def main():
print("Starting example")
mlops.init(run_in_non_pm_mode=True, mlops_mode=MLOpsMode.PYTHON)
# Line graphs
mlops.set_stat("myCounterDouble", 5.5)
mlops.set_stat("myCounterDouble2", 7.3)
# Multi-line graphs
mlt = MultiLineGraph().name("Multi Line").labels(["l1",
"l2"]).data([5, 16])
mlops.set_stat(mlt)
tbl = Table().name("MyTable").cols(["Date", "Some number"])
tbl.add_row(["2001Q1", "55"])
tbl.add_row(["2001Q2", "66"])
tbl.add_row(["2003Q3", "33"])
tbl.add_row(["2003Q2", "22"])
mlops.set_stat(tbl)
bar = BarGraph().name("MyBar").cols(["aa", "bb", "cc", "dd",
"ee"]).data([10, 15, 12, 9, 8])
mlops.set_stat(bar)
mlops.done()
print("Example done")
def main():
pm_options = parse_args()
# Initialize MLOps Library
mlops.init()
# Load the model
if pm_options.input_model is not None:
try:
filename = pm_options.input_model
file_obj = open(filename, 'rb')
mlops.set_stat("model_file", 1)
except Exception as e:
print("Model not found")
print("Got exception: {}".format(e))
mlops.set_stat("model_file", 0)
mlops.done()
return 0
classifier = pickle.load(file_obj)
# Create synthetic data (Gaussian Distribution, Poisson Distribution and Beta Distribution)
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
np.random.seed(0)
g = np.random.normal(0, 1, (num_samples, num_features))
p = np.random.poisson(0.7, (num_samples, num_features))
b = np.random.beta(2, 2, (num_samples, num_features))
test_data = np.concatenate((g, p, b), axis=0)
np.random.seed()
test_features = test_data[np.random.choice(test_data.shape[0],
num_samples,
replace=False)]
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data and compare it automatically with the ones
# reported during training to generate the similarity score.
mlops.set_data_distribution_stat(test_features)
# Output the number of samples being processed using MCenter
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT, num_samples,
st.TIME_SERIES)
# Predict labels
result = classifier.predict(test_features)
# Label distribution in prediction
value, counts = np.unique(result, return_counts=True)
label_distribution = np.asarray((value, counts)).T
column_names = value.astype(str).tolist()
print("Label distributions: \n {0}".format(label_distribution))
# Output label distribution as a BarGraph using MCenter
bar = BarGraph().name("Label Distribution").cols(
(label_distribution[:, 0]).astype(str).tolist()).data(
(label_distribution[:, 1]).tolist())
mlops.set_stat(bar)
# Terminate MLOPs
mlops.done()
def feature_importance(self,
model_obj,
feature_importance_vector=None,
feature_names=None,
model=None, df=None,
num_significant_features=100):
"""
present feature importance, either according to the provided vector or generated from
the provided model if available.
Feature importance bar graph is attached to the current model and can be fetched later for
this model.
this function implements:
1) use feature_importance_vector if exists
2) feature_names from the model if available
3) get feature names vector if exists
4) extract feature name from pipeline model or dataframe if exists -
(code different to pyspark and sklearn)
5) sort the vector.
6) take first k elements
7) create a bar graph for feature importance
:param model_obj: model object
:param feature_importance_vector: feature importance vector optional
:param feature_names: feature names vector optional
:param model: optional pipeline model for pyspark, sklearn model for python
:param df: optional dataframe for analysis
:param num_significant_features: Number of significant features
:raises: MLOpsException
"""
self._validate_feature_importance_inputs(feature_importance_vector, feature_names, model, df)
important_named_features = self._output_channel.feature_importance(feature_importance_vector, feature_names,
model, df)
if important_named_features:
# Sort the feature importance vector
important_named_features_sorted = sorted(important_named_features,
key=lambda x: x[1], reverse=True)
self._logger.info("Full important_named_features_sorted = {}"
.format(important_named_features_sorted))
# output k significant features
if int(num_significant_features) < len(important_named_features_sorted):
important_named_features_sorted = important_named_features_sorted[0:int(num_significant_features)]
# Plot results in a bar graph
self._logger.info("Important_named_features_sorted = {}"
.format(important_named_features_sorted))
col_names = [v[0] for i, v in enumerate(important_named_features_sorted)]
col_value = [v[1] for i, v in enumerate(important_named_features_sorted)]
bar = BarGraph().name("Feature Importance").cols(col_names).data(col_value)
model_obj.set_stat(bar)
def export_bar_table(bar_names, bar_data, title_name):
"""
This function provides a bar_graph for a bar type data at MCenter data scientist view
:param bar_names: Bar graph names
:param bar_data: Bar graph data.
:param title_name: Title of the bar Graph
:return:
"""
bar_graph_data = BarGraph().name(title_name).cols(
bar_names.astype(str).tolist()).data(bar_data.tolist())
mlops.set_stat(bar_graph_data)
def _report_bar_graph_metric(self, metric_meta, metrics):
cols = []
data = []
for related_m, bar_name in metric_meta.related_metric:
cols.append(bar_name)
data.append(metrics[related_m.metric_name])
if not all(v == 0
for v in data) or not metric_meta.metric_already_displayed:
metric_meta.metric_already_displayed = True
mlt = BarGraph().name(metric_meta.title).cols(cols).data(data)
mlops.set_stat(mlt)
def test_bar_graph():
pm.init(ctx=None, mlops_mode=MLOpsMode.STAND_ALONE)
with pytest.raises(MLOpsException):
BarGraph().name("bar").cols(["g1", "g2"]).data(["aa", "bb"])
with pytest.raises(MLOpsException):
BarGraph().name("bar").data(["aa", "bb"])
with pytest.raises(MLOpsException):
mlt = BarGraph().name("mlt").cols(["g1"]).data([55, 66])
pm.set_stat(mlt)
with pytest.raises(MLOpsException):
mlt_cont = BarGraph().name("mlt").cols([1,
2]).data([55,
66]).as_continuous()
pm.set_stat(mlt_cont)
mlt = BarGraph().name("mlt").cols(["g1", "g2"]).data([55, 66])
pm.set_stat(mlt)
mlt_cont = BarGraph().name("mlt").cols([1, 2,
3]).data([55, 66]).as_continuous()
pm.set_stat(mlt_cont)
pm.done()
def _materialize(self, parent_data_objs, user_data):
for param in parent_data_objs:
prent_param = "parent param is: {param}".format(param=param)
print(prent_param)
self._logger.info(prent_param)
for k,v in self._params.items():
params_info = "key: {key} ==> value: {value}".format(key=k, value=v)
print(params_info)
self._logger.info(params_info)
mlt = BarGraph().name("Kenshoo Bar graph example").cols(["bar", "bar2"]).data([1500, 2000])
mlops.set_stat(mlt)
return []
def _report_stats(self, file_path):
self._logger.info(" *** generate stats .. params:{}".format(
self._params))
self._logger.info(" *** Source file {}".format(file_path))
# Read the file
data = pd.read_csv(file_path, sep=' |,', header=None, skiprows=1)
data = data.rename(index=str,
columns={
1: "label",
2: "confidence0",
3: "confidence1"
})
prediction_distribution = data['label'].value_counts()
column_names = np.array(
prediction_distribution.index).astype(str).tolist()
# Initialize mlops
mlops.init()
# Report a bar graph
bar = BarGraph().name("Prediction Distribution").cols(
np.array(prediction_distribution.index).astype(str).tolist()).data(
prediction_distribution.values.tolist())
mlops.set_stat(bar)
# Generate an alert on low confidence if the argument is set to true
if (self._params["alert"]):
index = data.values[:, 1].astype(int)
confidence = data.values[:, 2:4]
confidence_per_prediction = confidence[:, index][:, 0] * 100
