def spark_transform(self):
"""
transforms Spark RDD with raw data into the RDD that contains
top-n pickup spots for each block and time slot
"""
BatchTransformer.spark_transform(self)
n = self.psql_config["topntosave"]
# calculation of top-n spots for each block and time slot
self.data = (self.data.map(lambda x: (
(x["block_id"], x["time_slot"], x["sub_block_id"]), x["passengers"]
)).reduceByKey(lambda x, y: x + y).map(lambda x: ((x[0][0], x[0][
1]), [(x[0][2], x[1])])).reduceByKey(lambda x, y: x + y).mapValues(
lambda vals: heapq.nlargest(n, vals, key=lambda x: x[1])).map(
lambda x: {
"block_id": x[0][0],
"time_slot": x[0][1],
"subblocks_psgcnt": x[1]
}))
self.data.persist(pyspark.StorageLevel(True, True, False, False,
3)).count() # MEMORY_AND_DISK_3
# recalculation of top-n, where for each key=(block_id, time_slot) top-n is calculated
# based on top-n of (block_id, time_slot) and top-ns of (adjacent_block, time_slot+1)
# from all adjacent blocks
maxval = self.psql_config["upperBound"]
self.data = (self.data.map(lambda x: ((x["block_id"], x[
"time_slot"]), x["subblocks_psgcnt"])).flatMap(lambda x: [x] + [
((bl, (x[0][1] - 1) % maxval), x[1])
for bl in helpers.get_neighboring_blocks(x[0][0])
]).reduceByKey(lambda x, y: x + y).mapValues(
lambda vals: heapq.nlargest(n, vals, key=lambda x: x[1])).map(
lambda x: {
"block_latid":
x[0][0][0],
"block_lonid":
x[0][0][1],
"time_slot":
x[0][1],
"longitude": [
helpers.determine_subblock_lonlat(el[0])[0]
for el in x[1]
],
"latitude": [
helpers.determine_subblock_lonlat(el[0])[1]
for el in x[1]
],
"passengers": [el[1] for el in x[1]]
}))
self.data.persist(pyspark.StorageLevel(True, True, False, False,
3)).count() # MEMORY_AND_DISK_3
def generate_digraph(edge_count=40, batch_size=50,
steps_to_python_gc=50, domain_size=50):
sc = pyspark.context.SparkContext.getOrCreate()
sc.setCheckpointDir("~/.transitive_closure")
spark = pyspark.sql.SparkSession(sc)
# Translation to Spark format is ludicrously slow without PyArrow
spark.conf.set("spark.sql.execution.arrow.enabled", "true")
batches_to_python_gc = steps_to_python_gc // batch_size
batch_count = int(np.ceil(edge_count/batch_size))
digraph=[]
for i in range(batch_count):
if i%batches_to_python_gc == 0:
gc.collect()
if i == batch_count-1 and edge_count % batch_size > 0:
batch_size = edge_count % batch_size
new_origins = np.random.randint(0,domain_size,dtype=np.int32,
size=(batch_size,1))
new_termini = np.random.randint(0,domain_size,dtype=np.int32,
size=(batch_size,1))
partial_digraph = pd.DataFrame(np.concatenate(
[new_origins, new_termini], 1), columns=("origin", "terminus")
)
partial_digraph = spark.createDataFrame(partial_digraph).distinct()
if digraph != []:
digraph[0] = digraph[1].union(digraph[0])\
.orderBy(["origin", "terminus"],
ascending=[True, True])\
.distinct()\
.persist(pyspark.StorageLevel(True, False,
False, True,
1))\
.checkpoint()
digraph[1] = partial_digraph
else:
digraph = [partial_digraph,partial_digraph]
if i == batch_count-1:
