def skew_drift_validator(mode, gcp_bucket, control_set_path,
treatment_set_path, feature_list_str, Linf_value):
logging.basicConfig(level=logging.INFO)
logging.info('Starting skew drift validator ..')
logging.info('Input data:')
logging.info('mode:{}'.format(mode))
logging.info('gcp_bucket:{}'.format(gcp_bucket))
logging.info('control_set_path:{}'.format(control_set_path))
logging.info('treatment_set_path:{}'.format(treatment_set_path))
logging.info('Linf_value:{}'.format(Linf_value))
feature_list = eval(feature_list_str)
control_set_df = pd.read_csv("gs://" + gcp_bucket + "/" + control_set_path,
sep=',')
treat_set_df = pd.read_csv("gs://" + gcp_bucket + "/" + treatment_set_path,
sep=',')
control_stats = tfdv.generate_statistics_from_dataframe(
dataframe=control_set_df)
treat_stats = tfdv.generate_statistics_from_dataframe(
dataframe=treat_set_df)
control_schema = tfdv.infer_schema(control_stats)
treat_schema = tfdv.infer_schema(treat_stats)
for feature in feature_list:
if (mode == "skew"):
if (tfdv.get_feature(control_schema, feature).domain
): # if we have domain it is a categorical variable
tfdv.get_feature(
control_schema, feature
).skew_comparator.infinity_norm.threshold = Linf_value
else:
logging.critical(
"feature: {} is not categorical".format(feature))
sys.exit(1)
elif (mode == "drift"):
tfdv.get_feature(
control_schema,
feature).drift_comparator.infinity_norm.threshold = Linf_value
else:
logging.critical("mode: {} not supported".format(mode))
sys.exit(1)
anomalies = tfdv.validate_statistics(statistics=control_stats,
schema=control_schema,
serving_statistics=treat_stats)
if (anomalies.anomaly_info):
logging.info("Data-{} detected:".format(anomalies))
return anomalies
else:
logging.info("No data-{} detected".format(mode))
def tfdv_skew_validator(feature_name, train_stats, serve_stats, schema, threshold):
"""
Validate skew for the csv dataset
"""
#this doesn't display skew anomalies as the book shows
tfdv.get_feature(schema, feature_name).skew_comparator.infinity_norm.threshold = threshold
skew_anomalies = tfdv.validate_statistics(statistics = train_stats,
schema = schema,
serving_statistics = serve_stats)
tfdv.display_anomalies(skew_anomalies)
return skew_anomalies
def tfdv_drift_validator(feature_name, train_stats, previous_stats, schema, threshold):
"""
Validate drift for the csv dataset
"""
#this doesn't display drift anomalies as the book shows
tfdv.get_feature(schema, feature_name).drift_comparator.infinity_norm.threshold = threshold
drift_anomalies = tfdv.validate_statistics(statistics=train_stats,
schema=schema,
previous_statistics= previous_stats
)
tfdv.display_anomalies(drift_anomalies)
return drift_anomalies
def tfdv_detect_drift(
stats_older_path: str, stats_new_path: str
) -> NamedTuple('Outputs', [('drift', str)]):
import logging
import time
import tensorflow_data_validation as tfdv
import tensorflow_data_validation.statistics.stats_impl
logging.getLogger().setLevel(logging.INFO)
logging.info('stats_older_path: %s', stats_older_path)
logging.info('stats_new_path: %s', stats_new_path)
if stats_older_path == 'none':
return ('true', )
stats1 = tfdv.load_statistics(stats_older_path)
stats2 = tfdv.load_statistics(stats_new_path)
schema1 = tfdv.infer_schema(statistics=stats1)
tfdv.get_feature(schema1, 'duration').drift_comparator.jensen_shannon_divergence.threshold = 0.01
drift_anomalies = tfdv.validate_statistics(