low_conf_percent = len(confidence_per_prediction[
confidence_per_prediction < self._params["confidence"]]) / len(
confidence_per_prediction) * 100
if low_conf_percent > self._params["samples"]:
msg = "Low confidence: {}% of inferences had confidence below {}%".format(
low_conf_percent, self._params["confidence"])
print(msg)
mlops.health_alert("Low confidence alert", msg)
mlops.done()
return []
def __init__(self, track_conf, conf_thresh, conf_percent, output):
self._track_conf = track_conf
self._conf_thresh = conf_thresh
self._conf_percent = conf_percent
self._output_low_confidence_predictions = output
self._low_confidence_predictions = 0
print("track_conf: {}".format(track_conf))
if track_conf > 0:
print("conf_thresh: {}".format(conf_thresh))
print("conf_percent: {}".format(conf_percent))
categories = [
"10", "20", "30", "40", "50", "60", "70", "80", "90", "100"
]
self._conf_hist = []
for i in range(0, 10):
self._conf_hist.append(0)
## MLOps start
self._conf_graph = BarGraph().name(
"Confidence Distribution Bar Graph").cols(categories)
def __init__(self,
print_interval,
stats_type,
num_categories,
conf_thresh,
conf_percent,
hot_label=True):
super(CategoricalStatistics, self).__init__(print_interval)
self._num_categories = num_categories
self._hot_label = hot_label
self._stats_type = stats_type
self._conf_thresh = conf_thresh / 100.0
self._conf_percent = conf_percent
# These are useful for development, but should be replaced by mlops library functions
self._label_hist = []
self._infer_hist = []
for i in range(0, self._num_categories):
self._label_hist.append(0)
self._infer_hist.append(0)
if self._stats_type == "python":
mlops.init(ctx=None,
connect_mlops=True,
mlops_mode=MLOpsMode.AGENT)
elif self._stats_type == "file":
mlops.init(ctx=None,
connect_mlops=False,
mlops_mode=MLOpsMode.STAND_ALONE)
else:
self._stats_type = "none"
if self._stats_type != "none":
column_names = ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
self._infer_tbl = Table().name("categories").cols(column_names)
self._infer_bar = BarGraph().name("categories bar").cols(
column_names)
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: # Sample: [{}]".format(
pm_options.num_samples))
print("PM: # Features: [{}]".format(
pm_options.num_features))
print("PM: C: [{}]".format(pm_options.C))
print("PM: Kernel: [{}]".format(pm_options.kernel))
print("PM: Degree: [{}]".format(pm_options.degree))
print("PM: Gamma: [{}]".format(pm_options.gamma))
print("PM: Tolerance: [{}]".format(pm_options.tol))
print("PM: Maximum iterations: [{}]".format(pm_options.max_iter))
print("PM: Output model: [{}]".format(
pm_options.output_model))
# Initialize MLOps Library
mlops.init()
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
# Create synthetic data using scikit learn
X, y = make_classification(n_samples=num_samples,
n_features=num_features,
n_informative=2,
n_redundant=1,
n_classes=3,
n_clusters_per_class=1,
random_state=42)
# Separate into features and labels
features = X
labels = y
# Add noise to the data
noisy_features = np.random.uniform(0, 10) * \
np.random.normal(0, 1,
(num_samples, num_features))
features = features + noisy_features
# Create a model that should be deployed into production
final_model = SVC(C=float(pm_options.C),
kernel=pm_options.kernel,
degree=int(pm_options.degree),
gamma=str(pm_options.gamma),
tol=float(pm_options.tol),
max_iter=int(pm_options.max_iter))
final_model.fit(features, labels)
# Accuracy for the chosen model
accuracy = final_model.score(features, labels)
print("Accuracy values: \n {0}".format(accuracy))
# Label distribution in training
value, counts = np.unique(labels, return_counts=True)
label_distribution = np.asarray((value, counts)).T
column_names = value.astype(str).tolist()
print("Label distributions: \n {0}".format(label_distribution))
# Output label distribution as a BarGraph using MCenter
bar = BarGraph().name("Label Distribution").cols(
(label_distribution[:, 0]).astype(str).tolist()).data(
(label_distribution[:, 1]).tolist())
mlops.set_stat(bar)
# Output accuracy of the chosen model using MCenter
mlops.set_stat("Accuracy", accuracy, st.TIME_SERIES)
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data
mlops.set_data_distribution_stat(features)
# Save the model
import pickle
model_file = open(pm_options.output_model, 'wb')
pickle.dump(final_model, model_file)
model_file.close()
# Terminate MLOPs
mlops.done()
def main():
pm_options = parse_args()
# Initialize MLOps Library
mlops.init()
# Load the model
if pm_options.input_model is not None:
try:
filename = pm_options.input_model
file_obj = open(filename, 'rb')
mlops.set_stat("model_file", 1)
except Exception as e:
print("Model not found")
print("Got exception: {}".format(e))
mlops.set_stat("model_file", 0)
mlops.done()
return 0
regression = pickle.load(file_obj)
# Create synthetic data (Gaussian Distribution, Poisson Distribution and Beta Distribution)
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
mae_threshold = float(pm_options.threshold)
# Create synthetic data using scikit learn
X, y = make_regression(n_samples=num_samples,
n_features=num_features,
n_informative=2,
random_state=42)
# for making labels all positive
y = y + -1 * np.min(y)
# Separate into features and labels
features = X
labels = y
# Add noise to the data
noisy_features = np.random.uniform(0, 10) * \
np.random.normal(0, 1,
(num_samples, num_features))
features = features + noisy_features
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data and compare it automatically with the ones
# reported during training to generate the similarity score.
mlops.set_data_distribution_stat(features)
# Output the number of samples being processed using MCenter
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT, num_samples,
st.TIME_SERIES)
# Predict labels
labels_pred = regression.predict(features)
hist_pred, bin_edges_pred = np.histogram(labels_pred)
# Output prediction label distribution as a BarGraph using MCenter
pred_label_bar = BarGraph().name("User Defined: Prediction Label Distribution") \
.cols(bin_edges_pred.astype(str).tolist()) \
.data(hist_pred.tolist()) \
.as_continuous()
mlops.set_stat(pred_label_bar)
##########################################################################
#################### Start: Output Sample/Conversions ####################
########################################################################@@
mae = np.absolute(labels_pred - labels)
conversions = sum(i < mae_threshold for i in mae)
samples = num_samples
mlops.set_stat("samples", samples)
mlops.set_stat("conversions", conversions)
########################################################################
#################### End: Output Sample/Conversions ####################
########################################################################
# Terminate MLOPs
mlops.done()
def infer_loop(model, input, output_file, stats_interval, conf_tracker):
output = open(output_file, "w")
# Initialize statistics
total_predictions = 0
low_confidence_predictions = 0
categories = ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
prediction_hist = []
for i in range(0, model.get_num_categories()):
prediction_hist.append(0)
### MLOPS start
# Create a bar graph and table for reporting prediction distributions and set the column names
infer_bar = BarGraph().name("Prediction Distribution Bar Graph").cols(
categories)
infer_tbl = Table().name("Prediction Distribution Table").cols(categories)
### MLOPS end
while True:
try:
sample, label = input.get_next_input()
