def process(spark, train_data, test_data):
df_train = spark.read.parquet(train_data)
df_test = spark.read.parquet(test_data)
features = VectorAssembler(inputCols=df_train.columns[1:-1],
outputCol='features')
evaluator = RegressionEvaluator(labelCol='ctr',
predictionCol='prediction',
metricName='rmse')
lr_model_base = LinearRegression(labelCol='ctr', **LR_PARAMS_BASE)
lr_model_to_tune = LinearRegression(labelCol='ctr')
lr_param_grid = ParamGridBuilder() \
.addGrid(lr_model_to_tune.maxIter, [5, 10, 20, 40, 50]) \
.addGrid(lr_model_to_tune.regParam, [0.4, 0.1, 0.01, 0.001]) \
.addGrid(lr_model_to_tune.fitIntercept, [False, True]) \
.addGrid(lr_model_to_tune.elasticNetParam, [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]) \
.build()
tvs = TrainValidationSplit(estimator=lr_model_to_tune,
estimatorParamMaps=lr_param_grid,
evaluator=evaluator,
trainRatio=0.8)
pipeline_model_base = Pipeline(
stages=[features, lr_model_base]).fit(df_train)
prediction_base = pipeline_model_base.transform(df_test)
rmse_base = evaluator.evaluate(prediction_base)
print(f'Base lr model params: {LR_PARAMS_BASE}')
print(f'RMSE at base lr model = {rmse_base}')
print('Tuning lr model...')
pipeline_model_tuned = Pipeline(stages=[features, tvs]).fit(df_train)
prediction_tuned = pipeline_model_tuned.transform(df_test)
rmse_tuned = evaluator.evaluate(prediction_tuned)
model_java_obj = pipeline_model_tuned.stages[-1].bestModel._java_obj
lr_params_tuned = {
'maxIter': model_java_obj.getMaxIter(),
'regParam': model_java_obj.getRegParam(),
'elasticNetParam': model_java_obj.getElasticNetParam(),
'fitIntercept': model_java_obj.getFitIntercept()
}
print(f'Base lr model params: {lr_params_tuned}')
print(f'RMSE at tuned lr model = {rmse_tuned}')
if rmse_tuned < rmse_base:
pipeline_model_tuned.write().overwrite().save(MODEL_PATH)
print(f'Tuned model has better RMSE value')
else:
pipeline_model_base.write().overwrite().save(MODEL_PATH)
print(f'Base model has better RMSE value')
print(f'Model saved at "{MODEL_PATH}"')
spark.stop()
def transform_data_in_pipeline(df):
"""
:param df:
:return:
"""
# Initialise pipeline variables
stages = []
assembler_inputs = []
# Assemble features vector from Spark dataframe fields
assembler = VectorAssembler(
inputCols=['x', 'y', 'star_rating_number', 'avg_adr'],
outputCol='features')
stages += [assembler]
assembler_inputs += [assembler.getOutputCol()]
# Apply standard scaling with unit std and centroid about the mean
scaler = StandardScaler(inputCol=assembler.getOutputCol(),
outputCol='scaledFeatures')
stages += [scaler]
assembler_inputs += [scaler.getOutputCol()]
# Execute the pipeline
pipeline_model = Pipeline() \
.setStages(stages) \
.fit(df)
# Return the dataframe with the additional transformed features vector
return pipeline_model.transform(df)
def date_conversion():
df = sql.read.csv("./run/date_test_res.csv", inferSchema=True, header=True)
datetime_formatting = DatetimeFormatting()
model = Pipeline(stages=[datetime_formatting]).fit(df)
res = model.transform(df)
print("resulted_df")
print(res.show())
def remove_skewness():
df = sql.read.csv("./run/file1.csv", inferSchema=True, header=True)
min_skewness = MinimizeSkewness(['Purpose'])
model = Pipeline(stages=[min_skewness]).fit(df)
res = model.transform(df)
print(res.show())
def process_df(df):
time_seq.append(['start process-df', time.time()])
model = Pipeline(stages=[
RegexTokenizer(pattern=" ",
inputCol="instruments",
outputCol="instruments_tokenized",
minTokenLength=1),
NGram(n=1,
inputCol="instruments_tokenized",
outputCol="instruments_ngrams"),
HashingTF(inputCol="instruments_ngrams",
outputCol="instruments_vectors"),
MinHashLSH(inputCol="instruments_vectors",
outputCol="instruments_lsh",
numHashTables=10)
]).fit(df)
df_hashed = model.transform(df)
df_matches = model.stages[-1].approxSimilarityJoin(df_hashed, df_hashed, 0.5, distCol="distance") \
.filter("datasetA.filename != datasetB.filename AND datasetA.filename < datasetB.filename") \
.select(f.col('datasetA.filename').alias('filename_A'),
f.col('datasetB.filename').alias('filename_B'),
f.col('distance'))
time_seq.append(['process-df df_matches', time.time()])
write_df_to_pgsql(df_matches, 'filepair_similarity_run3')
time_seq.append(['write pgsql', time.time()])
print('time_seq', time_seq)
def main(sc, spark):
# Load and vectorize the corpus
corpus = load_corpus(sc, spark)
vector = make_vectorizer().fit(corpus)
# Index the labels of the classification
labelIndex = StringIndexer(inputCol="label", outputCol="indexedLabel")
labelIndex = labelIndex.fit(corpus)
# Split the data into training and test sets
training, test = corpus.randomSplit([0.8, 0.2])