gc.collect()
digraph = digraph[1].union(digraph[0])\
.orderBy(["origin", "terminus"],
ascending=[True, True])\
.distinct()\
.persist(pyspark.StorageLevel(True, False,
False, True,
1))\
.checkpoint()
df=digraph.toPandas()
sc.stop()
return df
def create_df_from_generator(gen, names):
a = sc.parallelize(gen, 20)
a.persist(pyspark.StorageLevel(True, True, False, True, 1))
df = sqlContext.createDataFrame(a, schema=names,
samplingRatio=None).repartition(20)
#df.persist(pyspark.StorageLevel(True, True, False, True, 1))
return df
def transitive_closure_from_dataframe(digraph):
dir_path = os.path.dirname(os.path.realpath(__file__))
save_file = os.path.join(dir_path,'solution.pickle')
try:
shutil.rmtree(save_file)
except:
pass
sc = pyspark.context.SparkContext.getOrCreate()
sc.setCheckpointDir("~/.transitive_closure")
spark = pyspark.sql.SparkSession(sc)
# Translation to Spark format is ludicrously slow with PyArrow
spark.conf.set("spark.sql.execution.arrow.enabled", "true")
start_time=time.time()
orig_digraph = spark.createDataFrame(digraph.copy())
new_edges = spark.createDataFrame(digraph.copy()).checkpoint()
new_edges_mem = digraph
closed_digraph = spark.createDataFrame(digraph).checkpoint()
while not new_edges_mem.empty:
new_edges = spark.createDataFrame(new_edges_mem).persist(
pyspark.StorageLevel(False, True, False, False, 1)).checkpoint()
new_edges = new_edges.join(orig_digraph,
(new_edges.terminus == orig_digraph.origin))\
.select(new_edges.origin, orig_digraph.terminus)\
.union(orig_digraph.join(new_edges,
(new_edges.origin == orig_digraph.terminus))\
.select(orig_digraph.origin, new_edges.terminus))\
.distinct()\
.exceptAll(closed_digraph)
# I don't see any copy method, and PyArrow is nearly
# instantaneous, so I'm going to use pandas to copy the new
# edges to memory.
new_edges_mem = new_edges.toPandas().copy()
closed_digraph = closed_digraph.union(new_edges.persist(
pyspark.StorageLevel(True, False, False, True, 1)
).checkpoint())
# Putting this in just because of how much trouble I had with the
# generate_digraph function
gc.collect()
# Ideally, we would be able to use the fact that closed_digraph is
# already stored on the disk, but I'm not quite sure how to do it.
# Also, for clusters with separate hard drives,
# we would want to mount it somewhere with HDFS or similar.
closed_digraph.rdd.saveAsPickleFile(save_file)
df = closed_digraph.toPandas()
sc.stop()
return df
def create_dataframe(spark_context, sql_context, table, column_names):
data = spark_context.parallelize(table, 20)
data.persist(pyspark.StorageLevel(True, True, False, True, 1))
df = sql_context.createDataFrame(data,
schema=column_names,
samplingRatio=None).repartition(20)
# df.persist(pyspark.StorageLevel(True, True, False, True, 1))
return df
return result
#print(personName+","+changeManager+","+promote+","+str(ManagerCata.index(lastManager)) )
def getCataDict():
return {0: 2, 1: 2, 2: len(ManagerCata)}
couchbase_host = '10.1.193.189'
couchbase_bucket = 'persona'