# Get the inference. This is an array of probabilities for each output value.
inference = model.infer(sample)
# The prediction is the class with the highest probability
prediction = ny.argmax(inference)
# The confidence for that prediction
confidence = inference[prediction] * 100
# Append the prediction to the output file
output.write("{}\n".format(prediction))
# Calculate statistics
total_predictions += 1
prediction_hist[prediction] += 1
conf_tracker.check_confidence(confidence, sample)
# Report statistics
if total_predictions % stats_interval == 0:
# Report the prediction distribution
for i in range(0, model.get_num_categories()):
print("category: {} predictions: {}".format(
categories[i], prediction_hist[i]))
### MLOPS start
# Update total prediction count with the all new predictions since we last reported
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT,
stats_interval)
# Show the prediction distribution as a table
infer_tbl.add_row(str(total_predictions), prediction_hist)
# Show the prediction distribution as a bar graph
infer_bar.data(prediction_hist)
# Report the stats
mlops.set_stat(infer_tbl)
mlops.set_stat(infer_bar)
### MLOPS end
conf_tracker.report_confidence(stats_interval)
except EOFError:
# stop when we hit end of input
print("Reached end of input")
output.close()
break
def infer_loop(model, input, output_file, stats_interval, conf_thresh, conf_percent):
output = open(output_file, "w")
# Initialize statistics
total_predictions = 0
low_confidence_predictions = 0
categories = ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
prediction_hist = []
for i in range(0, len(categories)):
prediction_hist.append(0)
### MLOPS start
# Create a bar graph and table for reporting prediction distributions and set the column names
infer_bar = BarGraph().name("Prediction Distribution Bar Graph").cols(categories)
infer_tbl = Table().name("Prediction Distribution Table").cols(categories)
### MLOPS end
while True:
try:
sample, label = input.get_next_input()
sample_np = ny.array(sample).reshape(1, -1)
# The prediction is the class with the highest probability
prediction = model.predict(sample_np)
# Append the prediction to the output file
output.write("{}\n".format(prediction))
# Calculate statistics
total_predictions += 1
prediction_hist[ny.int(prediction[0])] += 1
# Report statistics
if total_predictions % stats_interval == 0:
# Report the prediction distribution
for i in range(0, len(categories)):
print("category: {} predictions: {}".format(categories[i], prediction_hist[i]))
### MLOPS start
# Show the prediction distribution as a table
infer_tbl.add_row(str(total_predictions), prediction_hist)
# Show the prediction distribution as a bar graph
infer_bar.data(prediction_hist)
except EOFError:
# stop when we hit end of input
# Report the stats
mlops.set_stat(infer_tbl)
mlops.set_stat(infer_bar)
### MLOPS end
output.close()
### MLOPS start
mlops.done()
### MLOPS end
break
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: # Sample: [{}]".format(
pm_options.num_samples))
print("PM: # Features: [{}]".format(
pm_options.num_features))
print("PM: # KS Threshold: [{}]".format(
pm_options.ks_threshold))
print("PM: # PSI Threshold: [{}]".format(
pm_options.psi_threshold))
print("PM: # Input File: [{}]".format(
pm_options.input_file))
print("PM: # Model File: [{}]".format(
pm_options.input_model))
# Initialize MLOps Library
mlops.init()
# Load the model
if pm_options.input_model is not None:
try:
filename = pm_options.input_model
model_file_obj = open(filename, 'rb')
mlops.set_stat("# Model Files Used", 1)
except Exception as e:
print("Model Not Found")
print("Got Exception: {}".format(e))
mlops.set_stat("# Model Files Used", 0)
mlops.done()
return 0
final_model = pickle.load(model_file_obj)
try:
data_filename = pm_options.input_file
data_file_obj = open(data_filename, 'rb')
data = np.loadtxt(data_file_obj)
X = data # select columns 1 through end
except Exception as e:
print("Generating Synthetic Data Because {}".format(e))
# Create synthetic data (Gaussian Distribution, Poisson Distribution and Beta Distribution)
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
# Create synthetic data using scikit learn
X, y = make_classification(
n_samples=num_samples,
n_features=num_features,
# binary classification only!
n_classes=2,
random_state=42)
# Add random noise to the data randomly
import random
if random.randint(1, 21) / 2 == 0:
print("Adding Random Noise!")
noisy_features = np.random.uniform(0, 1) * \
np.random.normal(0, 1,
(num_samples, num_features))
X = X + noisy_features
# Separate into features and labels
features = X
max_ks_requirement = float(pm_options.ks_threshold)
min_psi_requirement = float(pm_options.psi_threshold)
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data and compare it automatically with the ones
mlops.set_data_distribution_stat(features)
# Output the number of samples being processed using MCenter
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT, len(features),
st.TIME_SERIES)
# Accuracy for the chosen model
pred_labels = final_model.predict(features)
pred_probs = final_model.predict_proba(features)
print("Pred Labels: ", pred_labels) # Remove printout can be huge
print("Pred Probabilities: ", pred_probs) # Remove printout can be huge
# Pred Label distribution
pred_value, pred_counts = np.unique(pred_labels, return_counts=True)
pred_label_distribution = np.asarray((pred_value, pred_counts)).T
# pred_column_names = pred_value.astype(str).tolist()
print("Pred Label distributions: \n {0}".format(pred_label_distribution))
# Output Pred label distribution as a BarGraph using MCenter
pred_bar = BarGraph().name("Pred Label Distribution").cols(
(pred_label_distribution[:, 0]).astype(str).tolist()).data(
(pred_label_distribution[:, 1]).tolist())
mlops.set_stat(pred_bar)
# Pred Label confidence per label
label_number = len(pred_counts)
average_confidence = np.zeros(label_number)
max_pred_probs = pred_probs.max(axis=1)
for i in range(0, label_number):
index_class = np.where(pred_labels == i)[0]
print(" np.sum(confidence[index_class])",
np.sum(max_pred_probs[index_class]))
print("counts_elements[i] ", pred_counts[i])
if pred_counts[i] > 0:
average_confidence[i] = np.sum(
max_pred_probs[index_class]) / (float(pred_counts[i]))
else:
average_confidence[i] = 0
# BarGraph showing confidence per class
pred_values1 = [str(i) for i in pred_value]
bar = BarGraph().name("Average Confidence Per Class").cols(
pred_values1).data(average_confidence.tolist())
mlops.set_stat(bar)
# KS for the chosen model
ks = ks_2samp(max_pred_probs[pred_labels == 1],
max_pred_probs[pred_labels == 0])
ks_stat = ks.statistic
ks_pvalue = ks.pvalue
print("KS values: \n Statistics: {} \n pValue: {}\n".format(
ks_stat, ks_pvalue))
# Output KS Stat of the chosen model using MCenter
if not np.isnan(ks_stat):
print("printing KS_stat ")
mlops.set_stat("KS Stat", ks_stat, st.TIME_SERIES)
else:
print("not printing KS_stat ")
# Raising alert if ks-stat goes above required threshold
if ks_stat >= max_ks_requirement:
mlops.health_alert(
"[Inference] KS Violation From Inference Node",
"KS Stat Went Above {}. Current KS Stat Is {}".format(
max_ks_requirement, ks_stat))
ks_table = Table().name("KS Stats").cols(["Statistic", "pValue"])
ks_table.add_row([ks_stat, ks_pvalue])
mlops.set_stat(ks_table)
# Calculating PSI
total_psi, psi_table = get_psi(max_pred_probs[pred_labels == 1],
max_pred_probs[pred_labels == 0])
psi_table_stat = Table().name("PSI Stats").cols([
"Base Pop", "Curr Pop", "Lower Bound", "Upper Bound", "Base Percent",
"Curr Percent", "Segment PSI"
])
row_num = 1
for each_value in psi_table.values:
str_values = [str(i) for i in each_value]
psi_table_stat.add_row(str(row_num), str_values)
row_num += 1
mlops.set_stat(psi_table_stat)
print("Total PSI values: \n {}".format(total_psi))
# Output Total PSI of the chosen model using MCenter
mlops.set_stat("Total PSI ", total_psi, st.TIME_SERIES)
# Raising alert if total_psi goes below required threshold
if total_psi <= min_psi_requirement:
mlops.health_alert(
"[Inference] PSI Violation From Inference Node",
"PSI Went Below {}. Current PSI Is {}".format(
min_psi_requirement, total_psi))
# Terminate MLOPs
mlops.done()
"l2"]).data([5, 16])
mlops.set_stat(mlt)