# Create the classifier
clf = LogisticRegression(
maxIter=10, regParam=0.3, elasticNetParam=0.8,
family="multinomial", labelCol="indexedLabel", featuresCol="tfidf")
# Create the model
model = Pipeline(stages=[
vector, labelIndex, clf
]).fit(training)
# Make predictions
predictions = model.transform(test)
predictions.select("prediction", "indexedLabel", "tfidf").show(5)
# Select (prediction, true label) and compute test error
evaluator = MulticlassClassificationEvaluator(
labelCol="indexedLabel", predictionCol="prediction", metricName="accuracy")
accuracy = evaluator.evaluate(predictions)
print("Test Error = %g" % (1.0 - accuracy))
gbtModel = model.stages[2]
print(gbtModel) # summary only
def train_als(ratings_data, split_prop, max_iter, reg_param, rank, cold_start_strategy):
seed = 42
spark = pyspark.sql.SparkSession.builder.getOrCreate()
ratings_df = spark.read.parquet(ratings_data)
(training_df, test_df) = ratings_df.randomSplit([split_prop, 1 - split_prop], seed=seed)
training_df.cache()
test_df.cache()
mlflow.log_metric("training_nrows", training_df.count())
mlflow.log_metric("test_nrows", test_df.count())
print("Training: {0}, test: {1}".format(training_df.count(), test_df.count()))
als = (
ALS()
.setUserCol("userId")
.setItemCol("movieId")
.setRatingCol("rating")
.setPredictionCol("predictions")
.setMaxIter(max_iter)
.setSeed(seed)
.setRegParam(reg_param)
.setColdStartStrategy(cold_start_strategy)
.setRank(rank)
)
als_model = Pipeline(stages=[als]).fit(training_df)
reg_eval = RegressionEvaluator(predictionCol="predictions", labelCol="rating", metricName="mse")
predicted_test_dF = als_model.transform(test_df)
test_mse = reg_eval.evaluate(predicted_test_dF)
train_mse = reg_eval.evaluate(als_model.transform(training_df))
print("The model had a MSE on the test set of {0}".format(test_mse))
print("The model had a MSE on the (train) set of {0}".format(train_mse))
mlflow.log_metric("test_mse", test_mse)
mlflow.log_metric("train_mse", train_mse)
mlflow.spark.log_model(als_model, "als-model")
def remove_url_duplication():
df = sql.read.csv("./run/date_test_res.csv", inferSchema=True, header=True)
url_duplication = RemovingDuplicationUrl()
model = Pipeline(stages=[url_duplication]).fit(df)
result = model.transform(df)
result.toPandas().to_csv('./run/pipeline_url.csv')
print("resulted_df")
print(result.show())
def main():
input_dataset = sys.argv[1]
output_dir = sys.argv[2]
start_time = time.time()
#stackoverflow_df = sqlContext.read.csv("../Datasource/stackOverFlow_ID_Title_SMALL.csv", header=True).toDF('id', 'text')
stackoverflow_df = sqlContext.read.csv(input_dataset,
header=True).toDF('id', 'text')
# stackoverflow_df.show()
# stackoverflow_df.head(10).show()
# stack_df = stack_rdd.toDF(['id','text'])
# stackoverflow_df.show()
# stackoverflow_df.printSchema()
model = Pipeline(stages=[
RegexTokenizer(
pattern="", inputCol="text", outputCol="tokens", minTokenLength=1),
NGram(n=3, inputCol="tokens", outputCol="ngrams"),
HashingTF(inputCol="ngrams", outputCol="vectors"),
MinHashLSH(
inputCol="vectors", outputCol="lsh"
) #MinHashLSH(inputCol="vectors", outputCol="lsh", numHashTables=5)
]).fit(stackoverflow_df)
db_hashed = model.transform(stackoverflow_df)
# db_hashed.show()
# query_hashed = model.transform(query)
# db_hashed.show()
# query_hashed.show()
#res = model.stages[-1].approxSimilarityJoin(db_hashed, db_hashed, 0.90).filter("datasetA.id < datasetB.id")
res = model.stages[-1].approxSimilarityJoin(db_hashed, db_hashed,
0.70).filter("distCol > 0")
#print res
#print res.count()
res.show()
elapsed_time = time.time() - start_time
print 'Elapsed Time ==> ', elapsed_time
def model(self, pandas, column):
rdd = self.sqlctx.createDataFrame(pandas.astype(str))
model = Pipeline(stages=[
Tokenizer(inputCol=column, outputCol="tokens"),
StopWordsRemover(inputCol='tokens',
outputCol="tokens_stop",
stopWords=self.STOP_WORDS),
HashingTF(inputCol="tokens_stop", outputCol="vectors")
]).fit(rdd)
db_1 = model.transform(rdd)
db_1.cache()
return db_1
def test_serialize_to_bundle(self):
string_map = StringMap(
labels={'a': 1.0},
inputCol='key_col',
outputCol='value_col',
)
pipeline = Pipeline(stages=[string_map]).fit(self.input)
serialization_dataset = pipeline.transform(self.input)
jar_file_path = _serialize_to_file(pipeline, serialization_dataset)
deserialized_pipeline = _deserialize_from_file(jar_file_path)
result = deserialized_pipeline.transform(self.input)
expected = StringMapTest.spark.createDataFrame([['a', 'b', 1.0]], OUTPUT_SCHEMA)
assert_df(expected, result)
def columns_same_value():
try:
df = sql.read.csv("./run/rem_test.csv", inferSchema=True, header=True)
columns_with_same_val = ColumnsDroppingSameValue()
model = Pipeline(stages=[columns_with_same_val]).fit(df)
result = model.transform(df)
result.toPandas().to_csv('./run/pipeline_same_value.csv')
print(df.show())
print("#####################")
print("resulted_df")
print(result.show())
except Exception as e:
logger.error(e)
def Indexer(spark,
train_address,
val_address,
tst_address,
repartition_size=10000):
df_train = spark.read.parquet(train_address)
df_val = spark.read.parquet(val_address)
df_test = spark.read.parquet(tst_address)
# user_indexer = StringIndexer(inputCol="user_id", outputCol="user_id_numeric").fit(df_train)
# track_indexer = StringIndexer(inputCol="track_id", outputCol="track_id_numeric").fit(df_train.union(df_val))
# df_train = user_indexer.transform(df_train)
# df_train = track_indexer.transform(df_train)
# df_val = user_indexer.transform(df_val)
# df_val = track_indexer.transform(df_val)
# df_train = df_train.select("user_id_numeric","track_id_numeric","count")
# df_val = df_val.select("user_id_numeric","track_id_numeric","count")
# df_test = user_indexer.transform(df_test)
# df_test = track_indexer.transform(df_test)
user_indexer = StringIndexer(inputCol="user_id",
outputCol="user_id_numeric")
track_indexer = StringIndexer(inputCol="track_id",
outputCol="track_id_numeric")
# fit df_train only
model = Pipeline(stages=[user_indexer, track_indexer]).fit(df_train)
df_train, df_val, df_test = [
model.transform(x) for x in (df_train, df_val, df_test)
]
df_train = df_train.select("user_id_numeric", "track_id_numeric", "count")
df_val = df_val.select("user_id_numeric", "track_id_numeric", "count")
df_test = df_test.select("user_id_numeric", "track_id_numeric", "count")
# df_train = df_train.repartition(repartition_size,"user_id_numeric","track_id_numeric")
# df_val = df_val.repartition(repartition_size,"user_id_numeric","track_id_numeric")
# df_test = df_test.repartition(repartition_size,"user_id_numeric","track_id_numeric")
# df_train.write.parquet("./train_formatted.parquet", mode='overwrite')
# df_val.write.parquet("./val_formatted.parquet", mode='overwrite')
# df_test.write.parquet("./test_formatted.parquet", mode='overwrite')
print('Indexer succeed.')
return df_train, df_val, df_test
def main():
potential_clones = sys.argv[1]
outDir = sys.argv[2]
start_time = time.time()
potential_clones = '../Datasource/pc.xml'
output_csv = 'csvCodes.csv'
df = convertAndSaveAsCSV(potential_clones, output_csv, True)
# spark context
sc = SparkContext.getOrCreate()
sqlContext = SQLContext(sc)
spark_df = sqlContext.createDataFrame(df)
transformed_spark_df = spark_df.rdd.map(distributedSourceTransform)
pysparkdf_transformedClones = transformed_spark_df.toDF(
['filepath', 'startline', 'endline', 'source'])
#pysparkdf_transformedClones.show()
model = Pipeline(stages=[
RegexTokenizer(pattern=" ",
inputCol="source",
outputCol="tokens",
minTokenLength=1),
NGram(n=3, inputCol="tokens", outputCol="ngrams"),
HashingTF(inputCol="ngrams", outputCol="vectors", numFeatures=262144),
MinHashLSH(
inputCol="vectors", outputCol="lsh", numHashTables=105
) #MinHashLSH(inputCol="vectors", outputCol="lsh", numHashTables=5)
]).fit(pysparkdf_transformedClones)
hashed_clones = model.transform(pysparkdf_transformedClones)
clone_pairs = model.stages[-1].approxSimilarityJoin(
hashed_clones, hashed_clones, 0.70).filter("distCol > 0")
clone_pairs.show()
elapsed_time = time.time() - start_time
print 'Elapsed Time ==> ', elapsed_time
def main(argv):
spark = SparkSession.builder \
.master("local[*]") \
.config("spark.driver.memory", "4g") \
.config("spark.executor.memory", "1g") \
.getOrCreate()
features_df = ParquetDataFrame(
f'data/processed/{Phase.train.name}/features', spark)
test_data_frac = 0.1
test_features_df, train_features_df = features_df.randomSplit(
[test_data_frac, 1 - test_data_frac])
label_col = 'duration_min'
model = Pipeline(stages=[
StringIndexer(inputCol='pickup_cell_6',
handleInvalid='keep',
outputCol='pickup_cell_6_idx'),
StringIndexer(inputCol='dropoff_cell_6',
handleInvalid='keep',
outputCol='dropoff_cell_6_idx'),
VectorAssembler(inputCols=[
'pickup_cell_6_idx', 'dropoff_cell_6_idx', 'distance', 'month',
'day_of_month', 'day_of_week', 'hour', 'requests_pickup_cell',
'requests_dropoff_cell'
],
outputCol="features"),
DecisionTreeRegressor(
maxDepth=7, featuresCol='features', labelCol=label_col)
]).fit(train_features_df)
model_path = 'model/trip_duration_min'
print(f'Saving model to {model_path}')
model.write().overwrite().save(model_path)
print(f'Model saved...')