couchbucket = Couchbase.connect(bucket=couchbase_bucket, host=couchbase_host)
# reference http://spark.apache.org/docs/latest/api/python/pyspark.mllib.html#module-pyspark.mllib.tree
pyspark.StorageLevel(True, False, False, False, 1)
sc = SparkContext("local", "ad RD leave predict")
if __name__ == '__main__':
prepareLeaverList('../var/all2015rdleaves')
data = []
l = 500
for i in range(6):
data.extend(buildDataFromCouchbase(l, l * i))
cataDict = getCataDict()
model = GradientBoostedTrees.trainClassifier(sc.parallelize(data),
cataDict,
numIterations=10,
maxBins=500)
def __init__(self):
pyspark.StorageLevel(True, True, False, False, 2)
def main(result_dir_master, result_dir_s3):
CON_CONFIGS = {}
CON_CONFIGS["result_dir_master"] = result_dir_master
CON_CONFIGS["result_dir_s3"] = result_dir_s3
#
## user to specify: hyper-params
# clustering
CON_CONFIGS["n_clusters"] = 3
CON_CONFIGS["warn_threshold_np_ratio"] = 1
# classification
CON_CONFIGS["n_eval_folds"] = 5
CON_CONFIGS["n_cv_folds"] = 5
CON_CONFIGS["lambdas"] = list(10.0 ** numpy.arange(-2, 2, 1.0))
CON_CONFIGS["alphas"] = list(numpy.linspace(0, 1, 3))
CON_CONFIGS["desired_recalls"] = [0.025,0.05,0.075,0.1,0.125,0.15,0.175,0.2,0.225,0.25]
# CON_CONFIGS["desired_recalls"] = [0.05,0.10]
#
## read data and some meta studff
# user to specify : seed in Random Forest model
CON_CONFIGS["seed"] = 42
CON_CONFIGS["data_path"] = "s3://emr-rwes-pa-spark-dev-datastore/lichao.test/data/BI/smaller_data/"
CON_CONFIGS["pos_file"] = "pos_70.0pct.csv"
CON_CONFIGS["neg_file"] = "neg_70.0pct.csv"
CON_CONFIGS["ss_file"] = "ss_70.0pct.csv"
#reading in the data from S3
spark = SparkSession.builder.appName(os.path.basename(__file__)).getOrCreate()
org_pos_data = spark.read.option("header", "true")\
.option("inferSchema", "true")\
.csv(CON_CONFIGS["data_path"] + CON_CONFIGS["pos_file"])
org_neg_data = spark.read.option("header", "true")\
.option("inferSchema", "true")\
.csv(CON_CONFIGS["data_path"] + CON_CONFIGS["neg_file"])\
.select(org_pos_data.columns)
org_ss_data = spark.read.option("header", "true")\
.option("inferSchema", "true")\
.csv(CON_CONFIGS["data_path"] +CON_CONFIGS["ss_file"])\
.select(org_pos_data.columns)
# user to specify: original column names for predictors and output in data
orgOutputCol = "label"
matchCol = "matched_positive_id"
patIDCol = "patid"
nonFeatureCols = [matchCol, orgOutputCol, patIDCol]
# orgPredictorCols = ["PATIENT_AGE", "LOOKBACK_DAYS", "LVL3_CHRN_ISCH_HD_FLAG", "LVL3_ABN_CHST_XRAY_FLAG"]
# org_pos_data = org_pos_data.select(nonFeatureCols + orgPredictorCols)
# org_neg_data = org_neg_data.select(nonFeatureCols + orgPredictorCols)
# org_ss_data = org_ss_data.select(nonFeatureCols + orgPredictorCols)
# sanity check
if type(org_pos_data.select(orgOutputCol).schema.fields[0].dataType) not in (DoubleType, IntegerType):
raise TypeError("The output column is not of type integer or double. ")
org_pos_data = org_pos_data.withColumn(orgOutputCol, org_pos_data[orgOutputCol].cast("double"))