statistics=stats2, schema=schema1, previous_statistics=stats1)
logging.info('drift analysis results: %s', drift_anomalies.drift_skew_info)
from google.protobuf.json_format import MessageToDict
d = MessageToDict(drift_anomalies)
val = d['driftSkewInfo'][0]['driftMeasurements'][0]['value']
thresh = d['driftSkewInfo'][0]['driftMeasurements'][0]['threshold']
logging.info('value %s and threshold %s', val, thresh)
res = 'true'
if val < thresh:
res = 'false'
logging.info('train decision: %s', res)
return (res, )
# Compare evaluation data with training data
displayHTML(
get_statistics_html(lhs_statistics=stats_serve,
rhs_statistics=stats_train,
lhs_name='SERVE_DATASET',
rhs_name='TRAIN_DATASET'))
# COMMAND ----------
anomalies = tfdv.validate_statistics(statistics=stats_serve, schema=schema)
tfdv.display_anomalies(anomalies)
# COMMAND ----------
# Add skew and drift comparators
temp_f = tfdv.get_feature(schema, 'avg_temp_f')
temp_f.skew_comparator.jensen_shannon_divergence.threshold = 0
temp_f.drift_comparator.jensen_shannon_divergence.threshold = 0
precip_mm = tfdv.get_feature(schema, 'tot_precip_mm')
precip_mm.skew_comparator.jensen_shannon_divergence.threshold = 0
precip_mm.drift_comparator.jensen_shannon_divergence.threshold = 0
_anomalies = tfdv.validate_statistics(stats_train,
schema,
serving_statistics=stats_serve)
tfdv.display_anomalies(_anomalies)
# COMMAND ----------
def data_validation(EPOCHS: int, STEPS: int, BATCH_SIZE: int, HIDDEN_LAYER_SIZE: str, LEARNING_RATE: float):
import os
import shutil
from kale.utils import pod_utils
from kale.marshal import resource_save as _kale_resource_save
from kale.marshal import resource_load as _kale_resource_load
_kale_data_directory = "/marshal"
if not os.path.isdir(_kale_data_directory):
os.makedirs(_kale_data_directory, exist_ok=True)
import os
import shutil
import logging
import apache_beam as beam
import tensorflow as tf
import tensorflow_transform as tft
import tensorflow_model_analysis as tfma
import tensorflow_data_validation as tfdv
from apache_beam.io import textio
from apache_beam.io import tfrecordio
from tensorflow_transform.beam import impl as beam_impl
from tensorflow_transform.beam.tft_beam_io import transform_fn_io
from tensorflow_transform.coders.csv_coder import CsvCoder
from tensorflow_transform.coders.example_proto_coder import ExampleProtoCoder
from tensorflow_transform.tf_metadata import dataset_metadata
from tensorflow_transform.tf_metadata import metadata_io
DATA_DIR = 'data/'
TRAIN_DATA = os.path.join(DATA_DIR, 'taxi-cab-classification/train.csv')
EVALUATION_DATA = os.path.join(
DATA_DIR, 'taxi-cab-classification/eval.csv')
# Categorical features are assumed to each have a maximum value in the dataset.
MAX_CATEGORICAL_FEATURE_VALUES = [24, 31, 12]
CATEGORICAL_FEATURE_KEYS = ['trip_start_hour',
'trip_start_day', 'trip_start_month']
DENSE_FLOAT_FEATURE_KEYS = ['trip_miles', 'fare', 'trip_seconds']
# Number of buckets used by tf.transform for encoding each feature.
FEATURE_BUCKET_COUNT = 10
BUCKET_FEATURE_KEYS = [
'pickup_latitude', 'pickup_longitude', 'dropoff_latitude', 'dropoff_longitude']
# Number of vocabulary terms used for encoding VOCAB_FEATURES by tf.transform
VOCAB_SIZE = 1000
# Count of out-of-vocab buckets in which unrecognized VOCAB_FEATURES are hashed.
OOV_SIZE = 10
VOCAB_FEATURE_KEYS = ['pickup_census_tract', 'dropoff_census_tract', 'payment_type', 'company',
'pickup_community_area', 'dropoff_community_area']
