#!/usr/bin/env python
# -*- encoding: utf-8 -*-
#
# Copyright 2016 H2O.ai; Apache License Version 2.0 (see LICENSE for details)
#
from __future__ import absolute_import, division, print_function, unicode_literals
from datetime import datetime
import inspect
import types
import warnings
import h2o
from h2o.exceptions import H2OValueError, H2OResponseError
from h2o.frame import H2OFrame
from h2o.job import H2OJob
from h2o.utils.compatibility import * # NOQA
from h2o.utils.shared_utils import quoted
from h2o.utils.typechecks import assert_is_type, is_type, numeric, FunctionType
from ..model.autoencoder import H2OAutoEncoderModel
from ..model.binomial import H2OBinomialModel
from ..model.clustering import H2OClusteringModel
from ..model.dim_reduction import H2ODimReductionModel
from ..model.metrics_base import (H2OBinomialModelMetrics, H2OClusteringModelMetrics, H2ORegressionModelMetrics,
H2OMultinomialModelMetrics, H2OAutoEncoderModelMetrics, H2ODimReductionModelMetrics,
H2OWordEmbeddingModelMetrics, H2OOrdinalModelMetrics, H2OAnomalyDetectionModelMetrics,
H2OCoxPHModelMetrics)
from ..model.model_base import ModelBase
from ..model.multinomial import H2OMultinomialModel
from ..model.ordinal import H2OOrdinalModel
from ..model.regression import H2ORegressionModel
from ..model.word_embedding import H2OWordEmbeddingModel
from ..model.anomaly_detection import H2OAnomalyDetectionModel
from ..model.coxph import H2OCoxPHModel
class EstimatorAttributeError(AttributeError):
def __init__(self, obj, method):
super(AttributeError, self).__init__("No {} method for {}".format(method, obj.__class__.__name__))
[docs]class H2OEstimator(ModelBase):
"""
Base class for H2O Estimators.
H2O Estimators implement the following methods for model construction:
- ``start()`` - Top-level user-facing API for asynchronous model build
- ``join()`` - Top-level user-facing API for blocking on async model build
- ``train()`` - Top-level user-facing API for model building.
- ``fit()`` - Used by scikit-learn.
Because H2OEstimator instances are instances of ModelBase, these objects can use the H2O model API.
"""
[docs] def start(self, x, y=None, training_frame=None, offset_column=None, fold_column=None,
weights_column=None, validation_frame=None, **params):
"""
Train the model asynchronously (to block for results call :meth:`join`).
:param x: A list of column names or indices indicating the predictor columns.
:param y: An index or a column name indicating the response column.
:param H2OFrame training_frame: The H2OFrame having the columns indicated by x and y (as well as any
additional columns specified by fold, offset, and weights).
:param offset_column: The name or index of the column in training_frame that holds the offsets.
:param fold_column: The name or index of the column in training_frame that holds the per-row fold
assignments.
:param weights_column: The name or index of the column in training_frame that holds the per-row weights.
:param validation_frame: H2OFrame with validation data to be scored on while training.
"""
self._future = True
self.train(x=x,
y=y,
training_frame=training_frame,
offset_column=offset_column,
fold_column=fold_column,
weights_column=weights_column,
validation_frame=validation_frame,
**params)
[docs] def join(self):
"""Wait until job's completion."""
self._future = False
self._job.poll()
model_key = self._job.dest_key
self._job = None
model_json = h2o.api("GET /%d/Models/%s" % (self._rest_version, model_key))["models"][0]
self._resolve_model(model_key, model_json)
[docs] def train(self, x=None, y=None, training_frame=None, offset_column=None, fold_column=None,
weights_column=None, validation_frame=None, max_runtime_secs=None, ignored_columns=None,
model_id=None, verbose=False):
"""
Train the H2O model.
:param x: A list of column names or indices indicating the predictor columns.
:param y: An index or a column name indicating the response column.
:param H2OFrame training_frame: The H2OFrame having the columns indicated by x and y (as well as any
additional columns specified by fold, offset, and weights).
:param offset_column: The name or index of the column in training_frame that holds the offsets.
:param fold_column: The name or index of the column in training_frame that holds the per-row fold
assignments.
:param weights_column: The name or index of the column in training_frame that holds the per-row weights.
:param validation_frame: H2OFrame with validation data to be scored on while training.