orgPredictorCols = [x for x in org_pos_data.columns if x not in nonFeatureCols]
orgPredictorCols4Clustering = [x for x in orgPredictorCols if "FLAG" in x]
if type(org_neg_data.select(orgOutputCol).schema.fields[0].dataType) not in (DoubleType, IntegerType):
raise TypeError("The output column is not of type integer or double. ")
org_neg_data = org_neg_data.withColumn(orgOutputCol, org_neg_data[orgOutputCol].cast("double"))
if type(org_ss_data.select(orgOutputCol).schema.fields[0].dataType) not in (DoubleType, IntegerType):
raise TypeError("The output column is not of type integer or double. ")
org_ss_data = org_ss_data.withColumn(orgOutputCol, org_ss_data[orgOutputCol].cast("double"))
#
clusterFeatureCol = "cluster_features"
clusterCol = "cluster_id"
# user to specify: the collective column name for all predictors
collectivePredictorCol = "features"
# in-cluster distance
distCol = "dist"
# user to specify: the column name for prediction
predictionCol = "probability"
CON_CONFIGS["orgPredictorCols"] = orgPredictorCols
CON_CONFIGS["orgPredictorCols4Clustering"] = orgPredictorCols4Clustering
CON_CONFIGS["n_predictors_classification"] = len(orgPredictorCols)
CON_CONFIGS["n_rows_pos"] = org_pos_data.count()
CON_CONFIGS["n_rows_neg"] = org_neg_data.count()
CON_CONFIGS["n_rows_ss"] = org_ss_data.count()
save_analysis_info(\
result_dir_master,
"analysis_info.txt",
CON_CONFIGS
)
# convert to ml-compatible format
assembler = VectorAssembler(inputCols=orgPredictorCols, outputCol=collectivePredictorCol)
posFeatureAssembledData = assembler.transform(org_pos_data)\
.select(nonFeatureCols + [collectivePredictorCol])
negFeatureAssembledData = assembler.transform(org_neg_data)\
.select(nonFeatureCols + [collectivePredictorCol])
#
evalIDCol = "evalFoldID"
cvIDCol = "cvFoldID"
pos_neg_data = posFeatureAssembledData.union(negFeatureAssembledData)
pos_neg_data_with_eval_ids = AppendDataMatchingFoldIDs(pos_neg_data, CON_CONFIGS["n_eval_folds"], matchCol, foldCol=evalIDCol)
# the model (pipeline)
classifier_spec = LogisticRegression(maxIter=1e5, featuresCol = collectivePredictorCol,
labelCol = orgOutputCol, standardization = True)
evaluator = BinaryClassificationEvaluatorWithPrecisionAtRecall(\
rawPredictionCol=predictionCol,
labelCol=orgOutputCol,
metricName="precisionAtGivenRecall",
metricParams={"recallValue":0.05}\
)
paramGrid = ParamGridBuilder()\
.addGrid(classifier_spec.regParam, CON_CONFIGS["lambdas"])\
.addGrid(classifier_spec.elasticNetParam, CON_CONFIGS["alphas"])\
.build()
# cross-evaluation
predictionsAllData = None
kmeans = KMeans(featuresCol=clusterFeatureCol, predictionCol=clusterCol).setK(CON_CONFIGS["n_clusters"])
cluster_assembler = VectorAssembler(inputCols=orgPredictorCols4Clustering, outputCol=clusterFeatureCol)
metricSets = [{"metricName": "precisionAtGivenRecall", "metricParams": {"recallValue": x}} for x in CON_CONFIGS["desired_recalls"]]
filename_loop_info = result_dir_master + "loop_info.txt"