model = PipelineModel.load(model_path)
predictions_df = model.transform(test_features_df)
mae_cv = RegressionEvaluator(labelCol=label_col,
metricName='mae').evaluate(predictions_df)
print(f'Mean absolutre error: {mae_cv}')
spark.stop()
def remove_cols_containing_nan():
try:
logger.debug("this is debug")
df = sql.read.csv("./run/column_rem.csv",
inferSchema=True,
header=True)
col_contains_nan = ColumnsDroppingContainsNan()
model = Pipeline(stages=[col_contains_nan]).fit(df)
result = model.transform(df)
result.toPandas().to_csv('./run/pipeline_nan_value.csv')
print(df.show())
print("#####################")
print("resulted_df")
print(result.show())
except Exception as e:
logger.error(e)
def main(sc, spark):
# Load and vectorize the corpus
corpus = load_corpus(sc, spark)
vector = make_vectorizer().fit(corpus)
# Index the labels of the classification
labelIndex = StringIndexer(inputCol="label", outputCol="indexedLabel")
labelIndex = labelIndex.fit(corpus)
# Split the data into training and test sets
training, test = corpus.randomSplit([0.8, 0.2])
# Create the classifier
clf = LogisticRegression(maxIter=10,
regParam=0.3,
elasticNetParam=0.8,
family="multinomial",
labelCol="indexedLabel",
featuresCol="tfidf")
# Create the model
model = Pipeline(stages=[vector, labelIndex, clf]).fit(training)
# Make predictions
predictions = model.transform(test)
predictions.select("prediction", "indexedLabel", "tfidf").show(5)
# Select (prediction, true label) and compute test error
evaluator = MulticlassClassificationEvaluator(labelCol="indexedLabel",
predictionCol="prediction",
metricName="accuracy")
accuracy = evaluator.evaluate(predictions)
print("Test Error = %g" % (1.0 - accuracy))
gbtModel = model.stages[2]
print(gbtModel) # summary only
sw_filter = StopWordsRemover()\
.setStopWords(stop_words)\
.setCaseSensitive(False)\
.setInputCol("words")\
.setOutputCol("filtered")
from pyspark.ml.feature import CountVectorizer
# we will remove words that appear in 5 docs or less
cv = CountVectorizer(minTF=1., minDF=5., vocabSize=2**17)\
.setInputCol("filtered")\
.setOutputCol("tf")
# we now create a pipelined transformer
cv_pipeline = Pipeline(stages=[tokenizer, sw_filter, cv]).fit(review)
cv_pipeline.transform(review).show(5)
from pyspark.ml.feature import IDF
idf = IDF().\
setInputCol('tf').\
setOutputCol('tfidf')
idf_pipeline = Pipeline(stages=[cv_pipeline, idf]).fit(review)
tfidf_df = idf_pipeline.transform(review)
tfidf_df.show(10)
#training_df, validation_df, testing_df = review.randomSplit([0.6, 0.3, 0.1], seed=0)
#training_df, validation_df, testing_df = review.randomSplit([0.6, 0.3, 0.1], seed=0)
#[training_df.count(), validation_df.count(), testing_df.count()]
# A linear regression object regression = LinearRegression(labelCol='duration') -------------------------------------------------- # Exercise_2 # Import class for creating a pipeline from pyspark.ml import Pipeline # Construct a pipeline pipeline = Pipeline(stages=[indexer, onehot, assembler, regression]) # Train the pipeline on the training data pipeline = pipeline.fit(flights_train) # Make predictions on the testing data predictions = pipeline.transform(flights_test) -------------------------------------------------- # Exercise_3 from pyspark.ml.feature import Tokenizer, StopWordsRemover, HashingTF, IDF # Break text into tokens at non-word characters tokenizer = Tokenizer(inputCol='text', outputCol='words') # Remove stop words remover = StopWordsRemover(inputCol=tokenizer.getOutputCol(), outputCol='terms') # Apply the hashing trick and transform to TF-IDF hasher = HashingTF(inputCol=remover.getOutputCol(), outputCol="hash") idf = IDF(inputCol=hasher.getOutputCol(), outputCol="features")
def Indexer(spark, train_address, val_address, tst_address):
beg = time()
df_train = spark.read.parquet(train_address)
df_val = spark.read.parquet(val_address)
df_test = spark.read.parquet(tst_address)
print('File Reading Finished')