:param float max_runtime_secs: Maximum allowed runtime in seconds for model training. Use 0 to disable.
:param bool verbose: Print scoring history to stdout. Defaults to False.
"""
self._train(x=x, y=y, training_frame=training_frame, offset_column=offset_column, fold_column=fold_column,
weights_column=weights_column, validation_frame=validation_frame, max_runtime_secs=max_runtime_secs,
ignored_columns=ignored_columns, model_id=model_id, verbose=verbose)
def _train(self, x=None, y=None, training_frame=None, offset_column=None, fold_column=None,
weights_column=None, validation_frame=None, max_runtime_secs=None, ignored_columns=None,
model_id=None, verbose=False, extend_parms_fn=None):
has_default_training_frame = hasattr(self, 'training_frame') and self.training_frame is not None
training_frame = H2OFrame._validate(training_frame, 'training_frame',
required=self._requires_training_frame() and not has_default_training_frame)
validation_frame = H2OFrame._validate(validation_frame, 'validation_frame')
assert_is_type(y, None, int, str)
assert_is_type(x, None, int, str, [str, int], {str, int})
assert_is_type(ignored_columns, None, [str, int], {str, int})
assert_is_type(offset_column, None, int, str)
assert_is_type(fold_column, None, int, str)
assert_is_type(weights_column, None, int, str)
assert_is_type(max_runtime_secs, None, numeric)
assert_is_type(model_id, None, str)
assert_is_type(verbose, bool)
assert_is_type(extend_parms_fn, None, FunctionType)
override_default_training_frame = training_frame is not None
if not override_default_training_frame:
self._verify_training_frame_params(offset_column, fold_column, weights_column, validation_frame)
training_frame = self.training_frame if has_default_training_frame else None
algo = self.algo
if verbose and algo not in ["drf", "gbm", "deeplearning", "xgboost"]:
raise H2OValueError("Verbose should only be set to True for drf, gbm, deeplearning, and xgboost models")
parms = self._parms.copy()
if "__class__" in parms: # FIXME: hackt for PY3
del parms["__class__"]
is_auto_encoder = bool(parms.get("autoencoder"))
is_supervised = not(is_auto_encoder or algo in {"aggregator", "pca", "svd", "kmeans", "glrm", "word2vec", "isolationforest", "generic"})
names = training_frame.names if training_frame is not None else []
ncols = training_frame.ncols if training_frame is not None else 0
types = training_frame.types if training_frame is not None else {}
if is_supervised:
if y is None: y = "response"
if is_type(y, int):
if not (-ncols <= y < ncols):
raise H2OValueError("Column %d does not exist in the training frame" % y)
y = names[y]
else:
if y not in names:
raise H2OValueError("Column %s does not exist in the training frame" % y)
self._estimator_type = "classifier" if types[y] == "enum" else "regressor"
else:
# If `y` is provided for an unsupervised model we'll simply ignore
# it. This way an unsupervised model can be used as a step in
# sklearn's pipeline.
y = None
if override_default_training_frame:
assert_is_type(y, str, None)
ignored_columns_set = set()
if ignored_columns is None and "ignored_columns" in parms:
ignored_columns = parms['ignored_columns']
if ignored_columns is not None:
if x is not None:
raise H2OValueError("Properties x and ignored_columns cannot be specified simultaneously")
for ic in ignored_columns:
if is_type(ic, int):
if not (-ncols <= ic < ncols):
raise H2OValueError("Column %d does not exist in the training frame" % ic)
ignored_columns_set.add(names[ic])
else:
if ic not in names:
raise H2OValueError("Column %s not in the training frame" % ic)
ignored_columns_set.add(ic)
if x is None:
xset = set(names) - {y} - ignored_columns_set
else:
xset = set()
if is_type(x, int, str): x = [x]
for xi in x:
if is_type(xi, int):
if not (-ncols <= xi < ncols):
raise H2OValueError("Column %d does not exist in the training frame" % xi)
xset.add(names[xi])
else:
if xi not in names:
raise H2OValueError("Column %s not in the training frame" % xi)
xset.add(xi)
x = list(xset)
self._check_and_save_parm(parms, "offset_column", offset_column)
self._check_and_save_parm(parms, "weights_column", weights_column)
self._check_and_save_parm(parms, "fold_column", fold_column)
if max_runtime_secs is not None: parms["max_runtime_secs"] = max_runtime_secs
# Overwrites the model_id parameter only if model_id is passed
if model_id is not None:
parms["model_id"] = model_id
# Step 2
is_auto_encoder = "autoencoder" in parms and parms["autoencoder"]
is_unsupervised = is_auto_encoder or self.algo in {"aggregator", "pca", "svd", "kmeans", "glrm", "word2vec", "isolationforest"}
if is_auto_encoder and y is not None:
raise ValueError("y should not be specified for autoencoder.")