file_loop_info = open(filename_loop_info, "w")
inputTrivialNegPredCols = ["_pos_prob", "_neg_prob"]
trivial_neg_pred_assembler = VectorAssembler(inputCols=inputTrivialNegPredCols, outputCol=predictionCol)
for iFold in range(CON_CONFIGS["n_eval_folds"]):
condition = pos_neg_data_with_eval_ids[evalIDCol] == iFold
leftoutFold = pos_neg_data_with_eval_ids.filter(condition).drop(evalIDCol)
trainFolds = pos_neg_data_with_eval_ids.filter(~condition).drop(evalIDCol)
file_loop_info.write("####################################################################\n\n".format(iFold))
file_loop_info.write("iFold: {}\n\n".format(iFold))
file_loop_info.write("n_rows of leftoutFold: {}\n".format(leftoutFold.count()))
file_loop_info.write("n_rows of trainFolds: {}\n".format(trainFolds.count()))
#
## clustering to be done here
pos_data_4_clustering = trainFolds\
.filter(F.col(orgOutputCol)==1)\
.select(patIDCol)\
.join(org_pos_data, patIDCol)
pos_data_4_clustering_assembled = cluster_assembler.transform(pos_data_4_clustering)\
.select([patIDCol, matchCol] + [clusterFeatureCol])
cluster_model, clustered_pos = clustering(pos_data_4_clustering_assembled, kmeans,
clusterFeatureCol, clusterCol, distCol)
nPosesAllClusters = clustered_pos.count()
predictionsOneFold = None
file_loop_info.write("nPosesAllClusters: {}\n".format(nPosesAllClusters))
for i_cluster in range(CON_CONFIGS["n_clusters"]):
file_loop_info.write("i_cluster: {}\n\n".format(i_cluster))
# the positive data for training the classifier
train_pos = clustered_pos\
.filter(clustered_pos[clusterCol]==i_cluster)\
.select(patIDCol)\
.join(trainFolds, patIDCol)
file_loop_info.write("n_rows of train_pos: {}\n".format(train_pos.count()))
posPctThisClusterVSAllClusters = float(train_pos.count()) / nPosesAllClusters
file_loop_info.write("posPctThisClusterVSAllClusters: {}\n".format(posPctThisClusterVSAllClusters))
# select negative training data based on the clustering result
corresponding_neg = train_pos\
.select(matchCol)\
.join(org_neg_data, matchCol)
corresponding_neg_4_clustering_assembled = cluster_assembler.transform(corresponding_neg)\
.select([patIDCol, matchCol] + [clusterFeatureCol])
similar_neg_ids = select_certain_pct_ids_per_positive_closest_to_cluster_centre(\
corresponding_neg_4_clustering_assembled,
clusterFeatureCol,
cluster_model.clusterCenters()[i_cluster],
posPctThisClusterVSAllClusters,
patIDCol,
matchCol
)
train_data = similar_neg_ids\
.join(trainFolds, patIDCol)\
.select(train_pos.columns)\
.union(train_pos)
file_loop_info.write("n_rows of train_data: {}\n".format(train_data.count()))
trainDataWithCVFoldID = AppendDataMatchingFoldIDs(train_data, CON_CONFIGS["n_cv_folds"], matchCol, foldCol=cvIDCol)
trainDataWithCVFoldID.coalesce(int(trainFolds.rdd.getNumPartitions() * posPctThisClusterVSAllClusters) + 1)
# sanity check: if there are too few negatives for any positive
# thresh_n_neg_per_fold = round(train_pos.count() / float(CON_CONFIGS["n_cv_folds"])) * CON_CONFIGS["warn_threshold_np_ratio"]