# subsample
# here we include the last 110K users (for validation/testing)
# want to include all 110K users at the end and randomly draw 10% others
subsample_frac = 0.1
all_user_ids = [
row['user_id']
for row in df_train.select('user_id').distinct().collect()
]
val_user_ids = [
row['user_id']
for row in df_val.select('user_id').distinct().collect()
]
test_user_ids = [
row['user_id']
for row in df_test.select('user_id').distinct().collect()
]
train_user_ids = list(
set(all_user_ids) - set(val_user_ids) - set(test_user_ids))
selected_train_ids = sample(train_user_ids,
round(len(train_user_ids) * 0.2))
# >>> len(all_user_ids)
# 1129318
# >>> len(val_user_ids)
# 10000
# >>> len(test_user_ids)
# 100000
# >>> len(train_user_ids)
# 1019318
df_train = df_train.where(
df_train.user_id.isin(selected_train_ids + val_user_ids +
test_user_ids))
print('Sampling Finished')
user_indexer = StringIndexer(inputCol="user_id",
outputCol="user_id_numeric")
track_indexer = StringIndexer(inputCol="track_id",
outputCol="track_id_numeric")
model = Pipeline(stages=[user_indexer, track_indexer]).fit(
df_train.union(df_val).union(df_test))
df_train, df_val, df_test = [
model.transform(x) for x in (df_train, df_val, df_test)
]
df_train = df_train.select("user_id_numeric", "track_id_numeric", "count")
df_val = df_val.select("user_id_numeric", "track_id_numeric", "count")
df_test = df_test.select("user_id_numeric", "track_id_numeric", "count")
print('Formatting Finished')
df_train_subsampled.write.parquet("./train_formatted.parquet",
mode='overwrite')
df_val.write.parquet("./val_formatted.parquet", mode='overwrite')
df_test.write.parquet("./test_formatted.parquet", mode='overwrite')
end = time()
print('Indexer and Subsampler succeed. Took %f s' % (end - beg))
return
assembler = VectorAssembler(inputCols=['weight_kg', 'cyl', 'type_dummy'], outputCol='features')
# Split the data into training and testing sets
kars_train, kars_test = kars.randomSplit([0.8, 0.2], seed=23)
# Fit a Logistic Regression model to the training data
regression = LinearRegression(labelCol='consumption')
# Combine steps into a pipeline
pipeline = Pipeline(stages=[indexer, onehot, assembler, regression])
# run fit on training data
pipeline = pipeline.fit(kars_train)
# Make predictions on the testing data
prediction = pipeline.transform(kars_test)
# Create a confusion matrix, comparing predictions to known labels
prediction.groupBy("consumption", 'prediction').count().show(8)
# Print the coefficients and intercept for linear regression
print("Coefficients: %s" % str(pipeline.stages[REGRESSION_STAGE].coefficients))
print("Intercept: %s" % str(pipeline.stages[REGRESSION_STAGE].intercept))
# Summarize the model over the training set and print out some metrics
trainingSummary = pipeline.stages[REGRESSION_STAGE].summary
print("RMSE: %f" % trainingSummary.rootMeanSquaredError)
print("r2: %f" % trainingSummary.r2)
spark.stop()
'''
mlSourceDF = featureeddf
mlSourceDF.printSchema()
mlSourceDF=mlSourceDF.fillna(0, subset= [x for x in mlSourceDF.columns if 'Lag' in x])
# after creating all lag features, we can drop NA columns on the key columns
# drop na to avoid error in StringIndex
mlSourceDF = mlSourceDF.na.drop(subset=["ServerIP","SessionStartHourTime"])
# indexing
columnsForIndex = ['dayofweek', 'ServerIP', 'year', 'month', 'weekofyear', 'dayofmonth', 'hourofday',
'Holiday', 'BusinessHour', 'Morning']
mlSourceDF=mlSourceDF.fillna(0, subset= [x for x in columnsForIndex ])
sIndexers = [StringIndexer(inputCol=x, outputCol=x + '_indexed').setHandleInvalid("skip") for x in columnsForIndex]
indexModel = Pipeline(stages=sIndexers).fit(mlSourceDF)
mlSourceDF = indexModel.transform(mlSourceDF)
# save model for operationalization
indexModel.write().overwrite().save(stringIndexModelFile)
# encoding for categorical features
catVarNames=[x + '_indexed' for x in columnsForIndex ]
columnOnlyIndexed = [ catVarNames[i] for i in range(0,len(catVarNames)) if len(indexModel.stages[i].labels)<2 ]