# Example of sending a table to pm system.
# Multi-line graphs
mlt = MultiLineGraph().name("Multi Line").labels(["l1",
"l2"]).data([5, 16])
mlops.set_stat(mlt)
# Table example
tbl = Table().name("MyTable").cols(["", "Date"])
tbl.add_row(["line 1", "2001Q1"])
tbl.add_row(["line 2", "2014Q3"])
mlops.set_stat(tbl)
bar = BarGraph().name("MyBar").cols(["aa", "bb", "cc", "dd",
"ee"]).data([10, 15, 12, 9, 8])
mlops.set_stat(bar)
partitions = int(sys.argv[1]) if len(sys.argv) > 1 else 2
n = 100000 * partitions
def f(_):
x = random() * 2 - 1
y = random() * 2 - 1
return 1 if x**2 + y**2 <= 1 else 0
count = spark.sparkContext.parallelize(range(1, n + 1),
partitions).map(f).reduce(add)
print("Pi is roughly %f" % (4.0 * count / n))
spark.stop()
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: # Sample: [{}]".format(
pm_options.num_samples))
print("PM: # Features: [{}]".format(
pm_options.num_features))
print("PM: # Validation Split: [{}]".format(
pm_options.validation_split))
print("PM: # AUC Threshold: [{}]".format(
pm_options.auc_threshold))
print("PM: # KS Threshold: [{}]".format(
pm_options.ks_threshold))
print("PM: # PSI Threshold: [{}]".format(
pm_options.psi_threshold))
print("PM: # Estimators: [{}]".format(
pm_options.n_estimators))
print("PM: # Max Depth: [{}]".format(pm_options.max_depth))
print("PM: # Learning Rate: [{}]".format(
pm_options.learning_rate))
print("PM: # Min Child Weight: [{}]".format(
pm_options.min_child_weight))
print("PM: # Objective: [{}]".format(pm_options.objective))
print("PM: # Gamma: [{}]".format(pm_options.gamma))
print("PM: # Max Delta Step: [{}]".format(
pm_options.max_delta_step))
print("PM: # Subsample: [{}]".format(pm_options.subsample))
print("PM: # Reg Alpha: [{}]".format(pm_options.reg_alpha))
print("PM: # Reg Lambda: [{}]".format(
pm_options.reg_lambda))
print("PM: # Scale Pos Weight: [{}]".format(
pm_options.scale_pos_weight))
print("PM: # Input File: [{}]".format(
pm_options.input_file))
print("PM: Output model: [{}]".format(
pm_options.output_model))
min_auc_requirement = float(pm_options.auc_threshold)
max_ks_requirement = float(pm_options.ks_threshold)
min_psi_requirement = float(pm_options.psi_threshold)
# Initialize MLOps Library
mlops.init()
try:
data_filename = pm_options.input_file
data_file_obj = open(data_filename, 'rb')
data = np.loadtxt(data_file_obj)
X = data[:, 1:] # select columns 1 through end
y = data[:, 0]
except Exception as e:
print("Generating Synthetic Data Because {}".format(e))
# Create synthetic data (Gaussian Distribution, Poisson Distribution and Beta Distribution)
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
# Create synthetic data using scikit learn
X, y = make_classification(
n_samples=num_samples,
n_features=num_features,
# binary classification only!
n_classes=2,
random_state=42)
print("Adding Random Noise!")
noisy_features = np.random.uniform(0, 1) * \
np.random.normal(0, 1,
(num_samples, num_features))
X = X + noisy_features
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=float(pm_options.validation_split), random_state=42)
import xgboost as xgb
# Create a model that should be deployed into production
final_model = xgb.XGBClassifier(
max_depth=int(pm_options.max_depth),
min_child_weight=int(pm_options.min_child_weight),
learning_rate=float(pm_options.learning_rate),
n_estimators=int(pm_options.n_estimators),
silent=True,
objective=str(pm_options.objective),
gamma=float(pm_options.gamma),
max_delta_step=int(pm_options.max_delta_step),
subsample=float(pm_options.subsample),
colsample_bytree=1,
colsample_bylevel=1,
reg_alpha=float(pm_options.reg_alpha),
reg_lambda=float(pm_options.reg_lambda),
scale_pos_weight=float(pm_options.scale_pos_weight),
seed=1,
missing=None)
final_model.fit(X_train, y_train)
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data
mlops.set_data_distribution_stat(X_train)
# Accuracy for the chosen model
pred_labels = final_model.predict(X_test)
pred_probs = final_model.predict_proba(X_test)
print("Pred Labels: ", pred_labels)
print("Pred Probabilities: ", pred_probs)
accuracy = accuracy_score(y_test, pred_labels)
print("Accuracy values: \n {0}".format(accuracy))
# Output accuracy of the chosen model using MCenter
mlops.set_stat("Accuracy", accuracy, st.TIME_SERIES)
# Label distribution in training
value, counts = np.unique(y_test, return_counts=True)
label_distribution = np.asarray((value, counts)).T
# column_names = value.astype(str).tolist()
print("Validation Actual Label distributions: \n {0}".format(
label_distribution))
# Output Label distribution as a BarGraph using MCenter
bar = BarGraph().name("Validation Actual Label Distribution").cols(
(label_distribution[:, 0]).astype(str).tolist()).data(
(label_distribution[:, 1]).tolist())
mlops.set_stat(bar)
# Pred Label distribution in training
pred_value, pred_counts = np.unique(pred_labels, return_counts=True)
pred_label_distribution = np.asarray((pred_value, pred_counts)).T
# pred_column_names = pred_value.astype(str).tolist()
print("Validation Prediction Label Distributions: \n {0}".format(
pred_label_distribution))
# Output Pred label distribution as a BarGraph using MCenter
pred_bar = BarGraph().name(
"Validation Prediction Label Distributions").cols(
(pred_label_distribution[:, 0]).astype(str).tolist()).data(
(pred_label_distribution[:, 1]).tolist())
mlops.set_stat(pred_bar)
# ROC for the chosen model
roc_auc = roc_auc_score(y_test, pred_probs[:, 1])
print("ROC AUC values: \n {}".format(roc_auc))
# Output ROC of the chosen model using MCenter
mlops.set_stat("ROC AUC", roc_auc, st.TIME_SERIES)
if roc_auc <= min_auc_requirement:
mlops.health_alert(
"[Training] AUC Violation From Training Node",
"AUC Went Below {}. Current AUC Is {}".format(
min_auc_requirement, roc_auc))
# ROC Curve
fpr, tpr, thr = roc_curve(y_test, pred_probs[:, 1])