# allow nan values in these features.
OPTIONAL_FEATURES = ['dropoff_latitude', 'dropoff_longitude', 'pickup_census_tract', 'dropoff_census_tract',
'company', 'trip_seconds', 'dropoff_community_area']
LABEL_KEY = 'tips'
FARE_KEY = 'fare'
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.INFO)
# tf.get_logger().setLevel(logging.ERROR)
vldn_output = os.path.join(DATA_DIR, 'validation')
# TODO: Understand why this was used in the conversion to the output json
# key columns: list of the names for columns that should be treated as unique keys.
key_columns = ['trip_start_timestamp']
# read the first line of the cvs to have and ordered list of column names
# (the Schema will scrable the features)
with open(TRAIN_DATA) as f:
column_names = f.readline().strip().split(',')
stats = tfdv.generate_statistics_from_csv(data_location=TRAIN_DATA)
schema = tfdv.infer_schema(stats)
eval_stats = tfdv.generate_statistics_from_csv(
data_location=EVALUATION_DATA)
anomalies = tfdv.validate_statistics(eval_stats, schema)
# Log anomalies
for feature_name, anomaly_info in anomalies.anomaly_info.items():
logging.getLogger().error(
'Anomaly in feature "{}": {}'.format(
feature_name, anomaly_info.description))
# show inferred schema
tfdv.display_schema(schema=schema)
# Resolve anomalies
company = tfdv.get_feature(schema, 'company')
company.distribution_constraints.min_domain_mass = 0.9
# Add new value to the domain of feature payment_type.
payment_type_domain = tfdv.get_domain(schema, 'payment_type')
payment_type_domain.value.append('Prcard')
# Validate eval stats after updating the schema
updated_anomalies = tfdv.validate_statistics(eval_stats, schema)
tfdv.display_anomalies(updated_anomalies)
# -----------------------DATA SAVING START---------------------------------
if "column_names" in locals():
_kale_resource_save(column_names, os.path.join(
_kale_data_directory, "column_names"))
else:
print("_kale_resource_save: `column_names` not found.")
if "schema" in locals():
_kale_resource_save(schema, os.path.join(
_kale_data_directory, "schema"))
else:
print("_kale_resource_save: `schema` not found.")
args = parser.parse_args()
logging.basicConfig(format='%(asctime)s - %(message)s',
level=logging.INFO,
datefmt='%d-%m-%y %H:%M:%S')
logging.info(f"Args: {args}")
stats1 = tfdv.load_stats_text(input_path=args.stats_file_1)
stats2 = tfdv.load_stats_text(input_path=args.stats_file_2)
schema1 = tfdv.infer_schema(statistics=stats1)
# Custom rules, tweak this as required. This is just an example
tfdv.get_feature(
schema1,
'I1').drift_comparator.jensen_shannon_divergence.threshold = 0.06
# Calculate drift between the reference stats stats1, and the statistics from new data in stats2
drift_anomalies = tfdv.validate_statistics(statistics=stats2,
schema=schema1,
previous_statistics=stats1)
# Convert the .pb2 to dict
drift = MessageToDict(drift_anomalies)
value = drift['driftSkewInfo'][0]['driftMeasurements'][0]['value']
threshold = drift['driftSkewInfo'][0]['driftMeasurements'][0]['threshold']
logging.info(
f"JS divergence value: {value}, and JS divergence threshold: {threshold}"
)
# Displaying all detected anomalies
# Integer larger than 10
# STRING type when expected INT type
# FLOAT type when expected INT type
# Integer smaller than 0
print(tfdv.display_anomalies(anomalies))
# New data WITH anomalies
test_set_copy = test_set.copy()
test_set_copy.drop("soot", axis=1, inplace=True)
# Statistics based on data with anomalies
test_set_copy_stats = tfdv.generate_statistics_from_dataframe(
dataframe=test_set_copy)
anomalies_new = tfdv.validate_statistics(statistics=test_set_copy_stats,
schema=schema)
print(tfdv.display_anomalies(anomalies_new))
# Prepare the schema for Serving,environment
schema.default_environment.append("TRAINING")
schema.default_environment.append("SERVING")
tfdv.get_feature(schema, "soot").not_in_environment.append("SERVING")
serving_env_anomalies = tfdv.validate_statistics(test_set_copy_stats,