columnForEncode = [ catVarNames[i] for i in range(0,len(catVarNames)) if len(indexModel.stages[i].labels)>=2 ]
info['columnOnlyIndexed'] = columnOnlyIndexed
info['columnForEncode'] = columnForEncode
# save info to blob storage
write_blob(info, infoFile, storageContainer, storageAccount, storageKey)
outputCol="features")
############ Classifiers
rfC = RandomForestClassifier(labelCol="Survived",
featuresCol="features",
numTrees=300,
maxDepth=5)
gbtC = GBTClassifier(labelCol="Survived", featuresCol="features", maxIter=50)
pipeline = Pipeline().setStages([
sex_stringIndexer, age_discretizer, fare_discretizer,
embarked_stringIndexer, embarked_encoder, VectorAssembler, rfC
]).fit(train_df)
##### Applying pipeline
train_piped = pipeline.transform(train_df)
test_piped = pipeline.transform(test_df)
############################################### Feature importances
print("\n----------- Feature importances")
rfCmodel = pipeline.stages[6]
for feature_name, feature_importance in sorted(zip(
features_column, rfCmodel.featureImportances),
key=lambda x: -x[1]):
print("%20s: %s" % (feature_name, feature_importance))
############################################## Exporting
df_predictions = test_piped.select("prediction").toPandas().reset_index()
df_predictions['index'] = df_predictions['index'] + 892
df_predictions.columns = ['PassengerId', 'Survived']
tree = DecisionTreeClassifier(labelCol='Survived')
rf = RandomForestClassifier(labelCol='Survived')
# 4. Create pipeline
from pyspark.ml import Pipeline
#pipeline = Pipeline(stages=[indexer, onehot, assembler, tree])
pipeline = Pipeline(stages=[
title_extractor, indexer1, indexer2, indexer3, onehot, assembler, rf
])
# 5. Fit the model
pipeline = pipeline.fit(passengers_train)
# 6. Make predictions
#from pyspark.ml.evaluation import BinaryClassEvaluator
prediction = pipeline.transform(passengers_train)
prediction.show(5)
prediction.select('Survived', 'prediction', 'probability').show(5, False)
# Create a confusion matrix
prediction.groupBy('Survived', 'prediction').count().show()
TP = prediction.filter('Survived == 1 AND prediction == 1').count()
TN = prediction.filter('Survived == 0 AND prediction == 0').count()
FP = prediction.filter('Survived == 0 AND prediction == 1').count()
FN = prediction.filter('Survived == 1 AND prediction == 0').count()
# Compute accuracy
accuracy = (TP + TN) / (TP + TN + FP + FN)
print('Accuracy is %f' % accuracy)
df.show()
df.cache()
from pyspark.ml import Pipeline
from pyspark.ml.feature import RegexTokenizer, NGram, HashingTF, MinHashLSH
import pyspark.sql.functions as f
model = Pipeline(stages=[
RegexTokenizer(
pattern="", inputCol="title", outputCol="tokens", minTokenLength=1),
NGram(n=3, inputCol="tokens", outputCol="ngrams"),
HashingTF(inputCol="ngrams", outputCol="vectors"),
MinHashLSH(inputCol="vectors", outputCol="lsh", numHashTables=10)
]).fit(df)
df_hashed = model.transform(df)
df_matches = model.stages[-1].approxSimilarityJoin(df_hashed, df_hashed, 0.9)
#show all matches (including duplicates)
df_matches.select(
f.col('datasetA.id').alias('id_A'),
f.col('datasetB.id').alias('id_B'), f.col('distCol')).show()
#show non-duplicate matches
df_matches.select(
f.col('datasetA.id').alias('id_A'),
f.col('datasetB.id').alias('id_B'),
f.col('distCol')).filter('id_A < id_B').show()
rf_pipeline = Pipeline(stages=[va, rf]).fit(training_df) # In[26]: from pyspark.ml.evaluation import BinaryClassificationEvaluator # In[27]: bce = BinaryClassificationEvaluator() # In[28]: bce.evaluate(lr_pipeline.transform(validation_df)) # In[29]: bce.evaluate(rf_pipeline.transform(validation_df)) # In[30]: lr_model = lr_pipeline.stages[-1] # In[31]: pd.DataFrame(list(zip(airlineCleanDF.columns[12:19], lr_model.coefficients.toArray())),
def train(
self,
df: DataFrame,
params_map: Optional[Dict[str, List[Any]]] = None,
num_folds: Optional[int] = 10,
collect_sub_models: Optional[bool] = False,
return_cv: Optional[bool] = False
) -> Union[PipelineModel, Tuple[PipelineModel, CrossValidatorModel]]:
"""
Train model.
