def test_confusion_matrix(sdf):
assem = VectorAssembler(inputCols=['Fare', 'Pclass', 'Age'],
outputCol='features')
rf = RandomForestClassifier(featuresCol='features',
labelCol='Survived',
numTrees=20)
pipeline = Pipeline(stages=[assem, rf])
model = pipeline.fit(sdf.fillna(0.0))
predictions = model.transform(sdf.fillna(0.0)).select(
'probability', 'Survived')
bcm = BinaryClassificationMetrics(predictions,
scoreCol='probability',
labelCol='Survived')
predictions = predictions.toHandy().to_metrics_RDD('probability',
'Survived')
predictions = np.array(predictions.collect())
scm = bcm.confusionMatrix().toArray()
pcm = confusion_matrix(predictions[:, 1], predictions[:, 0] > .5)
npt.assert_array_almost_equal(scm, pcm)
scm = bcm.confusionMatrix(.3).toArray()
pcm = confusion_matrix(predictions[:, 1], predictions[:, 0] > .3)
npt.assert_array_almost_equal(scm, pcm)
def model_train(input,model_path):
tmax_schema = types.StructType([
types.StructField('station', types.StringType()),
types.StructField('date', types.DateType()),
types.StructField('latitude', types.FloatType()),
types.StructField('longitude', types.FloatType()),
types.StructField('elevation', types.FloatType()),
types.StructField('tmax', types.FloatType()),
])
data = spark.read.csv(input,schema= tmax_schema)
train, validation = data.randomSplit([0.75,0.25])
train = train.cache()
validation = validation.cache()
sql_query = """SELECT today.latitude, today.longitude, today.elevation, dayofyear(today.date) AS dy,yesterday.tmax AS yesterday_tmax, today.tmax
FROM __THIS__ as today
INNER JOIN __THIS__ as yesterday
ON date_sub(today.date, 1) = yesterday.date
AND today.station = yesterday.station"""
transformer = SQLTransformer(statement=sql_query)
assemble_features = VectorAssembler(inputCols=['latitude','longitude','elevation','dy','yesterday_tmax'],outputCol='features')
regressor = DecisionTreeRegressor(featuresCol='features',labelCol='tmax')
weather_pipeline = Pipeline(stages=[transformer,assemble_features,regressor])
model = weather_pipeline.fit(train)
model.write().overwrite().save(model_path)
prediction = model.transform(validation)
#Scoring the model
evaluator = RegressionEvaluator(predictionCol='prediction',labelCol='tmax',metricName='rmse')
score = evaluator.evaluate(prediction)
print("Score of the weather model is",score)
def main():
data = spark.range(100000)
data = data.select(
(functions.rand()*100).alias('length'),
(functions.rand()*100).alias('width'),
(functions.rand()*100).alias('height'),
)
data = data.withColumn('volume', data['length']*data['width']*data['height'])
training, validation = data.randomSplit([0.75, 0.25], seed=42)
assemble_features = VectorAssembler(
inputCols=['length', 'width', 'height'],
outputCol='features')
classifier = GBTRegressor(
featuresCol='features', labelCol='volume')
pipeline = Pipeline(stages=[assemble_features, classifier])
model = pipeline.fit(training)
predictions = model.transform(validation)
predictions.show()
r2_evaluator = RegressionEvaluator(
predictionCol='prediction', labelCol='volume',
metricName='r2')
r2 = r2_evaluator.evaluate(predictions)
print(r2)
def main(spark, logger, **kwargs):
logger.info("Creating a simple DataFrame ...")
schema_names = ["id", "german_text"]
fields = [
T.StructField(field_name, T.StringType(), True) for field_name in schema_names
]
schema = T.StructType(fields)
data = [
("abc", "Hallo Herr Mustermann"),
("xyz", "Deutsch ist das Ding!"),
]
df = spark.createDataFrame(data, schema)
df.show()
logger.info("Building the ML pipeline ...")
tokenizer = RegexTokenizer(
inputCol="german_text", outputCol="tokens", pattern="\\s+"
)
stemmer = SnowballStemmer(
inputCol="tokens", outputCol="stemmed_tokens", language="German"
)
stemming_pipeline = Pipeline(
stages=[
tokenizer,
stemmer,
]
)
logger.info("Running the stemming ML pipeline ...")
stemmed_df = stemming_pipeline.fit(df).transform(df)
stemmed_df.show()
def pipeline_dataframe(self, stages, dataframe):
print(stages)
dataframe.printSchema()
pipeline = Pipeline(stages=stages)
pipelineModel = pipeline.fit(dataframe)
model = pipelineModel.transform(dataframe)
return model
def test_nested_pipeline_persistence(self):
"""
Pipeline[HashingTF, Pipeline[PCA]]
"""
temp_path = tempfile.mkdtemp()
try:
df = self.spark.createDataFrame([(["a", "b", "c"],), (["c", "d", "e"],)], ["words"])
tf = HashingTF(numFeatures=10, inputCol="words", outputCol="features")
pca = PCA(k=2, inputCol="features", outputCol="pca_features")
p0 = Pipeline(stages=[pca])
pl = Pipeline(stages=[tf, p0])
model = pl.fit(df)
pipeline_path = temp_path + "/pipeline"
pl.save(pipeline_path)
loaded_pipeline = Pipeline.load(pipeline_path)
self._compare_pipelines(pl, loaded_pipeline)
model_path = temp_path + "/pipeline-model"
model.save(model_path)
loaded_model = PipelineModel.load(model_path)
self._compare_pipelines(model, loaded_model)
finally:
try:
rmtree(temp_path)
except OSError:
pass
def fit(self, sdf):
"""
:param sdf:
:return:
"""
if self.weighter is None:
raise NotImplementedError(
"The weighter parameter has not been defined.")