# neg_counts_all_cv_folds = trainDataWithCVFoldID\
# .filter(F.col(orgOutputCol)==0)\
# .groupBy(cvIDCol)\
# .agg(F.count(orgOutputCol).alias("_tmp"))\
# .select("_tmp")\
# .collect()
# if any(map(lambda x: x["_tmp"] < thresh_n_neg_per_fold, neg_counts_all_cv_folds)):
# raise ValueError("Insufficient number of negative data in at least one cv fold.")
#
## train the classifier
validator = CrossValidatorWithStratificationID(\
estimator=classifier_spec,
estimatorParamMaps=paramGrid,
evaluator=evaluator,
stratifyCol=cvIDCol\
)
cvModel = validator.fit(trainDataWithCVFoldID)
#
## test data
entireTestData = org_ss_data\
.join(leftoutFold.filter(F.col(orgOutputCol)==1).select(matchCol), matchCol).select(org_pos_data.columns)\
.union(org_pos_data.join(leftoutFold.select(patIDCol), patIDCol).select(org_pos_data.columns))\
.union(org_neg_data.join(leftoutFold.select(patIDCol), patIDCol).select(org_pos_data.columns))
entireTestDataAssembled4Clustering = cluster_assembler.transform(entireTestData)\
.select([patIDCol, matchCol] + [clusterFeatureCol])
file_loop_info.write("n_rows of entireTestData: {}\n".format(entireTestData.count()))
filteredTestData = select_certain_pct_overall_ids_closest_to_cluster_centre(\
entireTestDataAssembled4Clustering,
clusterFeatureCol,
cluster_model.clusterCenters()[i_cluster],
posPctThisClusterVSAllClusters,
patIDCol
).join(entireTestData, patIDCol)
file_loop_info.write("n_rows of filteredTestData: {}\n".format(filteredTestData.count()))
filteredTestDataAssembled = assembler.transform(filteredTestData)\
.select(nonFeatureCols + [collectivePredictorCol])
# testing
predictions = cvModel\
.transform(filteredTestDataAssembled)\
.select(nonFeatureCols + [collectivePredictorCol, predictionCol])
# need to union the test data filtered away (all classified as negative)
discarded_test_ids = entireTestData\
.select(patIDCol)\
.subtract(filteredTestData.select(patIDCol))
discardedTestData = discarded_test_ids\
.join(entireTestData, patIDCol)
discardedTestDataAssembled = assembler.transform(discardedTestData, )\
.select(nonFeatureCols + [collectivePredictorCol])
predictionsDiscardedTestData = discardedTestDataAssembled\
.withColumn(inputTrivialNegPredCols[0], F.lit(0.0))\
.withColumn(inputTrivialNegPredCols[1], F.lit(1.0))
predictionsDiscardedTestDataAssembled = trivial_neg_pred_assembler\
.transform(predictionsDiscardedTestData)\
.select(predictions.columns)
predictionsEntireTestData = predictions.union(predictionsDiscardedTestDataAssembled)
metricValuesOneCluster = evaluator\
.evaluateWithSeveralMetrics(predictionsEntireTestData, metricSets = metricSets)
file_name_metrics_one_cluster = result_dir_master + "metrics_cluster_" + str(i_cluster) + "fold_" + str(iFold) + "_.csv"
save_metrics(file_name_metrics_one_cluster, metricValuesOneCluster)
predictionsEntireTestData.write.csv(result_dir_s3 + "predictions_fold_" + str(iFold) + "_cluster_" + str(i_cluster) + ".csv")
predictionsEntireTestData.persist(pyspark.StorageLevel(True, False, False, False, 1))
if predictionsOneFold is not None:
predictionsOneFold = predictionsOneFold.union(predictionsEntireTestData)
else:
predictionsOneFold = predictionsEntireTestData
# save the metrics for all hyper-parameter sets in cv
cvMetrics = cvModel.avgMetrics
cvMetricsFileName = result_dir_s3 + "cvMetrics_cluster_" + str(i_cluster) + "_fold_" + str(iFold)
cvMetrics.coalesce(4).write.csv(cvMetricsFileName, header="true")
# save the hyper-parameters of the best model
bestParams = validator.getBestModelParams()
file_best_params = result_dir_master + "bestParams_cluster_" + str(i_cluster) + "_fold_" + str(iFold) + ".txt"
with open(file_best_params, "w") as fileBestParams:
fileBestParams.write(str(bestParams))
os.chmod(file_best_params, 0o777)
# summarise all clusters from the fold
metricValuesOneFold = evaluator\
.evaluateWithSeveralMetrics(predictionsOneFold, metricSets = metricSets)
file_name_metrics_one_fold = result_dir_master + "metrics_fold_" + str(iFold) + "_.csv"
save_metrics(file_name_metrics_one_fold, metricValuesOneFold)
if predictionsAllData is not None:
predictionsAllData = predictionsAllData.union(predictionsOneFold)
else:
predictionsAllData = predictionsOneFold
# save all predictions
predictionsFileName = result_dir_s3 + "predictionsAllData"
predictionsAllData.select(orgOutputCol,
getitem(1)(predictionCol).alias('prob_1'))\
.write.csv(predictionsFileName, header="true")
# metrics of predictions on the entire dataset
metricValues = evaluator\
.evaluateWithSeveralMetrics(predictionsAllData, metricSets = metricSets)
save_metrics(result_dir_master + "metricValuesEntireData.csv", metricValues)
file_loop_info.close()
os.chmod(file_loop_info, 0o777)
spark.stop()