Params
------
df: Spark DataFrame
Input train data
params_map: Optional[Dict[str, List[Any]]] (default=None)
Parameters mapping to grid search over
num_folds: Optional[int] (default=10)
Number of cross-validation folds
collect_sub_models: Optional[bool] (default=False)
Collect models per fold per parameter
combination
return_cv: Optional[bool] (default=False)
Additionally return the CrossValidatorModel
object or not
Returns
-------
self: PipelineModel
The (best) model trained on df.
cv_model: Optional[CrossValidatorModel]
The CrossValidatorModel object.
"""
# get input features
binary, numeric, categorical = self._get_features(df)
# convert categorical to numeric labels
indexed_cols = [f'{c}_idx' for c in categorical]
indexers = [
StringIndexer(inputCol=c[:-6], outputCol=c) for c in indexed_cols
]
self.features = binary + numeric + indexed_cols
self.logger.info(f'Final model features list: {self.features}')
# assemble features into feature vector
assembler = VectorAssembler(inputCols=self.features,
outputCol=self.estimator.getFeaturesCol())
p = Pipeline(stages=indexers + [assembler]).fit(df)
self.logger.info('Index and vector assemble features')
df = p.transform(df)\
.select(self.estimator.getFeaturesCol(), self.estimator.getLabelCol())
# if provided, set estimator params map
if params_map:
self.params_map = params_map
# run cross-validation and choose the best set of parameters
self.logger.info('Start Cross Validation')
cv_params = {
'estimator': self.estimator,
'estimatorParamMaps': self.__params_grid,
'evaluator': self.evaluator,
'numFolds': num_folds,
'collectSubModels': collect_sub_models
}
cv_model = CrossValidator(**cv_params).fit(df)
# set the best model
p.stages.append(cv_model.bestModel)
self.best_model = p
self.logger.info(
f'Set the best model with best params: {self.best_params}')
if return_cv:
return self.best_model, cv_model
else:
return self.best_model
# define numerical assembler first for scaling
numericalAssembler = VectorAssembler(inputCols=numericalColumnsImputed,
outputCol='numerical_cols_imputed')
stages += [numericalAssembler]
# define the standard scaler stage for the numerical columns
scaler = StandardScaler(inputCol='numerical_cols_imputed',
outputCol="numerical_cols_imputed_scaled")
stages += [scaler] # already a list so no need for brackets
# Perform assembly stage to bring together features
assemblerInputs = [c + "classVec" for c in categoricalColumns
] + ["numerical_cols_imputed_scaled"]
# features contains everything, one hot encoded and numerical
assembler = VectorAssembler(inputCols=assemblerInputs,
outputCol="features")
stages += [assembler]
# define the model stage at the end of the pipeline
lr = LogisticRegression(labelCol="label",
featuresCol="features",
maxIter=10)
stages += [lr]
# Random train test split with seed
(trainingData, testData) = data.randomSplit([0.7, 0.3], seed=100)
# Define the entire pipeline and fit on the train data and transform on the test data
clfPipeline = Pipeline().setStages(stages).fit(trainingData)
clfPipeline.transform(testData)
"select * from cours_spark.meteoMensuelle order by 1").cache()
modelA = VectorAssembler().\
setInputCols(
['Janvier', 'Fevrier', 'Mars',
'Avril', 'Mai', 'Juin', 'Juillet',
'Aout', 'Septembre', 'Octobre',
'Novembre', 'Decembre']).\
setOutputCol('variables')
modelN = StandardScaler().\
setInputCol("variables").\
setOutputCol("vNormalisees").\
setWithStd(True).\
setWithMean(False)
modelACP = PCA().\
setInputCol("vNormalisees").\
setOutputCol("vACP").\
setK(2)
modelKM = KMeans().setK(7).\
setFeaturesCol("vACP").\
setPredictionCol("vKM")
modelPipe = Pipeline(stages=[modelA, modelN, modelACP, modelKM]).fit(donnees)
donneesKM = modelPipe.transform(donnees)
donneesKM.select("Ville", "vKM").show(5)
class Pipeline:
def __init__(self, spark, train_date, score_date, train_pct, validate_pct=0.):
assert (0. <= train_pct <= 1.)
assert (0. <= validate_pct <= 1.)
assert (train_pct + validate_pct <= 1.)