weights_arr = self.weighter.get_feature_importances(sdf)
pipeline_lst = [
VectorAssembler(inputCols=self.input_cols, outputCol="vec"),
StandardScaler(inputCol="vec", outputCol="standard_vec"),
ElementwiseProduct(scalingVec=weights_arr,
inputCol='standard_vec',
outputCol='scaled_vec')
]
_model = Pipeline(stages=pipeline_lst)
model = _model.fit(sdf)
self.model = model
return self
def test_python_transformer_pipeline_persistence(self):
"""
Pipeline[MockUnaryTransformer, Binarizer]
"""
temp_path = tempfile.mkdtemp()
try:
df = self.spark.range(0, 10).toDF('input')
tf = MockUnaryTransformer(shiftVal=2)\
.setInputCol("input").setOutputCol("shiftedInput")
tf2 = Binarizer(threshold=6, inputCol="shiftedInput", outputCol="binarized")
pl = Pipeline(stages=[tf, tf2])
model = pl.fit(df)
pipeline_path = temp_path + "/pipeline"
pl.save(pipeline_path)
loaded_pipeline = Pipeline.load(pipeline_path)
self._compare_pipelines(pl, loaded_pipeline)
model_path = temp_path + "/pipeline-model"
model.save(model_path)
loaded_model = PipelineModel.load(model_path)
self._compare_pipelines(model, loaded_model)
finally:
try:
rmtree(temp_path)
except OSError:
pass
def test_pipeline(self):
dataset = MockDataset()
estimator0 = MockEstimator()
transformer1 = MockTransformer()
estimator2 = MockEstimator()
transformer3 = MockTransformer()
pipeline = Pipeline() \
.setStages([estimator0, transformer1, estimator2, transformer3])
pipeline_model = pipeline.fit(dataset, {
estimator0.fake: 0,
transformer1.fake: 1
})
self.assertEqual(0, estimator0.dataset_index)
self.assertEqual(0, estimator0.fake_param_value)
model0 = estimator0.model
self.assertEqual(0, model0.dataset_index)
self.assertEqual(1, transformer1.dataset_index)
self.assertEqual(1, transformer1.fake_param_value)
self.assertEqual(2, estimator2.dataset_index)
model2 = estimator2.model
self.assertIsNone(
model2.dataset_index,
"The model produced by the last estimator should "
"not be called during fit.")
dataset = pipeline_model.transform(dataset)
self.assertEqual(2, model0.dataset_index)
self.assertEqual(3, transformer1.dataset_index)
self.assertEqual(4, model2.dataset_index)
self.assertEqual(5, transformer3.dataset_index)
self.assertEqual(6, dataset.index)
def test_pipeline(self):
dataset = MockDataset()
estimator0 = MockEstimator()
transformer1 = MockTransformer()
estimator2 = MockEstimator()
transformer3 = MockTransformer()
pipeline = Pipeline(
stages=[estimator0, transformer1, estimator2, transformer3])
pipeline_model = pipeline.fit(dataset, {
estimator0.fake: 0,
transformer1.fake: 1
})
model0, transformer1, model2, transformer3 = pipeline_model.stages
self.assertEqual(0, model0.dataset_index)
self.assertEqual(0, model0.getFake())
self.assertEqual(1, transformer1.dataset_index)
self.assertEqual(1, transformer1.getFake())
self.assertEqual(2, dataset.index)
self.assertIsNone(model2.dataset_index,
"The last model shouldn't be called in fit.")
self.assertIsNone(transformer3.dataset_index,
"The last transformer shouldn't be called in fit.")
dataset = pipeline_model.transform(dataset)
self.assertEqual(2, model0.dataset_index)
self.assertEqual(3, transformer1.dataset_index)
self.assertEqual(4, model2.dataset_index)
self.assertEqual(5, transformer3.dataset_index)
self.assertEqual(6, dataset.index)
def test_pipeline(self):
dataset = MockDataset()
estimator0 = MockEstimator()
transformer1 = MockTransformer()
estimator2 = MockEstimator()
transformer3 = MockTransformer()
pipeline = Pipeline() \
.setStages([estimator0, transformer1, estimator2, transformer3])
pipeline_model = pipeline.fit(dataset, {estimator0.fake: 0, transformer1.fake: 1})
self.assertEqual(0, estimator0.dataset_index)
self.assertEqual(0, estimator0.fake_param_value)
model0 = estimator0.model
self.assertEqual(0, model0.dataset_index)
self.assertEqual(1, transformer1.dataset_index)
self.assertEqual(1, transformer1.fake_param_value)
self.assertEqual(2, estimator2.dataset_index)
model2 = estimator2.model
self.assertIsNone(model2.dataset_index, "The model produced by the last estimator should "
"not be called during fit.")
dataset = pipeline_model.transform(dataset)
self.assertEqual(2, model0.dataset_index)
self.assertEqual(3, transformer1.dataset_index)
self.assertEqual(4, model2.dataset_index)
self.assertEqual(5, transformer3.dataset_index)
self.assertEqual(6, dataset.index)
def test_model_log(tmpdir):
conda_env = os.path.join(str(tmpdir), "conda_env.yml")
_mlflow_conda_env(conda_env, additional_pip_deps=["pyspark=={}".format(pyspark_version)])
iris = datasets.load_iris()
feature_names = ["0", "1", "2", "3"]
pandas_df = pd.DataFrame(iris.data, columns=feature_names) # to make spark_udf work
pandas_df['label'] = pd.Series(iris.target)
spark_session = pyspark.sql.SparkSession.builder \
.config(key="spark_session.python.worker.reuse", value=True) \
.master("local-cluster[2, 1, 1024]") \
.getOrCreate()
spark_df = spark_session.createDataFrame(pandas_df)
assembler = VectorAssembler(inputCols=feature_names, outputCol="features")
lr = LogisticRegression(maxIter=50, regParam=0.1, elasticNetParam=0.8)
pipeline = Pipeline(stages=[assembler, lr])
# Fit the model
model = pipeline.fit(spark_df)
# Print the coefficients and intercept for multinomial logistic regression
preds_df = model.transform(spark_df)
preds1 = [x.prediction for x in preds_df.select("prediction").collect()]
old_tracking_uri = mlflow.get_tracking_uri()