if not is_unsupervised and y is None and self.algo not in ["generic"]:
raise ValueError("Missing response")
# Step 3
if override_default_training_frame:
parms["training_frame"] = training_frame
offset = parms["offset_column"]
folds = parms["fold_column"]
weights = parms["weights_column"]
if validation_frame is not None:
parms["validation_frame"] = validation_frame
if is_type(y, int):
y = names[y]
if y is not None:
parms["response_column"] = y
if not isinstance(x, (list, tuple)):
x = [x]
if is_type(x[0], int):
x = [names[i] for i in x]
if override_default_training_frame:
ignored_columns = list(set(names) - set(x + [y, offset, folds, weights] + self._additional_used_columns(parms)))
parms["ignored_columns"] = None if ignored_columns == [] else [quoted(col) for col in ignored_columns]
parms["interactions"] = (None if "interactions" not in parms or parms["interactions"] is None
else [quoted(col) for col in parms["interactions"]])
parms["interaction_pairs"] = (None if "interaction_pairs" not in parms or parms["interaction_pairs"] is None
else [tuple(map(quoted, ip)) for ip in parms["interaction_pairs"]])
# internal hook allowing subclasses to extend train parms
if extend_parms_fn is not None:
extend_parms_fn(parms)
parms = {k: H2OEstimator._keyify_if_h2oframe(parms[k]) for k in parms}
if ("stopping_metric" in parms.keys()) and ("r2" in parms["stopping_metric"]):
raise H2OValueError("r2 cannot be used as an early stopping_metric yet. Check this JIRA https://0xdata.atlassian.net/browse/PUBDEV-5381 for progress.")
rest_ver = parms.pop("_rest_version") if "_rest_version" in parms else 3
model_builder_json = h2o.api("POST /%d/ModelBuilders/%s" % (rest_ver, self.algo), data=parms)
model = H2OJob(model_builder_json, job_type=(self.algo + " Model Build"))
if self._future:
self._job = model
self._rest_version = rest_ver
return
model.poll(poll_updates=self._print_model_scoring_history if verbose else None)
model_json = h2o.api("GET /%d/Models/%s" % (rest_ver, model.dest_key))["models"][0]
self._resolve_model(model.dest_key, model_json)
def _print_model_scoring_history(self, job, bar_progress=0):
"""
the callback function used to poll/print updates during model training.
"""
if int(bar_progress * 10) % 5 > 0:
return
try:
model = h2o.get_model(job.job['dest']['name'])
print("\nScoring History for Model " + str(model.model_id) + " at " + str(datetime.now()))
print("Model Build is {0:.0f}% done...".format(job.progress*100))
print(model.scoring_history().tail())
print("\n")
except H2OResponseError: # To catch 400 error
print("Model build is starting now...")
except AttributeError: # To catch NoneType error if scoring history is not available
print("Scoring History is not available yet...")