cg = MultiGraph().name(
"Receiver Operating Characteristic ").set_continuous()
cg.add_series(label='Random Curve ' '', x=fpr.tolist(), y=fpr.tolist())
cg.add_series(label='ROC Curve (Area = {0:0.2f})'
''.format(roc_auc),
x=fpr.tolist(),
y=tpr.tolist())
cg.x_title('False Positive Rate')
cg.y_title('True Positive Rate')
mlops.set_stat(cg)
max_pred_probs = pred_probs.max(axis=1)
# KS for the chosen model
ks = ks_2samp(max_pred_probs[y_test == 1], max_pred_probs[y_test == 0])
ks_stat = ks.statistic
ks_pvalue = ks.pvalue
print("KS values: \n Statistics: {} \n pValue: {}\n".format(
ks_stat, ks_pvalue))
# Output KS Stat of the chosen model using MCenter
mlops.set_stat("KS Stat", ks_stat, st.TIME_SERIES)
# Raising alert if ks-stat goes above required threshold
if ks_stat >= max_ks_requirement:
mlops.health_alert(
"[Training] KS Violation From Training Node",
"KS Stat Went Above {}. Current KS Stat Is {}".format(
max_ks_requirement, ks_stat))
ks_table = Table().name("KS Stats").cols(["Statistic", "pValue"])
ks_table.add_row([ks_stat, ks_pvalue])
mlops.set_stat(ks_table)
# Calculating PSI
total_psi, psi_table = get_psi(max_pred_probs[y_test == 1],
max_pred_probs[y_test == 0])
psi_table_stat = Table().name("PSI Stats").cols([
"Base Pop", "Curr Pop", "Lower Bound", "Upper Bound", "Base Percent",
"Curr Percent", "Segment PSI"
])
row_num = 1
for each_value in psi_table.values:
str_values = [str(i) for i in each_value]
psi_table_stat.add_row(str(row_num), str_values)
row_num += 1
mlops.set_stat(psi_table_stat)
print("Total PSI values: \n {}".format(total_psi))
# Output Total PSI of the chosen model using MCenter
mlops.set_stat("Total PSI ", total_psi, st.TIME_SERIES)
# Raising alert if total_psi goes below required threshold
if total_psi <= min_psi_requirement:
mlops.health_alert(
"[Training] PSI Violation From Training Node",
"PSI Went Below {}. Current PSI Is {}".format(
min_psi_requirement, total_psi))
# Save the model
import pickle
model_file = open(pm_options.output_model, 'wb')
pickle.dump(final_model, model_file)
model_file.close()
# Terminate MLOPs
mlops.done()
def canary_comparator(options, start_time, end_time, mode):
sc = None
if mode == RunModes.PYSPARK:
from pyspark import SparkContext
sc = SparkContext(appName="canary-comparator")
mlops.init(sc)
elif mode == RunModes.PYTHON:
mlops.init()
else:
raise Exception("Invalid mode " + mode)
not_enough_data = False
# Following are main and canary component names
main_prediction_component_name = options.nodeA
canary_prediction_component_name = options.nodeB
main_stat_name = options.predictionHistogramA
canary_stat_name = options.predictionHistogramB
main_agent = utils._get_agent_id(main_prediction_component_name,
options.agentA)
canary_agent = utils._get_agent_id(canary_prediction_component_name,
options.agentB)
if main_agent is None or canary_agent is None:
print("Invalid agent provided {} or {}".format(options.agentA,
options.agentB))
mlops.system_alert(
"PyException",
"Invalid Agent {} or {}".format(options.agentA, options.agentB))
return
try:
main_data_frame = mlops.get_stats(
name=main_stat_name,
mlapp_node=main_prediction_component_name,
agent=main_agent,
start_time=start_time,
end_time=end_time)
canary_data_frame = mlops.get_stats(
name=canary_stat_name,
mlapp_node=canary_prediction_component_name,
agent=canary_agent,
start_time=start_time,
end_time=end_time)
main_pdf = pd.DataFrame(main_data_frame)
canary_pdf = pd.DataFrame(canary_data_frame)
try:
row1 = main_pdf.tail(1).iloc[0]
row2 = canary_pdf.tail(1).iloc[0]
except Exception as e:
not_enough_data = True
print("Not enough histograms produced in pipelines")
raise ValueError("Not enough data to compare")
if row1['hist_type'] != row2['hist_type']:
raise ValueError(
'Canary and Main pipelines dont produce histograms' +
'of same type {} != {}'.format(row1['hist_type'],
row2['hist_type']))
if row1['hist_type'] == 'continuous':
rmse = _compare_cont_hist(row1['bin_edges'], row2['bin_edges'],
row1['hist_values'], row2['hist_values'])
gg2 = MultiGraph().name("Prediction Histograms").set_categorical()
gg2.x_title("Predictions")
gg2.y_title("Normalized Frequency")
gg2.add_series(label="Main",
x=[float(x) for x in row1['bin_edges']][:-1],
y=[y for y in row1['hist_values']])
gg2.add_series(label="Canary",
x=[float(x) for x in row2['bin_edges']][:-1],
y=[y for y in row2['hist_values']])
mlops.set_stat(gg2)
bar1 = BarGraph().name("Main Pipeline").cols([
"{} to {}".format(x, y)
for (x, y) in pairwise(row1['bin_edges'])
]).data([x for x in row1['hist_values']])
mlops.set_stat(bar1)
bar2 = BarGraph().name("Canary Pipeline").cols([
"{} to {}".format(x, y)
for (x, y) in pairwise(row2['bin_edges'])
]).data([x for x in row2['hist_values']])
mlops.set_stat(bar2)
elif row1['hist_type'] == 'categorical':
rmse = _compare_cat_hist(row1['bin_edges'], row2['bin_edges'],
row1['hist_values'], row2['hist_values'])
gg2 = MultiGraph().name("Prediction Histograms").set_categorical()
gg2.x_title("Predictions")
gg2.y_title("Normalized Frequency")
gg2.add_series(label="Main",
x=row1['bin_edges'],
y=[y for y in row1['hist_values']])
gg2.add_series(label="Canary",
x=row2['bin_edges'],
y=[y for y in row2['hist_values']])
mlops.set_stat(gg2)
bar1 = BarGraph().name("Main Pipeline").cols([
"{}".format(x) for x in row1['bin_edges']
]).data([x for x in row1['hist_values']])
mlops.set_stat(bar1)
bar2 = BarGraph().name("Canary Pipeline").cols([
"{}".format(x) for x in row2['bin_edges']
]).data([x for x in row2['hist_values']])
mlops.set_stat(bar2)
else:
raise ValueError('Invalid histogram type: {}'.format(
row1['hist_type']))
mlops.set_stat("RMSE", rmse, st.TIME_SERIES)
print("mlops policy {}".format(mlops.mlapp_policy))
if mlops.mlapp_policy.canary_threshold is None:
print("Canary health threshold not set")
raise ValueError("Canary health threshold not set in config")