cnt = 0
# should_start_run tests whether or not calling log_model() automatically starts a run.
for should_start_run in [False, True]:
for dfs_tmp_dir in [None, os.path.join(str(tmpdir), "test")]:
print("should_start_run =", should_start_run, "dfs_tmp_dir =", dfs_tmp_dir)
try:
tracking_dir = os.path.abspath(str(tmpdir.mkdir("mlruns")))
mlflow.set_tracking_uri("file://%s" % tracking_dir)
if should_start_run:
mlflow.start_run()
artifact_path = "model%d" % cnt
cnt += 1
sparkm.log_model(artifact_path=artifact_path, spark_model=model,
dfs_tmpdir=dfs_tmp_dir)
run_id = active_run().info.run_uuid
# test pyfunc
x = pyfunc.load_pyfunc(artifact_path, run_id=run_id)
preds2 = x.predict(pandas_df)
assert preds1 == preds2
# test load model
reloaded_model = sparkm.load_model(artifact_path, run_id=run_id,
dfs_tmpdir=dfs_tmp_dir)
preds_df_1 = reloaded_model.transform(spark_df)
preds3 = [x.prediction for x in preds_df_1.select("prediction").collect()]
assert preds1 == preds3
# test spar_udf
preds4 = score_model_as_udf(artifact_path, run_id, pandas_df)
assert preds1 == preds4
# We expect not to delete the DFS tempdir.
x = dfs_tmp_dir or sparkm.DFS_TMP
assert os.path.exists(x)
assert os.listdir(x)
shutil.rmtree(x)
finally:
mlflow.end_run()
mlflow.set_tracking_uri(old_tracking_uri)
shutil.rmtree(tracking_dir)
def train_model(training_size, mode):
print('Training model with records: ' + str(training_size))
spark = pyspark.sql.SparkSession.builder.appName(
'Model Prep').getOrCreate()
data_df = model_utils.get_player_df(spark, training_size, mode)
pipeline = Pipeline().setStages(transform_stages())
model = pipeline.fit(data_df)
model.write().overwrite().save(model_constants.MODEL_LOCATION)
def test_save_with_sample_input_containing_unsupported_data_type_raises_serialization_exception(
spark_context, model_path):
sql_context = SQLContext(spark_context)
unsupported_df = sql_context.createDataFrame([(1, "2016-09-30"), (2, "2017-02-27")])
unsupported_df = unsupported_df.withColumn("_2", unsupported_df._2.cast(DateType()))
pipeline = Pipeline(stages=[])
model = pipeline.fit(unsupported_df)
# The Spark `DateType` is not supported by MLeap, so we expect serialization to fail.
with pytest.raises(mleap.MLeapSerializationException):
sparkm.save_model(spark_model=model, path=model_path, sample_input=unsupported_df)
def test_identity_pipeline(self):
dataset = MockDataset()
def doTransform(pipeline):
pipeline_model = pipeline.fit(dataset)
return pipeline_model.transform(dataset)
# check that empty pipeline did not perform any transformation
self.assertEqual(dataset.index, doTransform(Pipeline(stages=[])).index)
# check that failure to set stages param will raise KeyError for missing param
self.assertRaises(KeyError, lambda: doTransform(Pipeline()))
def MachineLearning(df):
file_dataSVM = "G:/Projects/Spark-Machine-Learning/Spark Machine Learning/Spark Machine Learning/svm/"
data = df.select(['Summary','Sentiment']).withColumnRenamed('Sentiment','label')
data = data.withColumn('length',length(data['Summary']))
# Basic sentence tokenizer
tokenizer = Tokenizer(inputCol="Summary", outputCol="words")
#remove stop words
remover = StopWordsRemover(inputCol="words", outputCol="filtered_features")
#transoform dataset to vectors
cv = HashingTF(inputCol="filtered_features", outputCol="features1", numFeatures=1000)
#calculate IDF for all dataset
idf = IDF(inputCol= 'features1', outputCol = 'tf_idf')
normalizer = StandardScaler(inputCol="tf_idf", outputCol="normFeatures", withStd=True, withMean=False)
selector = ChiSqSelector(numTopFeatures=150, featuresCol="normFeatures",
outputCol="selectedFeatures", labelCol="label")
#prepare data for ML spark library
cleanUp = VectorAssembler(inputCols =['selectedFeatures'],outputCol='features')
