacks_index = all_arg_names.index("callbacks")
run_id = mlflow.active_run().info.run_id
with batch_metrics_logger(run_id) as metrics_logger:
callback = record_eval_results(eval_results, metrics_logger)
if num_pos_args >= callbacks_index + 1:
tmp_list = list(args)
tmp_list[callbacks_index] += [callback]
args = tuple(tmp_list)
elif "callbacks" in kwargs and kwargs["callbacks"] is not None:
kwargs["callbacks"] += [callback]
else:
kwargs["callbacks"] = [callback]
# training model
model = original(*args, **kwargs)
# If early_stopping_rounds is present, logging metrics at the best iteration
# as extra metrics with the max step + 1.
early_stopping_index = all_arg_names.index("early_stopping_rounds")
early_stopping = (num_pos_args >= early_stopping_index + 1
or "early_stopping_rounds" in kwargs)
if early_stopping:
extra_step = len(eval_results)
metrics_logger.record_metrics(
{"stopped_iteration": extra_step - 1})
metrics_logger.record_metrics(
{"best_iteration": model.best_iteration})
metrics_logger.record_metrics(
eval_results[model.best_iteration], extra_step)
# logging feature importance as artifacts.
for imp_type in importance_types:
imp = None
try:
imp = model.get_score(importance_type=imp_type)
features, importance = zip(*imp.items())
log_feature_importance_plot(features, importance, imp_type)
except Exception:
_logger.exception(
"Failed to log feature importance plot. XGBoost autologging "
"will ignore the failure and continue. Exception: ")
if imp is not None:
tmpdir = tempfile.mkdtemp()
try:
filepath = os.path.join(
tmpdir, "feature_importance_{}.json".format(imp_type))
with open(filepath, "w") as f:
json.dump(imp, f)
try_mlflow_log(mlflow.log_artifact, filepath)
finally:
shutil.rmtree(tmpdir)
# dtrain must exist as the original train function already ran successfully
dtrain = args[1] if len(args) > 1 else kwargs.get("dtrain")
# it is possible that the dataset was constructed before the patched
# constructor was applied, so we cannot assume the input_example_info exists
input_example_info = getattr(dtrain, "input_example_info", None)
def get_input_example():
if input_example_info is None:
raise Exception(ENSURE_AUTOLOGGING_ENABLED_TEXT)
if input_example_info.error_msg is not None:
raise Exception(input_example_info.error_msg)
return input_example_info.input_example
def infer_model_signature(input_example):
model_output = model.predict(xgboost.DMatrix(input_example))
model_signature = infer_signature(input_example, model_output)
return model_signature
# Only log the model if the autolog() param log_models is set to True.
if log_models:
# Will only resolve `input_example` and `signature` if `log_models` is `True`.
input_example, signature = resolve_input_example_and_signature(
get_input_example,
infer_model_signature,
log_input_examples,
log_model_signatures,
_logger,
)
try_mlflow_log(
log_model,
model,
artifact_path="model",
signature=signature,
input_example=input_example,
)
return model