# Following code perform comparison between the histograms.
# Here you can insert your own code
if rmse > mlops.mlapp_policy.canary_threshold:
print("Canary Alert {} > {}".format(
rmse, mlops.mlapp_policy.canary_threshold))
mlops.event(
CanaryAlert(label="CanaryAlert",
is_healthy=False,
score=rmse,
threshold=mlops.mlapp_policy.canary_threshold))
else:
print("Data matches {}".format(rmse))
mlops.event(
CanaryAlert(label="CanaryAlert",
is_healthy=True,
score=rmse,
threshold=mlops.mlapp_policy.canary_threshold))
except Exception as e:
if not_enough_data is False:
print("Got exception while getting stats: {}".format(e))
mlops.system_alert(
"PyException",
"Got exception {}".format(traceback.format_exc()))
if mode == RunModes.PYSPARK:
sc.stop()
mlops.done()
def main():
# Initialize spark and MLOps
spark = SparkSession.builder.appName("RandomForestClassifier").getOrCreate()
mlops.init(spark.sparkContext)
# parse the arguments to component
options = parse_args()
print("PM: Configuration:")
print("PM: Number of trees: [{}]".format(options.num_trees))
print("PM: Maximum depth: [{}]".format(options.max_depth))
print("PM: Output model: [{}]".format(options.output_model))
print("PM: Temp shared path: [{}]".format(options.temp_shared_path))
# Generate synthetic data using scikit learn
num_samples = 50
num_features = 20
num_classes = 3
X, y = make_classification(n_samples=num_samples, n_features=num_features, n_informative=2, n_redundant=1,
n_classes=num_classes, n_clusters_per_class=1, random_state=42)
X = X + np.random.uniform(0, 5) * np.random.normal(0, 1, (num_samples, num_features))
feature_names = ["".join(ascii_lowercase[a]) for a in range(num_features + 1)]
feature_names[0] = "label"
# Create a spark dataframe from the synthetic data generated
trainingData = spark.createDataFrame(
pd.DataFrame(np.concatenate((y.reshape(-1, 1), X), axis=1), columns=feature_names))
# Histogram of label distribution
value, counts = np.unique(y, return_counts=True)
label_distribution = np.asarray((value, counts)).T
column_names = value.astype(str).tolist()
print("Label distributions: \n {0}".format(label_distribution))
# Output label distribution as a BarGraph using MCenter
bar = BarGraph().name("Label Distribution").cols((label_distribution[:, 0]).astype(str).tolist()).data(
(label_distribution[:, 1]).tolist())
mlops.set_stat(bar)
# Output Health Statistics to MCenter
# Report features whose distribution should be compared during inference
mlops.set_data_distribution_stat(trainingData)
# Fit a random forest classifiction model
assembler = VectorAssembler(inputCols=feature_names[1:num_features + 1], outputCol="features")
layers = [num_features, 5, 4, num_classes]
classifier = RandomForestClassifier(numTrees=int(options.num_trees), maxDepth=int(options.max_depth))
pipeline = Pipeline(stages=[assembler, classifier])
model = pipeline.fit(trainingData)
predictions = model.transform(trainingData)
# Select (prediction, true label) and compute training error
evaluator = MulticlassClassificationEvaluator(
labelCol="label", predictionCol="prediction", metricName="accuracy")
accuracy = evaluator.evaluate(predictions)
# Report accuracy of the chosen model using MCenter
mlops.set_stat("Accuracy", accuracy, st.TIME_SERIES)
# Save the spark model
SparkPipelineModelHelper() \
.set_shared_context(spark_context=spark.sparkContext) \
.set_local_path(local_path=options.output_model) \
.set_shared_path_prefix(shared_path_prefix=options.temp_shared_path) \
.save_sparkml_model(model)
# Stop spark context and MLOps
spark.sparkContext.stop()
mlops.done()
def kmeans_train(pm_options, spark):
"""
Kmeans Training function
:param pm_options:
:param spark:
:return:
"""
# Import Data
##################################
input_data = (spark.read.format("csv").option(
"header", pm_options.with_headers).option(
"ignoreLeadingWhiteSpace",
"true").option("ignoreTrailingWhiteSpace", "true").option(
"inferschema",
"true").load(pm_options.data_file)).repartition(10)
column_names_all = input_data.columns
if not pm_options.with_headers == "true":
for col_index in range(0, len(column_names_all)):
input_data = input_data.withColumnRenamed(
column_names_all[col_index], 'c' + str(col_index))
input_data = input_data.cache()
input_train = input_data
input_test = input_data
# SparkML pipeline
##################################
exclude_cols = []
column_names = input_train.columns
input_col_names = []
for elmts in column_names:
ind = True
for excludes in exclude_cols:
if elmts == excludes:
ind = False
if ind:
input_col_names.append(elmts)
print(input_col_names)
vector_assembler = VectorAssembler(inputCols=input_col_names,
outputCol="features")
kmeans_pipe = KMeans(k=int(pm_options.K),
initMode="k-means||",
initSteps=2,
tol=1e-4,
maxIter=100,
featuresCol="features")
full_pipe = [vector_assembler, kmeans_pipe]
model_kmeans = Pipeline(stages=full_pipe).fit(input_train)
# Test validation and statistics collection
############################################################
predicted_df = model_kmeans.transform(input_test)
print("model_kmeans.stages(1) = ", model_kmeans.stages[1])
sum_errors = model_kmeans.stages[1].computeCost(predicted_df)
print("Sum of Errors for Kmeans = " + str(sum_errors))
# Shows the result.
kmeans_centers = model_kmeans.stages[1].clusterCenters()
print("Kmeans Centers: ")
for center in kmeans_centers:
print(center)
# calculating stats
############################################################
# Calculating Inter cluster distance
inter_cluster_distance = np.zeros(
(len(kmeans_centers), len(kmeans_centers)))
for centerIndex1 in range(0, len(kmeans_centers)):
for centerIndex2 in range(0, len(kmeans_centers)):
inter_cluster_distance[centerIndex1, centerIndex2] =\
eq_dist(kmeans_centers[centerIndex1], kmeans_centers[centerIndex2])
print("inter_cluster_distance = ", inter_cluster_distance)
# Calculating Intra cluster distances and the bars for the cluster distribution
intra_cluster_distance = np.zeros(len(kmeans_centers))
cluster_dist = np.zeros(len(kmeans_centers))
for centerIndex1 in range(0, len(kmeans_centers)):
filtered_df = predicted_df.filter(
predicted_df["prediction"] == centerIndex1)
cluster_dist[centerIndex1] = filtered_df.count()
if cluster_dist[centerIndex1] == 0:
intra_cluster_distance[centerIndex1] = 0
else:
filtered_df =\
filtered_df.withColumn('distance',
udf(eq_dist, FloatType())(col("features"),
array([lit(v) for v in kmeans_centers[centerIndex1]])))
intra_cluster_distance[centerIndex1] =\
filtered_df.agg(sum("distance")).first()[0] / cluster_dist[centerIndex1]
# calculating Davis-Boulding Index
############################################################
# R[i,j] = (S[i] + S[j])/M[i,j]
# D[i] = max(R[i,j]) for i !=j
# DB = (1/K) * sum(D[i])
r_index = np.zeros((len(kmeans_centers), len(kmeans_centers)))
for centerIndex1 in range(0, len(kmeans_centers)):
for centerIndex2 in range(0, len(kmeans_centers)):
r_index[centerIndex1, centerIndex2] = 0
if not inter_cluster_distance[centerIndex1, centerIndex2] == 0:
r_index[centerIndex1, centerIndex2] =\
(intra_cluster_distance[centerIndex1] + intra_cluster_distance[centerIndex2])\
/ inter_cluster_distance[centerIndex1, centerIndex2]
d_index = np.max(r_index, axis=0)
db_index = np.sum(d_index, axis=0) / len(kmeans_centers)
# pmml model generation
############################################################
pmml_file = toPMMLBytes(spark, input_train, model_kmeans).decode("UTF-8")
# PM stats
############################################################
print("Sum of Errors for Kmeans = " + str(sum_errors))
pm.set_stat("Sum of Errors for Kmeans", sum_errors, st.TIME_SERIES)
print("Davies-Bouldin index = " + str(db_index))
pm.set_stat("Davies-Bouldin index", db_index, st.TIME_SERIES)
# Tables
tbl_col_name = []
for j in range(0, len(kmeans_centers)):
tbl_col_name.append(str(j))
tbl = Table().name("Inter cluster distance").cols(tbl_col_name)
for j in range(0, len(kmeans_centers)):
tbl.add_row(
str(j) + ":", ["%.2f" % x for x in inter_cluster_distance[j, :]])
pm.set_stat(tbl)
tbl = Table().name("Intra cluster avg. distance").cols(tbl_col_name)
tbl.add_row("Distances:", ["%.2f" % x for x in intra_cluster_distance])
pm.set_stat(tbl)