# Normalize each Vector using $L^1$ norm.
pipeline = Pipeline(stages=[tokenizer, remover, cv, idf,normalizer,selector,cleanUp])
pipelineModel = pipeline.fit(data)
data = pipelineModel.transform(data)
data.printSchema()
train_data, test_data = data.randomSplit([0.7,0.3],seed=2018)
lr = LogisticRegression(featuresCol="features", labelCol='label')
lrModel = lr.fit(train_data)
beta = np.sort(lrModel.coefficients)
plt.plot(beta)
plt.ylabel('Beta Coefficients')
plt.show()
trainingSummary = lrModel.summary
roc = trainingSummary.roc.toPandas()
plt.plot(roc['FPR'],roc['TPR'])
plt.ylabel('False Positive Rate')
plt.xlabel('True Positive Rate')
plt.title('ROC Curve')
plt.show()
print('Training set areaUnderROC: ' + str(trainingSummary.areaUnderROC))
pr = trainingSummary.pr.toPandas()
plt.plot(pr['recall'],pr['precision'])
plt.ylabel('Precision')
plt.xlabel('Recall')
plt.show()
predictions = lrModel.transform(test_data)
evaluator = BinaryClassificationEvaluator()
print('Test Area Under ROC', evaluator.evaluate(predictions))
def spark_model_iris(iris_df):
feature_names, iris_pandas_df, iris_spark_df = iris_df
assembler = VectorAssembler(inputCols=feature_names, outputCol="features")
lr = LogisticRegression(maxIter=50, regParam=0.1, elasticNetParam=0.8)
pipeline = Pipeline(stages=[assembler, lr])
# Fit the model
model = pipeline.fit(iris_spark_df)
preds_df = model.transform(iris_spark_df)
preds = [x.prediction for x in preds_df.select("prediction").collect()]
return SparkModelWithData(
model=model, spark_df=iris_spark_df, pandas_df=iris_pandas_df, predictions=preds
)
def strat_scatterplot(sdf, col1, col2, n=30):
stages = []
for col in [col1, col2]:
splits = get_buckets(sdf.select(col).rdd.map(itemgetter(0)), n)
stages.append(Bucketizer(splits=splits,
inputCol=col,
outputCol="__{}_bucket".format(col),
handleInvalid="skip"))
pipeline = Pipeline(stages=stages)
model = pipeline.fit(sdf)
return model, sdf.count()
def test_mleap_module_model_save_with_unsupported_transformer_raises_serialization_exception(
spark_model_iris, model_path):
class CustomTransformer(JavaModel):
def _transform(self, dataset):
return dataset
unsupported_pipeline = Pipeline(stages=[CustomTransformer()])
unsupported_model = unsupported_pipeline.fit(spark_model_iris.spark_df)
with pytest.raises(mlflow.mleap.MLeapSerializationException):
mlflow.mleap.save_model(spark_model=unsupported_model,
path=model_path,
sample_input=spark_model_iris.spark_df)
def test_spark_module_model_save_with_mleap_and_unsupported_transformer_raises_exception(
spark_model_iris, model_path):
class CustomTransformer(JavaModel):
def _transform(self, dataset):
return dataset
unsupported_pipeline = Pipeline(stages=[CustomTransformer()])
unsupported_model = unsupported_pipeline.fit(spark_model_iris.spark_df)
with pytest.raises(ValueError):
sparkm.save_model(spark_model=unsupported_model,
path=model_path,
sample_input=spark_model_iris.spark_df)
def oneHotEncoding(clickDF , columns):
"""
ohe = OneHotEncoderEstimator
"""
allStages = [StringIndexer(inputCol=column, outputCol=column+STRING_INDEXER_OUT_SUFFIX).setHandleInvalid("skip") for column in columns]
oneHotEncodeInputOutputNames = [(column+STRING_INDEXER_OUT_SUFFIX , column+ONE_HOT_ENCODER_OUT_SUFFIX) for column in columns]
oneHotEncodeInputOutputNames = list(zip(*oneHotEncodeInputOutputNames))
ohe = OneHotEncoderEstimator(inputCols=oneHotEncodeInputOutputNames[0] , outputCols=oneHotEncodeInputOutputNames[1])
allStages.append(ohe);
pipeline = Pipeline(stages=allStages)
clickDF = pipeline.fit(clickDF).transform(clickDF)
deletedColumns = list(oneHotEncodeInputOutputNames[0])+columns;
return clickDF;
def dataToVectorForTree(clickDF,categoricalColumnsNames , numericColumnNames):
print ("===== Imputing=======")
clickDF , imputedColumnNames = impute(clickDF,numericColumnNames)
print ("===== String Indexer=======")
allStages = [StringIndexer(inputCol=column, outputCol=column+STRING_INDEXER_OUT_SUFFIX).setHandleInvalid("skip") for column in categoricalColumnsNames]
stringIndexderColumnsNames = [(column+STRING_INDEXER_OUT_SUFFIX , column+ONE_HOT_ENCODER_OUT_SUFFIX) for column in categoricalColumnsNames]
stringIndexderColumnsNames = list(zip(*stringIndexderColumnsNames))
pipeline = Pipeline(stages=allStages)
clickDF = pipeline.fit(clickDF).transform(clickDF)
all_feature_columns = imputedColumnNames + list(stringIndexderColumnsNames[0]);
print ("===== Assambler =======")
feature_assembler = VectorAssembler(inputCols=all_feature_columns,outputCol="features")
return feature_assembler.transform(clickDF);
def preprocess(data, data_all, num_ints, num_cats, encoder_pipeline=None):
# Index category strings
string_indexers = [StringIndexer(inputCol=str(col), outputCol='%d_idx' % col)
for col in range(1+num_ints, 1+num_ints+num_cats)]
# One hot encode category features
encoders = [OneHotEncoder(dropLast=True, inputCol='%d_idx' % col, outputCol='%d_cat' % col)
for col in range(1+num_ints, 1+num_ints+num_cats)]
# Build and fit pipeline
if not encoder_pipeline:
encoder_pipeline = Pipeline(stages=string_indexers + encoders).fit(data_all)
results = encoder_pipeline.transform(data)
return results, encoder_pipeline
def test_model_log(tmpdir):
conda_env = os.path.join(str(tmpdir), "conda_env.yml")
_mlflow_conda_env(
conda_env, additional_pip_deps=["pyspark=={}".format(pyspark_version)])
iris = datasets.load_iris()
X = iris.data # we only take the first two features.