schema,
environment="SERVING")
print(tfdv.display_anomalies(serving_env_anomalies))
# Freezing the schema
tfdv.write_schema_text(schema=schema, output_path='pollution_schema.pbtext')
anomalies = tfdv.validate_statistics(statistics=test_stats, schema=schema)
#display all detected anomalies in new data
anomalies = tfdv.validate_statistics(statistics=test_stats, schema=schema)
#Displayign all detected anomalies
tfdv.display_anomalies(anomalies)
#New data with anomalies
test_set_copy = test_set.copy()
test_set_copy.drop('soot', axis=1, inplace=True)
#stats based on data with anomalies
test_set_copy_stats = tfdv.generate_statistics_from_dataframe(dataframe=test_set_copy)
anomalies_new = tfdv.validate_statistics(statistics=test_set_copy_stats, schema=schema)
tfdv.display_anomalies(anomalies_new)
#Stage 5: prepare the schema for serving
schema.default_environment.append('TRAINING')
schema.default_environment.append('SERVING')
#Removing a target column from the Serving Schema
tfdv.get_feature(schema, 'soot').not_in_environment.append('SERVING')
#Checking for anomalies between serving environment and new test set
serving_env_anomalies = tfdv.validate_statistics(test_set_copy_stats, schema, environment='SERVING')
tfdv.display_anomalies(serving_env_anomalies)
def is_acceptable(train_batch, test_batch, method):
if method == 'baseline':
percent_null_range = determine_baseline_range(train_batch)
test_batch_percent_null, __ = percent_null_dataframes(
test_batch, test_batch)
if percent_null_range['min'].values[
0] <= test_batch_percent_null.values[0][
0] <= percent_null_range['max'].values[0]:
return True
else:
return False
elif method == 'original':
completeness_range, distinct_range = determine_acceptable_metric_range(
train_batch)
test_batch_completeness, __ = completeness_dataframes(
test_batch, test_batch)
test_batch_distinct, __ = distinct_counts_dataframes(
test_batch, test_batch)
completeness_within_range = 0
distinct_within_range = 0
for i in range(len(test_batch_distinct.values)):
if distinct_range[i:i + 1]['min'][0] <= test_batch_distinct.values[
i][0] <= distinct_range[i:i + 1]['max'][0]:
distinct_within_range += 1
if completeness_range[i:i + 1]['min'][
0] <= test_batch_completeness.values[i][
0] <= completeness_range[i:i + 1]['max'][0]:
completeness_within_range += 1
if completeness_within_range / test_batch_completeness.shape[
0] > .8 and distinct_within_range / test_batch_completeness.shape[
0] > .8:
return True
else:
return False
elif method == 'tfdv':
#set all flags (this is bc what we need to do with TFDV depends on the type of data we're dealing with)
flights = False
fb = False
dirty = False
#if this is flights data
if 'FLIGHTS' in test_batch[0]: flights = True
#if this is facebook data
if 'FBPosts' in test_batch[0]: fb = True
#if the test batch is a dirty file
if 'dirty' in test_batch[0]: dirty = True
######## now work on updating the schema #######
#update schema with all the files in the train batch
updated_schema = acceptable_schema(train_batch)
#make sure TFDV knows the columns that are only in the clean + not the dirty data
updated_schema.default_environment.append('TRAINING')
updated_schema.default_environment.append('TESTING')
#specify that 'for_key' feature is not in TESTING environment
if dirty and flights:
tfdv.get_feature(updated_schema,
'for_key').not_in_environment.append('TESTING')
#get the number of anomalies for the test batch prior to relaxing constraints
num_anomalies = get_num_anomalies(test_batch[0],
updated_schema,
environment='TESTING')
#don't need to do anything for the dirty facebook data (there are no extra columns to worry about)
else:
#if it's a clean file, default environment is TRAINING (or dirty facebook file since no extra cols)
num_anomalies = get_num_anomalies(test_batch[0], updated_schema)
#relax constraints for each type of data
if flights:
tfdv.get_feature(
updated_schema,