tbl_col_name1 = []
for j in range(0, len(kmeans_centers[0])):
tbl_col_name1.append(str(j))
tbl = Table().name("Centers (for K<6, Attr<11)").cols(tbl_col_name1)
for j in range(0, len(kmeans_centers)):
tbl.add_row("center" + str(j) + ":",
["%.2f" % x for x in kmeans_centers[j]])
pm.set_stat(tbl)
# BarGraph
bar = BarGraph().name("Cluster Destribution").cols(tbl_col_name).data(
cluster_dist.tolist())
pm.stat(bar)
print("PM: generating histogram from data-frame and model")
print("PM:" + pmml_file)
try:
pm.set_data_distribution_stat(data=input_train, model=pmml_file)
print("PM: done generating histogram")
except Exception as e:
print("PM: failed to generate histogram using pm.stat")
print(e)
return pmml_file
def main():
# Initialize spark and MLOps
spark = SparkSession.builder.appName(
"RandomForestClassifier").getOrCreate()
mlops.init(spark.sparkContext)
# parse the arguments to component
options = parse_args()
# Load the model, exit gracefully if model is not found
try:
model_rf = \
SparkPipelineModelHelper() \
.set_shared_context(spark_context=spark.sparkContext) \
.set_local_path(local_path=options.input_model) \
.set_shared_path_prefix(shared_path_prefix=options.temp_shared_path) \
.load_sparkml_model()
except Exception as e:
print(e)
mlops.done()
spark.sparkContext.stop()
exit()
# Generate synthetic data for inference (Gaussian Distribution, Poisson Distribution and Beta Distribution)
num_samples = 50
num_features = 20
np.random.seed(0)
g = np.random.normal(0, 1, (num_samples, num_features))
p = np.random.poisson(0.7, (num_samples, num_features))
b = np.random.beta(2, 2, (num_samples, num_features))
test_data = np.concatenate((g, p, b), axis=0)
np.random.seed()
test_features = test_data[np.random.choice(test_data.shape[0],
num_samples,
replace=False)]
feature_names = [
"".join(ascii_lowercase[a]) for a in range(num_features + 1)
]
# Create a spark dataframe from the synthetic data generated
inferenceData = spark.createDataFrame(
pd.DataFrame(test_features, columns=feature_names[1:num_features + 1]))
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data and compare it automatically with the ones
# reported during training to generate the similarity score
mlops.set_data_distribution_stat(inferenceData)
num_samples = inferenceData.count()
# Report the number of samples being processed using MCenter
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT, num_samples,
st.TIME_SERIES)
# Make inference predictions
predicted_df = model_rf.transform(inferenceData)
# Create a bar graph with label and confidence distributions
histogram_predictions = predicted_df.groupby("prediction").count()
prediction_values = np.array(
histogram_predictions.select("prediction").collect())
prediction_counts = np.array(
histogram_predictions.select("count").collect())
# Report label distribution as a BarGraph using MCenter
bar_predictions = BarGraph().name("Prediction Distribution").cols(
(prediction_values[0]).astype(str).tolist()).data(
(prediction_counts[0]).tolist())
mlops.set_stat(bar_predictions)
# Stop spark context and MLOps
spark.sparkContext.stop()
mlops.done()
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: Data file: [{}]".format(pm_options.data_file))
print("PM: Output model: [{}]".format(pm_options.output_model))
print("PM: regularization_range: [{}]".format(
pm_options.regularization_range))
mlops.init()
# Read the Samsung datafile
dataset = pd.read_csv(pm_options.data_file)
# Separate into features and labels
features = dataset.iloc[:, 1:].values
labels = dataset.iloc[:, 0].values
# Hyper-parameter search using k-fold cross-validation
# Applying k_fold cross validation
regularization_range = pm_options.regularization_range.split(',')
regularization = [
float(regularization_var)
for regularization_var in regularization_range
]
tune_parameters = [{'C': regularization}]
# Initialize logistic regression algorithm
LR = LogisticRegression(class_weight='balanced',
multi_class='multinomial',
solver='lbfgs')
clf = GridSearchCV(LR, tune_parameters, cv=5, scoring='accuracy')
clf.fit(features, labels)
print("best parameter = ", clf.best_params_)
accuracy = clf.cv_results_['mean_test_score']
print(
'Accuracy values: \n {0} \n for `Regularization values: \n{1}'.format(
accuracy, regularization))
########## Start of ParallelM instrumentation ##############
# Report Hyper-parameter Table
tbl = Table().name("Hyper-parameter Search Results").cols(
["Mean accuracy from k-fold cross-validation"])
print("length of regularization", len(regularization))
index_max = np.argmax(accuracy)
for a in range(0, len(regularization)):
print("adding row", regularization[a])
if a == index_max:
tbl.add_row("[Best] Regularization = " + np.str(regularization[a]),
[accuracy[a]])
else:
tbl.add_row("Regularization = " + np.str(regularization[a]),
[accuracy[a]])
mlops.set_stat(tbl)
########## End of ParallelM instrumentation ##############
# Label distribution in training
label_distribution = dataset['label'].value_counts()
column_names = np.array(label_distribution.index).astype(str).tolist()
print("Label distributions: \n {0}".format(label_distribution))
########## Start of ParallelM instrumentation ##############
# Report label distribution as a BarGraph
bar = BarGraph().name("Label Distribution").cols(
np.array(label_distribution.index).astype(str).tolist()).data(
label_distribution.values.tolist())
mlops.set_stat(bar)
########## Start of ParallelM instrumentation ##############
#################### Start of ParallelM instrumentation ################
# Report accuracy of the chosen model
mlops.set_stat("K-fold cross-validation Accuracy", accuracy[index_max],
st.TIME_SERIES)
#################### End of ParallelM instrumentation ################
# Histogram input
mlops.set_data_distribution_stat(dataset)
# Save the model
import pickle
model_file = open(pm_options.output_model, 'wb')
pickle.dump(clf, model_file)
model_file.close()
mlops.done()
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: # Sample: [{}]".format(
pm_options.num_samples))
print("PM: # Features: [{}]".format(
pm_options.num_features))
print("PM: # Classes: [{}]".format(
pm_options.num_classes))
print("PM: C: [{}]".format(pm_options.C))
print("PM: Kernel: [{}]".format(pm_options.kernel))
print("PM: Degree: [{}]".format(pm_options.degree))
print("PM: Gamma: [{}]".format(pm_options.gamma))
print("PM: Tolerance: [{}]".format(pm_options.tol))
print("PM: Maximum iterations: [{}]".format(pm_options.max_iter))
print("PM: Output model: [{}]".format(
pm_options.output_model))
# Initialize MLOps Library
mlops.init()
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
num_classes = int(pm_options.num_classes)
# Create synthetic data using scikit learn
X, y = make_classification(n_samples=num_samples,
n_features=num_features,
n_informative=2,
n_redundant=1,
n_classes=num_classes,
n_clusters_per_class=1,
random_state=42)
# Separate into features and labels
features = X
labels = y
# Add noise to the data
noisy_features = np.random.uniform(0, 10) * \
np.random.normal(0, 1,
(num_samples, num_features))
features = features + noisy_features
# Create a model that should be deployed into production
final_model = SVC(C=float(pm_options.C),
probability=True,
kernel=pm_options.kernel,
degree=int(pm_options.degree),
gamma=str(pm_options.gamma),
tol=float(pm_options.tol),
max_iter=int(pm_options.max_iter))
final_model.fit(features, labels)
value, counts = np.unique(labels, return_counts=True)
label_distribution = np.asarray((value, counts)).T
# Output actual label distribution as a BarGraph using MCenter
bar = BarGraph().name("User Defined: Actual Label Distribution") \
.cols((label_distribution[:, 0]).astype(str).tolist()) \
.data((label_distribution[:, 1]).tolist())
mlops.set_stat(bar)
pos_label = 1
# calculate classification prediction
labels_pred = final_model.predict(features)
# calculate decision scores [n_sample, n_class]
labels_decision_score = final_model.decision_function(features)
# calculate classification probabilities [n_sample, n_class]
labels_prob = final_model.predict_proba(features)
# calculate classification probabilities of positive labels
label_pos_class_prob = list(map(lambda x: x[pos_label], labels_prob))