y = iris.target
pandas_df = pd.DataFrame(X, columns=iris.feature_names)
pandas_df['label'] = pd.Series(y)
spark_session = pyspark.sql.SparkSession.builder \
.config(key="spark_session.python.worker.reuse", value=True) \
.master("local-cluster[2, 1, 1024]") \
.getOrCreate()
spark_df = spark_session.createDataFrame(pandas_df)
model_path = tmpdir.mkdir("model")
assembler = VectorAssembler(inputCols=iris.feature_names,
outputCol="features")
lr = LogisticRegression(maxIter=50, regParam=0.1, elasticNetParam=0.8)
pipeline = Pipeline(stages=[assembler, lr])
# Fit the model
model = pipeline.fit(spark_df)
# Print the coefficients and intercept for multinomial logistic regression
preds_df = model.transform(spark_df)
preds1 = [x.prediction for x in preds_df.select("prediction").collect()]
old_tracking_uri = tracking.get_tracking_uri()
# should_start_run tests whether or not calling log_model() automatically starts a run.
for should_start_run in [False, True]:
try:
tracking_dir = os.path.abspath(str(tmpdir.mkdir("mlruns")))
tracking.set_tracking_uri("file://%s" % tracking_dir)
if should_start_run:
tracking.start_run()
sparkm.log_model(artifact_path="model", spark_model=model)
run_id = tracking.active_run().info.run_uuid
x = pyfunc.load_pyfunc("model", run_id=run_id)
preds2 = x.predict(pandas_df)
assert preds1 == preds2
reloaded_model = sparkm.load_model("model", run_id=run_id)
preds_df_1 = reloaded_model.transform(spark_df)
preds3 = [
x.prediction
for x in preds_df_1.select("prediction").collect()
]
assert preds1 == preds3
finally:
tracking.end_run()
tracking.set_tracking_uri(old_tracking_uri)
shutil.rmtree(tracking_dir)
def main(spark, df, model_file):
# import metadata
df = spark.read.parquet(df)
print("imported meta data")
# make indexer on all the tracks in metadata
indexer1 = StringIndexer(inputCol="track_id",
outputCol="track_index",
handleInvalid="skip")
pipeline = Pipeline(stages=[indexer1])
model = pipeline.fit(df)
print("mapped user_index")
# output the model indexer
model.write().overwrite().save(model_file)
def test_simple_csv_and_sql_pipeline(spark_session: SparkSession) -> None:
# Arrange
data_dir: Path = Path(__file__).parent.joinpath("./")
flights_path: str = f"file://{data_dir.joinpath('flights.csv')}"
schema = StructType([])
df: DataFrame = spark_session.createDataFrame(
spark_session.sparkContext.emptyRDD(), schema)
spark_session.sql("DROP TABLE IF EXISTS default.flights")
# Act
parameters: Dict[str, Any] = {}
stages: List[Transformer] = create_steps([
FrameworkCsvLoader(view="flights", filepath=flights_path),
FeaturesCarriersV1(parameters=parameters),
])
pipeline: Pipeline = Pipeline(stages=stages) # type: ignore
transformer = pipeline.fit(df)
transformer.transform(df)
# Assert
result_df: DataFrame = spark_session.sql("SELECT * FROM flights2")
result_df.show()
assert result_df.count() > 0
def test_simple_csv_loader_pipeline(spark_session: SparkSession) -> None:
# Arrange
data_dir: Path = Path(__file__).parent.joinpath('./')
flights_path: str = f"file://{data_dir.joinpath('flights.csv')}"
schema = StructType([])
df: DataFrame = spark_session.createDataFrame(
spark_session.sparkContext.emptyRDD(), schema)
# noinspection SqlDialectInspection,SqlNoDataSourceInspection
spark_session.sql("DROP TABLE IF EXISTS default.flights")
# Act
# parameters = Dict[str, Any]({
# })
stages: List[Union[Estimator, Transformer]] = [
FrameworkCsvLoader(
view="flights",
path_to_csv=flights_path
),
SQLTransformer(statement="SELECT * FROM flights"),
]
pipeline: Pipeline = Pipeline(stages=stages)
transformer = pipeline.fit(df)
result_df: DataFrame = transformer.transform(df)
# Assert
result_df.show()
assert result_df.count() > 0
def train_with_tune(input_df):
# https://spark.apache.org/docs/latest/ml-tuning.html
# 构建模型训练流程
lr = LogisticRegression()
pipeline = Pipeline(stages=[lr])
# 构建超参空间
paramGrid = ParamGridBuilder() \
.addGrid(lr.regParam, [0.1, 0.01]) \
.build()
# 只做一次切分
# tvs = TrainValidationSplit(estimator=pipeline,
# estimatorParamMaps=paramGrid,
# evaluator=BinaryClassificationEvaluator(),
# # 80% of the data will be used for training, 20% for validation.
# trainRatio=0.8)
# k-fold cross validation
cross_val = CrossValidator(estimator=pipeline,
estimatorParamMaps=paramGrid,
evaluator=BinaryClassificationEvaluator(),
numFolds=3)
# train and find the best
cvModel = cross_val.fit(input_df)
return cvModel.bestModel
def define_pipeline():
temp_indexer = StringIndexer(inputCol="dtemp", outputCol="idtemp")
nausea_indexer = StringIndexer(inputCol="nausea", outputCol="inausea")
lumbar_indexer = StringIndexer(inputCol="lumbar", outputCol="ilumbar")
urine_indexer = StringIndexer(inputCol="urine", outputCol="iurine")
micturition_indexer = StringIndexer(inputCol="micturition",
outputCol="imicturition")
urethra_indexer = StringIndexer(inputCol="urethra", outputCol="iurethra")
bladder_indexer = StringIndexer(inputCol="bladder", outputCol="label")
vector_assembler = VectorAssembler(inputCols=[
"idtemp", "inausea", "ilumbar", "iurine", "imicturition", "iurethra"
],
outputCol="features")
label_indexer = StringIndexer(inputCol="bladder",
outputCol="label_bladder")
rf = RandomForestClassifier(predictionCol="rf_prediction",
probabilityCol="rf_probability",
rawPredictionCol="rf_rawPrediction",
numTrees=10)
dt = DecisionTreeClassifier(predictionCol="dt_prediction",
probabilityCol="dt_probability",
rawPredictionCol="dt_rawPrediction")
res_pipeline = Pipeline(stages=[
temp_indexer, nausea_indexer, lumbar_indexer, urine_indexer,
micturition_indexer, urethra_indexer, bladder_indexer,
vector_assembler, label_indexer, rf, dt
])
return res_pipeline
def buildModel(data, label):
"""