'ArrivalGate').distribution_constraints.min_domain_mass = 0.8
tfdv.get_feature(
updated_schema,
'DepartureGate').distribution_constraints.min_domain_mass = 0.8
if fb:
tfdv.get_feature(
updated_schema,
'outlet').distribution_constraints.min_domain_mass = 0.65
tfdv.get_feature(
updated_schema,
'domain').distribution_constraints.min_domain_mass = 0.6
#get anomalies after having relaxed constraints
if dirty and flights:
adjusted_num_anomalies = get_num_anomalies(test_batch[0],
updated_schema,
environment='TESTING')
else:
adjusted_num_anomalies = get_num_anomalies(test_batch[0],
updated_schema)
######## determine whether acceptable or not ########
#my definition of "acceptable" is just if the relaxed constraints had any affect on the anomalies
#being present (so relaxing constraints should yield less anomalies after adjustment)
if adjusted_num_anomalies < num_anomalies:
return True
else:
return False
elif method == 'exploratory':
#if flights data
if 'FLIGHTS' in test_batch[0]:
data = 'flight'
#if facebook data
if 'FBPosts' in test_batch[0]:
data = 'fb'
if data == 'flight':
test_batch_completeness, __ = completeness_dataframes(
test_batch, test_batch)
test_batch_distinct, __ = distinct_counts_dataframes(
test_batch, test_batch)
accepted = 0
not_accepted = 0
for i in range(len(test_batch_distinct.values)):
# values based off mean values for flights data of distinct values and completeness
if test_batch_distinct.values[i][
0] > 100 and test_batch_completeness.values[i][0] > 0.2:
not_accepted += 1 # if both values true than add to not_accepted
else:
accepted += 1 # if not then accepted
if not_accepted == 0:
return True
else:
return False
if data == 'fb':
test_batch_completeness, __ = completeness_dataframes(
test_batch, test_batch)
test_batch_distinct, __ = distinct_counts_dataframes(
test_batch, test_batch)
accepted = 0
not_accepted = 0
for i in range(len(test_batch_distinct.values)):
# values based off mean values for fb data of distinct values and completeness
if test_batch_distinct.values[i][
0] > 30 and test_batch_completeness.values[i][0] > 0.25:
not_accepted += 1 # if both values true than add to not_accepted
else:
accepted += 1 # if not then accepted
if not_accepted == 0:
return True
else:
return False
anomalies = tfdv.validate_statistics(statistics=test_stats, schema=schema)
# Display anomalies
tfdv.display_anomalies(anomalies)
"""Three features has inconsistent values, two of them has wrong type (net_manager, type_conn_perc), the purchase_are it is a feature expected to be outside of the train schema, it is due the fact correspond to a different region in Netherlands.
### Fixing the erros
The corrections will depend of our knowlegde of the data, looking at each feature and decide what action should be take, looking in the description of each feature will to decide how to correct the erros, **net_manager** is code of the regional network manager, **type_conn_perc: percentage** of presence of the principal type of connection in the zipcode range, **purchase_area** code of the area where the energy is purchased.
For the type_conn_perc let's covert it to BYTES and append the missing values to the purcha_are.
"""
# Convert the type_conn_perc to Bytes
type_conn_perc = tfdv.get_feature(schema, 'type_conn_perc')
type_conn_perc.type = 1 # Type 1 represent BYTES
# Add new value to the domain of feature net_manager.
purchase_area = tfdv.get_domain(schema, 'purchase_area')
purchase_area.value.append('Stedin')
purchase_area.value.append('Stedin Delfland')
purchase_area.value.append('Stedin Elektriciteit Zuid-Kennemerland')
purchase_area.value.append('Stedin Midden-Holland')
purchase_area.value.append('Stedin Schiedam')
purchase_area.value.append('Stedin Utrecht')
purchase_area.value.append('Stedin Weert')
# Validate eval stats after updating the schema
updated_anomalies = tfdv.validate_statistics(test_stats, schema)
tfdv.display_anomalies(updated_anomalies)