self.spark = spark
self.train_date = train_date
self.score_date = score_date
self.train_pct = train_pct
self.validate_pct = validate_pct
def _load_feature_df(self, feature_date, training):
phase = 'training' if training else 'scoring'
print('[ ' + str(datetime.utcnow()) + ' ] : Loading ' + phase + ' feature data')
table_suffix = 't365d' if training else 'scoring'
features_df = self.spark.sql('from grp_gdoop_clv_db.keep_cdf_final_features_' + table_suffix) \
.filter(F.col('record_date') == feature_date) \
.drop('record_date', 'zip_code_cat_x')
if training:
target_df = self.spark.sql('select * from grp_gdoop_clv_db.ce_keep_deact_target') \
.filter(F.col('record_date') == feature_date) \
.select('consumer_id', 'deactivated')
final = features_df \
.join(target_df, features_df.consumer_id == target_df.consumer_id, how='left') \
.drop(target_df.consumer_id)
return final
else:
return features_df
def _train_validate_split(self, features_df):
print('[ ' + str(datetime.utcnow()) + ' ] : Splitting training data into model training and validation data')
splits = features_df.randomSplit([self.train_pct, self.validate_pct, 1 - self.train_pct - self.validate_pct])
return splits[0], splits[1]
def _make_feature_list(self, all_cols, cat_cols, indexers):
features = list(filter(lambda x: x.endswith('_x') and not x.endswith('_cat_x'), all_cols))
for i, col in enumerate(cat_cols):
for label in indexers[i].labels:
features.extend([col + '_' + re.sub('\W+', '_', str(label).strip())])
self.feature_list = features
def _one_hot_encode_pl(self, train_raw):
print('[ ' + str(datetime.utcnow()) + ' ] : Creating feature engineering pipeline')
all_cols = train_raw.columns
cat_cols = list(filter(lambda x: x.endswith('_cat_x'), all_cols))
indexers = [StringIndexer(inputCol=c, outputCol=c.replace('_cat_x', '_index'),
handleInvalid='keep') for c in cat_cols]
one_hots = [OneHotEncoderEstimator(inputCols=[c.replace('_cat_x', '_index')],
outputCols=[c.replace('_cat_x', '_vec_x')], handleInvalid='keep',
dropLast=False) for c in cat_cols]
self.one_hot_plm = MLPipeline(stages=indexers + one_hots).fit(train_raw)
self._make_feature_list(all_cols, cat_cols, self.one_hot_plm.stages[:len(cat_cols)])
def _assemble_features(self, raw_df, data_type):
print('[ ' + str(datetime.utcnow()) + ' ] : Feature engineering ' + data_type + ' data')
cat_cols = list(filter(lambda x: x.endswith('_cat_x'), raw_df.columns))
df = self.one_hot_plm.transform(raw_df) \
.drop(*cat_cols)
features = list(filter(lambda x: x.endswith('_x'), df.columns))
assembler = VectorAssembler(inputCols=features, outputCol='features', handleInvalid='keep')
return assembler.transform(df)
def _training_data(self, validate_model):
train_features_df = self._load_feature_df(self.train_date, True)
train_raw, validate_raw = self._train_validate_split(train_features_df)
train_raw.cache()
# Create one-hot encoding pipeline that will be applied to all DFs
self._one_hot_encode_pl(train_raw)
train_df = self._assemble_features(train_raw, 'model training').cache()
train_raw.unpersist()
if validate_model:
validate_raw.cache()
validate_df = self._assemble_features(validate_raw, 'model validation').cache()
validate_raw.unpersist()
else:
validate_df = None
return train_df, validate_df
def _scoring_data(self):
score_features_df = self._load_feature_df(self.score_date, False).cache()
score_df = self._assemble_features(score_features_df, 'scoring').cache()
score_features_df.unpersist()
return score_df
def run(self, validate_model, score_active_users):
print('\nDATA PIPELINE\n')
train_df, validate_df = self._training_data(validate_model)
if score_active_users:
score_df = self._scoring_data()
else:
score_df = None
return {'training': train_df, 'validation': validate_df, 'scoring': score_df, 'features': self.feature_list}
def __repr__(self):
return '<Pipeline(train_date={0}, score_date={1})>'.format(self.train_date, self.score_date)
def __str__(self):
return '<Pipeline(train_date={0}, score_date={1})>'.format(self.train_date, self.score_date)
onthot = onthot.fit(cars_train)
cars_train = onehot.transform(cars_train)
cars_train =assemble.transform(cars_train)
# Fit model to training data
regression = regression.fit(cars_train)
# Testing data
cars_test = indexer.transform(cars_test)
cars_test = onthot.transform(cars_test)
cars_test = assemble.transform(cars_test)
# Make predictions on testing data
predictions = regression.transform(cars_test)
# Cars model: Pipeline
from pyspark.ml import Pipeline
pipeline = Pipeline(stages=[indexer, onehot, assemble, regression])
pipeline = pipenline.fit(cars_train)
predictions = pipeline.transform(cars_test)
# Cars model: stages
# The LinearRegression object (fourth stage -> index 3)
print(pipeline.stages[3].intercept)
print(pipeline.stages[3].coefficients)
# Convert categorical strings to index values
indexer = StringIndexer(inputCol='org', outputCol='org_idx')
# One-hot encode index values
onehot = OneHotEncoderEstimator(
inputCols=['org_idx', 'dow'],
outputCols=['org_dummy', 'dow_dummy']
)
# Assemble predictors into a single column