@staticmethod
def _keyify_if_h2oframe(item):
if isinstance(item, H2OFrame):
return item.frame_id
elif isinstance(item, list) and all(i is None or isinstance(i, H2OFrame) for i in item):
return [quoted(i) if i is None else quoted(i.frame_id) for i in item]
else:
return item
def _resolve_model(self, model_id, model_json):
metrics_class, model_class = H2OEstimator._metrics_class(model_json)
m = model_class()
m._id = model_id
m._model_json = model_json
m._have_pojo = model_json.get('have_pojo', True)
m._have_mojo = model_json.get('have_mojo', True)
m._metrics_class = metrics_class
m._parms = self._parms
m._estimator_type = self._estimator_type
m._start_time = model_json.get('output', {}).get('start_time', None)
m._end_time = model_json.get('output', {}).get('end_time', None)
m._run_time = model_json.get('output', {}).get('run_time', None)
if model_id is not None and model_json is not None and metrics_class is not None:
# build Metric objects out of each metrics
for metric in ["training_metrics", "validation_metrics", "cross_validation_metrics"]:
if metric in model_json["output"]:
if model_json["output"][metric] is not None:
if metric == "cross_validation_metrics":
m._is_xvalidated = True
model_json["output"][metric] = \
metrics_class(model_json["output"][metric], metric, model_json["algo"])
#if m._is_xvalidated:
if m._is_xvalidated and model_json["output"]["cross_validation_models"] is not None:
m._xval_keys = [i["name"] for i in model_json["output"]["cross_validation_models"]]
# build a useful dict of the params
for p in m._model_json["parameters"]:
m.parms[p["name"]] = p
H2OEstimator.mixin(self, model_class)
self.__dict__.update(m.__dict__.copy())
# TODO: replace with a property which is overriden in subclasses
def _compute_algo(self):
name = self.__class__.__name__
if name == "H2ODeepLearningEstimator": return "deeplearning"
if name == "H2ODeepWaterEstimator": return "deepwater"
if name == "H2OAutoEncoderEstimator": return "deeplearning"
if name == "H2OGradientBoostingEstimator": return "gbm"
if name == "H2OGeneralizedLinearEstimator": return "glm"
if name == "H2OGeneralizedLowRankEstimator": return "glrm"
if name == "H2OKMeansEstimator": return "kmeans"
if name == "H2ONaiveBayesEstimator": return "naivebayes"
if name == "H2ORandomForestEstimator": return "drf"
if name == "H2OXGBoostEstimator": return "xgboost"
if name == "H2OCoxProportionalHazardsEstimator": return "coxph"
if name in ["H2OPCA", "H2OPrincipalComponentAnalysisEstimator"]: return "pca"
if name in ["H2OSVD", "H2OSingularValueDecompositionEstimator"]: return "svd"
@staticmethod
def mixin(obj, cls):
for name in cls.__dict__:
if name.startswith("__") and name.endswith("__"): continue
if not isinstance(cls.__dict__[name], types.FunctionType): continue
obj.__dict__[name] = cls.__dict__[name].__get__(obj)
#------ Scikit-learn Interface Methods -------
[docs] def fit(self, X, y=None, **params):
"""
Fit an H2O model as part of a scikit-learn pipeline or grid search.
A warning will be issued if a caller other than sklearn attempts to use this method.
:param H2OFrame X: An H2OFrame consisting of the predictor variables.
:param H2OFrame y: An H2OFrame consisting of the response variable.
:param params: Extra arguments.
:returns: The current instance of H2OEstimator for method chaining.
"""
stk = inspect.stack()[1:]
warn = True
for s in stk:
mod = inspect.getmodule(s[0])
if mod:
warn = "sklearn" not in mod.__name__
if not warn: break
if warn:
warnings.warn("\n\n\t`fit` is not recommended outside of the sklearn framework. Use `train` instead.",
UserWarning, stacklevel=2)
training_frame = X.cbind(y) if y is not None else X
x = X.names
y = y.names[0] if y is not None else None
self.train(x, y, training_frame, **params)
return self
[docs] def get_params(self, deep=True):
"""
Obtain parameters for this estimator.
Used primarily for sklearn Pipelines and sklearn grid search.
:param deep: If True, return parameters of all sub-objects that are estimators.
:returns: A dict of parameters
"""
out = dict()
for key, value in self.parms.items():
if deep and isinstance(value, H2OEstimator):
deep_items = list(value.get_params().items())
out.update((key + "__" + k, val) for k, val in deep_items)
out[key] = value
return out
[docs] def set_params(self, **parms):
"""
Used by sklearn for updating parameters during grid search.
:param parms: A dictionary of parameters that will be set on this model.
:returns: self, the current estimator object with the parameters all set as desired.
"""
self._parms.update(parms)
return self
def _verify_training_frame_params(self, *args):
for param in args:
if param is not None:
raise H2OValueError("No training frame defined, yet the parameter %d is has been specified.", param)
def _requires_training_frame(self):
"""
Determines if a training frame is required for given algorithm.
:return: True as a default value. Can be overridden by any specific algorithm.
"""
return True
def _additional_used_columns(self, parms):
"""
Returns list of additional columns not to automatically add to ignored_columns parameter.