Build a pipeline to classify `label` against the rest of classes using Binary Regression Classification
:param data: the training data as a DF
:param label: 0..C-1 where C is the number of classes
:param shouldDisplayGraph: True to plot the graph illustrating the classification
:return: the model as a Transformer
"""
logging.info('building model for label = %d, type = %s' %
(label, type(label)))
lr = LogisticRegression()
pipeline = Pipeline(stages=[lr])
paramGrid = ParamGridBuilder()\
.addGrid(lr.maxIter, [100])\
.addGrid(lr.elasticNetParam, [0.0, 1.0])\
.addGrid(lr.fitIntercept, [True, False])\
.build()
crossValidator = CrossValidator(estimator=pipeline,
estimatorParamMaps=paramGrid,
evaluator=BinaryClassificationEvaluator(),
numFolds=15)
dataDF = data.map(lambda point: LabeledPoint(
0 if point.label == label else 1, point.features)).toDF()
model = crossValidator.fit(dataDF)
return model
def strat_scatterplot(sdf, col1, col2, n=30):
stages = []
for col in [col1, col2]:
splits = np.linspace(
*sdf.agg(F.min(col), F.max(col)).rdd.map(tuple).collect()[0],
n + 1)
bucket_name = '__{}_bucket'.format(col)
stages.append(
Bucketizer(splits=splits,
inputCol=col,
outputCol=bucket_name,
handleInvalid="skip"))
pipeline = Pipeline(stages=stages)
model = pipeline.fit(sdf)
return model, sdf.count()
def test_get_metrics_by_threshold(sdf):
assem = VectorAssembler(inputCols=['Fare', 'Pclass', 'Age'],
outputCol='features')
rf = RandomForestClassifier(featuresCol='features',
labelCol='Survived',
numTrees=20,
seed=13)
pipeline = Pipeline(stages=[assem, rf])
model = pipeline.fit(sdf.fillna(0.0))
predictions = model.transform(sdf.fillna(0.0)).select(
'probability', 'Survived')
bcm = BinaryClassificationMetrics(predictions,
scoreCol='probability',
labelCol='Survived')
metrics = bcm.getMetricsByThreshold()
predictions = predictions.toHandy().to_metrics_RDD('probability',
'Survived')
predictions = np.array(predictions.collect())
pr = np.array(bcm.pr().collect())
idx = pr[:, 0].argmax()
pr = pr[:idx + 1, :]
precision, recall, thresholds = precision_recall_curve(
predictions[:, 1], predictions[:, 0])
npt.assert_array_almost_equal(precision, pr[:, 1][::-1])
npt.assert_array_almost_equal(recall, pr[:, 0][::-1])
roc = np.array(bcm.roc().collect())
idx = roc[:, 1].argmax()
roc = roc[:idx + 1, :]
sroc = pd.DataFrame(np.round(roc, 6), columns=['fpr', 'tpr'])
sroc = sroc.groupby('fpr').agg({'tpr': [np.min, np.max]})
fpr, tpr, thresholds = roc_curve(predictions[:, 1], predictions[:, 0])
idx = tpr.argmax()
proc = pd.DataFrame({
'fpr': np.round(fpr[:idx + 1], 6),
'tpr': np.round(tpr[:idx + 1], 6)
})
proc = proc.groupby('fpr').agg({'tpr': [np.min, np.max]})
sroc = sroc.join(proc, how='inner', rsuffix='sk')
npt.assert_array_almost_equal(sroc.iloc[:, 0], proc.iloc[:, 0])
npt.assert_array_almost_equal(sroc.iloc[:, 1], proc.iloc[:, 1])
def main(input, model_file):
tmax_schema = types.StructType([
types.StructField('station', types.StringType()),
types.StructField('date', types.DateType()),
types.StructField('latitude', types.FloatType()),
types.StructField('longitude', types.FloatType()),
types.StructField('elevation', types.FloatType()),
types.StructField('tmax', types.FloatType()),
])
data = spark.read.csv(input, schema=tmax_schema)
train, validation = data.randomSplit([0.75, 0.25], seed=123)
train = train.cache()
validation = validation.cache()
y_tmax = SQLTransformer(
statement=
"SELECT today.station,today.latitude,today.longitude,today.elevation,today.date,today.tmax,yesterday.tmax AS yesterday_tmax FROM __THIS__ as today INNER JOIN __THIS__ as yesterday ON date_sub(today.date, 1) = yesterday.date AND today.station = yesterday.station"
)
getvalues = SQLTransformer(
statement=
"SELECT station,latitude,longitude,elevation,dayofyear(date) AS dayofyear,tmax,yesterday_tmax from __THIS__"
)
assemble_features = VectorAssembler(inputCols=[
'latitude', 'longitude', 'elevation', 'dayofyear', 'yesterday_tmax'
],
outputCol='features')
classifier = GBTRegressor(featuresCol='features', labelCol='tmax')
pipeline = Pipeline(
stages=[y_tmax, getvalues, assemble_features, classifier])
model = pipeline.fit(train)
predictions = model.transform(validation)
r2_evaluator = RegressionEvaluator(predictionCol='prediction',
labelCol='tmax',
metricName='r2')
r2 = r2_evaluator.evaluate(predictions)
print('-----------------------------------')
print('r2: %g' % (r2, ))
print('-----------------------------------')
rmse_evaluator = RegressionEvaluator(predictionCol='prediction',
labelCol='tmax',
metricName='rmse')
rmse = rmse_evaluator.evaluate(predictions)
print('rmse: %g' % (rmse, ))
model.write().overwrite().save(model_file)
def simple_train_model(input_df):
xgboost_params = {
"eta": 0.023,
"max_depth": 10,
"min_child_weight": 0.3,
"subsample": 0.7,
"colsample_bytree": 0.82,
"colsample_bylevel": 0.9,
"eval_metric": "auc",
"seed": 49,
"silent": 1,
"objective": "binary:logistic",
"round": 10,
"nWorkers": 2
}
xgb_model = XGBoostClassifier(xgboost_params)
pipeline = Pipeline(stages=[xgb_model])
return pipeline.fit(input_df)
def test_pipeline(self):
dataset = MockDataset()
estimator0 = MockEstimator()
transformer1 = MockTransformer()
estimator2 = MockEstimator()
transformer3 = MockTransformer()
pipeline = Pipeline(stages=[estimator0, transformer1, estimator2, transformer3])
pipeline_model = pipeline.fit(dataset, {estimator0.fake: 0, transformer1.fake: 1})
model0, transformer1, model2, transformer3 = pipeline_model.stages
self.assertEqual(0, model0.dataset_index)
self.assertEqual(0, model0.getFake())
self.assertEqual(1, transformer1.dataset_index)
self.assertEqual(1, transformer1.getFake())
self.assertEqual(2, dataset.index)
self.assertIsNone(model2.dataset_index, "The last model shouldn't be called in fit.")