:return: Empty list as default. Can be overridden by any specific algorithm.
"""
return []
@staticmethod
def _metrics_class(model_json):
model_type = model_json["output"]["model_category"]
if model_type == "Binomial":
metrics_class = H2OBinomialModelMetrics
model_class = H2OBinomialModel
elif model_type == "Clustering":
metrics_class = H2OClusteringModelMetrics
model_class = H2OClusteringModel
elif model_type == "Regression":
metrics_class = H2ORegressionModelMetrics
model_class = H2ORegressionModel
elif model_type == "Multinomial":
metrics_class = H2OMultinomialModelMetrics
model_class = H2OMultinomialModel
elif model_type == "Ordinal":
metrics_class = H2OOrdinalModelMetrics
model_class = H2OOrdinalModel
elif model_type == "AutoEncoder":
metrics_class = H2OAutoEncoderModelMetrics
model_class = H2OAutoEncoderModel
elif model_type == "DimReduction":
metrics_class = H2ODimReductionModelMetrics
model_class = H2ODimReductionModel
elif model_type == "WordEmbedding":
metrics_class = H2OWordEmbeddingModelMetrics
model_class = H2OWordEmbeddingModel
elif model_type == "AnomalyDetection":
metrics_class = H2OAnomalyDetectionModelMetrics
model_class = H2OAnomalyDetectionModel
elif model_type == "CoxPH":
metrics_class = H2OCoxPHModelMetrics
model_class = H2OCoxPHModel
else:
raise NotImplementedError(model_type)
return [metrics_class, model_class]
[docs] def convert_H2OXGBoostParams_2_XGBoostParams(self):
'''
In order to use convert_H2OXGBoostParams_2_XGBoostParams and convert_H2OFrame_2_DMatrix, you must import
the following toolboxes: xgboost, pandas, numpy and scipy.sparse.
Given an H2OXGBoost model, this method will generate the corresponding parameters that should be used by
native XGBoost in order to give exactly the same result, assuming that the same dataset
(derived from h2oFrame) is used to train the native XGBoost model.
Follow the steps below to compare H2OXGBoost and native XGBoost:
1. Train the H2OXGBoost model with H2OFrame trainFile and generate a prediction:
h2oModelD = H2OXGBoostEstimator(**h2oParamsD) # parameters specified as a dict()
h2oModelD.train(x=myX, y=y, training_frame=trainFile) # train with H2OFrame trainFile
h2oPredict = h2oPredictD = h2oModelD.predict(trainFile)
2. Derive the DMatrix from H2OFrame:
nativeDMatrix = trainFile.convert_H2OFrame_2_DMatrix(myX, y, h2oModelD)
3. Derive the parameters for native XGBoost:
nativeParams = h2oModelD.convert_H2OXGBoostParams_2_XGBoostParams()
4. Train your native XGBoost model and generate a prediction:
nativeModel = xgb.train(params=nativeParams[0], dtrain=nativeDMatrix, num_boost_round=nativeParams[1])
nativePredict = nativeModel.predict(data=nativeDMatrix, ntree_limit=nativeParams[1].
5. Compare the predictions h2oPredict from H2OXGBoost, nativePredict from native XGBoost.
:return: nativeParams, num_boost_round
'''
import xgboost as xgb
nativeParams = self._model_json["output"]["native_parameters"]
nativeXGBoostParams = dict()
for (a,keyname,keyvalue) in nativeParams.cell_values:
nativeXGBoostParams[keyname]=keyvalue
paramsSet = self.full_parameters
return nativeXGBoostParams, paramsSet['ntrees']['actual_value']
def _check_and_save_parm(self, parms, parameter_name, parameter_value):
"""
If a parameter is not stored in parms dict save it there (even though the value is None).
Else check if the parameter has been already set during initialization of estimator. If yes, check the new value is the same or not. If the values are different, set the last passed value to params dict and throw UserWarning.
"""
if parameter_name not in parms:
parms[parameter_name] = parameter_value
elif parameter_value is not None and parms[parameter_name] != parameter_value:
parms[parameter_name] = parameter_value
warnings.warn("\n\n\t`%s` parameter has been already set and had a different value in `train` method. The last passed value \"%s\" is used." % (parameter_name, parameter_value), UserWarning, stacklevel=2)