self.assertIsNone(transformer3.dataset_index,
"The last transformer shouldn't be called in fit.")
dataset = pipeline_model.transform(dataset)
self.assertEqual(2, model0.dataset_index)
self.assertEqual(3, transformer1.dataset_index)
self.assertEqual(4, model2.dataset_index)
self.assertEqual(5, transformer3.dataset_index)
self.assertEqual(6, dataset.index)
d1.printSchema()
d2 = d1.toDF("number", "name", "SI", "GOO", "DONG", "x", "y", "b_code", "h_code", "utmk_x", "utmk_y", "wtm_x", "wtm_y")
d3 = d2.select(d2.GOO.alias("loc"), d2.x, d2.y)
d3.show(5, False)
indexer = StringIndexer(inputCol="loc", outputCol="loccode")
assembler = VectorAssembler(inputCols=["loccode", "x", "y"], outputCol="features")
kmeans = KMeans(k=5, seed=1, featuresCol="features")
pipeline = Pipeline(stages=[indexer, assembler, kmeans])
model = pipeline.fit(d3)
d4 = model.transform(d3)
d4.groupBy("prediction") \
.agg(functions.collect_set("loc").alias("loc")) \
.orderBy("prediction").show(100, False)
WSSSE = model.stages[2].computeCost(d4)
print("Within Set Sum of Squared Errors = %d" % WSSSE)
print("Cluster Centers: ")
for v in model.stages[2].clusterCenters():
print(v)
.setOutputCol("widthFeatures")
.setSplits(splits))
irisBucketizedWidth = widthBucketizer.transform(irisBucketizedLength)
display(irisBucketizedWidth)
# COMMAND ----------
# MAGIC %md
# MAGIC Let's combine the two bucketizers into a [Pipeline](http://spark.apache.org/docs/latest/ml-guide.html#pipeline-components) that performs both bucketizations. A `Pipeline` is make up of stages which can be set using `setStages` and passing in a `list` of stages in Python or an `Array` of stages in `Scala`. `Pipeline` is an estimator, which means it implements a `fit` method which returns a `PipelineModel`. A `PipelineModel` is a transformer, which means that it implements a `transform` method which can be used to run the stages.
# COMMAND ----------
from pyspark.ml.pipeline import Pipeline
pipelineBucketizer = Pipeline().setStages([lengthBucketizer, widthBucketizer])
pipelineModelBucketizer = pipelineBucketizer.fit(irisSeparateFeatures)
irisBucketized = pipelineModelBucketizer.transform(irisSeparateFeatures)
display(irisBucketized)
# COMMAND ----------
# MAGIC %md
# MAGIC Now that we have created two new features through bucketing, let's combined those two features into a `Vector` with `VectorAssembler`.
# MAGIC
# MAGIC Set the params of `assembler` so that both "lengthFeatures" and "widthFeatures" are assembled into a column called "featuresBucketized".
# MAGIC
# MAGIC Then, set the stages of `pipeline` to include both bucketizers and the assembler as the last stage.
firstMlModel = firstMlKMeans.fit(ca1mlFeaturizedDF)
type(firstMlModel)
# In[65]:
firstMlModel.clusterCenters()
# Sudarome `Pipeline` žingsnių seką iš `vecAssembler` ir `kmeans` komponentų.
# In[66]:
from pyspark.ml.pipeline import Pipeline
firstPipeline = Pipeline(stages=[vecAssembler, firstMlKMeans])
# In[67]:
firstPipelineModel = firstPipeline.fit(ca1MlDF)
# In[73]:
firstPipelineModel.transform(ca1MlDF).show(5)
# In[74]:
MlKmeansWSSSEResults = collections.namedtuple("MlKmeansWSSSEResults", ["ks", "WSSSEs", "pipelineModels"])
.setOutputCol("widthFeatures")
.setSplits(splits))
irisBucketizedWidth = widthBucketizer.transform(irisBucketizedLength)
display(irisBucketizedWidth)
# COMMAND ----------
# MAGIC %md
# MAGIC Let's combine the two bucketizers into a [Pipeline](http://spark.apache.org/docs/latest/ml-guide.html#pipeline-components) that performs both bucketizations. A `Pipeline` is made up of stages which can be set using `setStages` and passing in a `list` of stages in Python or an `Array` of stages in Scala. `Pipeline` is an estimator, which means it implements a `fit` method which returns a `PipelineModel`. A `PipelineModel` is a transformer, which means that it implements a `transform` method which can be used to run the stages.
# COMMAND ----------
from pyspark.ml.pipeline import Pipeline
pipelineBucketizer = Pipeline().setStages([lengthBucketizer, widthBucketizer])
pipelineModelBucketizer = pipelineBucketizer.fit(irisSeparateFeatures)
irisBucketized = pipelineModelBucketizer.transform(irisSeparateFeatures)
display(irisBucketized)
# COMMAND ----------
# MAGIC %md
# MAGIC Now that we have created two new features through bucketing, let's combine those two features into a `Vector` with `VectorAssembler`. VectorAssembler can be found in [pyspark.ml.feature](https://spark.apache.org/docs/latest/api/python/pyspark.ml.html#pyspark.ml.feature.VectorAssembler) for Python and the [org.apache.spark.ml.feature](http://spark.apache.org/docs/latest/api/scala/#org.apache.spark.ml.feature.VectorAssembler) package for Scala.
# MAGIC
# MAGIC Set the params of `assembler` so that both "lengthFeatures" and "widthFeatures" are assembled into a column called "featuresBucketized".
# MAGIC
# MAGIC Then, set the stages of `pipeline` to include both bucketizers and the assembler as the last stage.
def simple_train_model(input_df):
lr_model = LogisticRegression(regParam=0.01)
pipeline = Pipeline(stages=[lr_model])
return pipeline.fit(input_df)
from pyspark.ml.regression import LinearRegression
lr = (LinearRegression()
.setLabelCol('sepalWidth')
.setMaxIter(1000))
print lr.explainParams()
# COMMAND ----------
# MAGIC %md
# MAGIC Next, we'll create a `Pipeline` that only contains one stage for the linear regression.
# COMMAND ----------
from pyspark.ml.pipeline import Pipeline
pipeline = Pipeline().setStages([lr])
pipelineModel = pipeline.fit(irisSepalSample)
sepalPredictions = pipelineModel.transform(irisSepalSample)
display(sepalPredictions)
# COMMAND ----------
# MAGIC %md
# MAGIC What does our resulting model look like?
# COMMAND ----------
lrModel = pipelineModel.stages[-1]
print type(lrModel)
evaluator = RegressionEvaluator(labelCol="weight", predictionCol="predic_weight")
# root mean squared error
rmse = evaluator.evaluate(d13)
# mean squared error
mse = evaluator.setMetricName("mse").evaluate(d13)
# R2 metric
r2 = evaluator.setMetricName("r2").evaluate(d13)
# mean absolute error
mae = evaluator.setMetricName("mae").evaluate(d13)
print("rmse:%d, mse:%d, r2:%d, mae:%d" % (rmse, mse, r2, mae))
# 파이프라인
pipeline = Pipeline(stages=[gradeIndexer, genderIndexer, assembler, lr])
samples2 = df9.randomSplit([0.7, 0.3])
training2 = samples2[0]
test2 = samples2[1]
# 파이프라인 모델 생성
pipelineModel = pipeline.fit(training2)
# 파이프라인 모델을 이용한 예측값 생성
pipelineModel.transform(test2).show(5, False)
spark.stop
d6 = d5.withColumn("label", label(d5.avr_velo_month, d5.avr_velo_total).cast("double"))
d6.select("road", "avr_velo_month", "avr_velo_total", "label").show(5, False)
d6.groupBy("label").count().show(truncate=False)
dataArr = d6.randomSplit([0.7, 0.3])
train = dataArr[0]
test = dataArr[1]
indexer = StringIndexer(inputCol="road", outputCol="roadcode")
assembler = VectorAssembler(inputCols=["roadcode", "mon", "tue", "wed", "thu", "fri", "sat", "sun"],
outputCol="features")
dt = DecisionTreeClassifier(labelCol="label", featuresCol="features")
pipeline = Pipeline(stages=[indexer, assembler, dt])
model = pipeline.fit(train)
predict = model.transform(test)
predict.select("label", "probability", "prediction").show(3, False)
# areaUnderROC, areaUnderPR
evaluator = BinaryClassificationEvaluator(labelCol="label", metricName="areaUnderROC")
print(evaluator.evaluate(predict))
treeModel = model.stages[2]
print("Learned classification tree model:%s" % treeModel.toDebugString)
.setOutputCol("widthFeatures")
.setSplits(splits))
irisBucketizedWidth = widthBucketizer.transform(irisBucketizedLength)
display(irisBucketizedWidth)
# COMMAND ----------
# MAGIC %md
# MAGIC Let's combine the two bucketizers into a [Pipeline](http://spark.apache.org/docs/latest/ml-guide.html#pipeline-components) that performs both bucketizations. A `Pipeline` is make up of stages which can be set using `setStages` and passing in a `list` of stages in Python or an `Array` of stages in `Scala`. `Pipeline` is an estimator, which means it implements a `fit` method which returns a `PipelineModel`. A `PipelineModel` is a transformer, which means that it implements a `transform` method which can be used to run the stages.
# COMMAND ----------
from pyspark.ml.pipeline import Pipeline
pipelineBucketizer = Pipeline().setStages([lengthBucketizer, widthBucketizer])
pipelineModelBucketizer = pipelineBucketizer.fit(irisSeparateFeatures)
irisBucketized = pipelineModelBucketizer.transform(irisSeparateFeatures)
display(irisBucketized)
# COMMAND ----------
# MAGIC %md
# MAGIC Now that we have created two new features through bucketing, let's combined those two features into a `Vector` with `VectorAssembler`.
# MAGIC
# MAGIC Set the params of `assembler` so that both "lengthFeatures" and "widthFeatures" are assembled into a column called "featuresBucketized".
# MAGIC
# MAGIC Then, set the stages of `pipeline` to include both bucketizers and the assembler as the last stage.
# COMMAND ---------- (rf .setMaxBins(10) .setMaxDepth(2) .setNumTrees(20) .setSeed(0)) # COMMAND ---------- # MAGIC %md # MAGIC Next, we'll build a pipeline that includes the `StringIndexer`, `PolynomialExpansion`, and `RandomForestClassifier`. # COMMAND ---------- rfPipeline = Pipeline().setStages([stringIndexer, px, rf]) rfModelPipeline = rfPipeline.fit(irisTrain) rfPredictions = rfModelPipeline.transform(irisTest) print multiEval.evaluate(rfPredictions) # COMMAND ---------- display(rfPredictions) # COMMAND ---------- # MAGIC %md # MAGIC So what exactly did `PolynomialExpansion` do? # COMMAND ----------