# list of sorted labels. i.e. [0, 1, 2, ..]
labels_ordered = sorted(set(labels))
value_pred, counts_pred = np.unique(labels_pred, return_counts=True)
label_distribution_pred = np.asarray((value_pred, counts_pred)).T
# Output prediction label distribution as a BarGraph using MCenter
bar_pred = BarGraph().name("User Defined: Prediction Label Distribution") \
.cols((label_distribution_pred[:, 0]).astype(str).tolist()) \
.data((label_distribution_pred[:, 1]).tolist())
mlops.set_stat(bar_pred)
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data
mlops.set_data_distribution_stat(features)
################################################################
#################### Start: Output Accuracy ####################
################################################################
accuracy = final_model.score(features, labels)
#################### OLD WAY ####################
# First Way
#
# # Output accuracy of the chosen model using MCenter
# mlops.set_stat("User Defined: Accuracy", accuracy, st.TIME_SERIES)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.ACCURACY_SCORE, accuracy)
# OR
# Third Way
mlops.metrics.accuracy_score(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
##############################################################
#################### End: Output Accuracy ####################
##############################################################
################################################################
#################### Start: Output AUC ####################
################################################################
fpr, tpr, thresholds = sklearn.metrics.roc_curve(labels,
labels_pred,
pos_label=pos_label)
auc = sklearn.metrics.auc(fpr, tpr)
#################### OLD WAY ####################
# First Way
#
# # Output auc of the chosen model using MCenter
# mlops.set_stat("User Defined: AUC", auc)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.AUC, auc)
# OR
# Third Way
mlops.metrics.auc(x=fpr, y=tpr)
#################### DONE NEW WAY ####################
##############################################################
#################### End: Output AUC ####################
##############################################################
###############################################################################
#################### Start: Output Average Precision Score ####################
###############################################################################
# average precision is not supported for multiclass
if len(labels_ordered) <= 2:
aps = sklearn.metrics.average_precision_score(labels,
labels_decision_score)
#################### OLD WAY ####################
# First Way
#
# # Output aps of the chosen model using MCenter
# mlops.set_stat("User Defined: Average Precision Score", aps)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.AVERAGE_PRECISION_SCORE, aps)
# OR
# Third Way
mlops.metrics.average_precision_score(y_true=labels,
y_score=labels_decision_score)
#################### DONE NEW WAY ####################
#############################################################################
#################### End: Output Average Precision Score ####################
#############################################################################
#########################################################################
#################### Start: Output Balanced Accuracy ####################
#########################################################################
bas = sklearn.metrics.balanced_accuracy_score(labels, labels_pred)
#################### OLD WAY ####################
# First Way
#
# # Output bas of the chosen model using MCenter
# mlops.set_stat("User Defined: Balanced Accuracy Score", bas)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.BALANCED_ACCURACY_SCORE, data=bas)
# OR
# Third Way
mlops.metrics.balanced_accuracy_score(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
#######################################################################
#################### End: Output Balanced Accuracy ####################
#######################################################################
########################################################################
#################### Start: Output Brier Score Loss ####################
########################################################################
bsl = sklearn.metrics.brier_score_loss(labels,
label_pos_class_prob,
pos_label=pos_label)
#################### OLD WAY ####################
# First Way
#
# # Output bsl of the chosen model using MCenter
# mlops.set_stat("User Defined: Brier Score Loss", bsl)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.BRIER_SCORE_LOSS, data=bsl)
# OR
# Third Way
mlops.metrics.brier_score_loss(y_true=labels,
y_prob=label_pos_class_prob,
pos_label=pos_label)
#################### DONE NEW WAY ####################
######################################################################
#################### End: Output Brier Score Loss ####################
######################################################################
#############################################################################
#################### Start: Output Classification Report ####################
#############################################################################
cr = sklearn.metrics.classification_report(labels, labels_pred)
print("Classification Report\n{}".format(cr))
#################### OLD WAY ####################
# First Way
#
# from parallelm.mlops.stats.table import Table
#
# arrayReport = list()
# for row in cr.split("\n"):
# parsed_row = [x for x in row.split(" ") if len(x) > 0]
# if len(parsed_row) > 0:
# arrayReport.append(parsed_row)
#
# header = arrayReport[0]
# cr_table = Table().name("User Defined: Classification Report").cols(header)
#
# for index in range(1, len(arrayReport)):
# row_title = arrayReport[index][0]
# row_value = arrayReport[index][:-1]
# cr_table.add_row(row_title, row_value)
#
# # output classification report using MCenter
# mlops.set_stat(cr_table)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.CLASSIFICATION_REPORT, data=cr)
# OR
# Third Way
mlops.metrics.classification_report(labels, labels_pred)
#################### DONE NEW WAY ####################
###########################################################################
#################### End: Output Classification Report ####################
###########################################################################
#########################################################################
#################### Start: Output Cohen Kappa Score ####################
#########################################################################
cks = sklearn.metrics.cohen_kappa_score(labels, labels_pred)
#################### OLD WAY ####################
# First Way
#
# # Output cks of the chosen model using MCenter
# mlops.set_stat("User Defined: Cohen Kappa Score", cks)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.COHEN_KAPPA_SCORE, data=cks)
# OR
# Third Way
mlops.metrics.cohen_kappa_score(labels, labels_pred)
#################### DONE NEW WAY ####################
#######################################################################
#################### End: Output Cohen Kappa Score ####################
#######################################################################
########################################################################
#################### Start: Output Confusion Matrix ####################
########################################################################
cm = sklearn.metrics.confusion_matrix(labels,
labels_pred,
labels=labels_ordered)
#################### OLD WAY ####################
# First Way
# from parallelm.mlops.stats.table import Table
# labels_string = [str(i) for i in labels_ordered]
# cm_matrix = Table().name("User Defined: Confusion Matrix").cols(labels_string)
#
# for index in range(len(cm)):
# cm_matrix.add_row(labels_string[index], list(cm[index]))
#
# mlops.set_stat(cm_matrix)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.CONFUSION_MATRIX,
cm,
labels=labels_ordered)
# OR
# Third Way
mlops.metrics.confusion_matrix(y_true=labels,
y_pred=labels_pred,
labels=labels_ordered)
#################### DONE NEW WAY ####################
######################################################################
#################### End: Output Confusion Matrix ####################
######################################################################
################################################################
#################### Start: Output F1 Score ####################
################################################################
f1 = sklearn.metrics.f1_score(labels,
labels_pred,
pos_label=pos_label,
average=None)
#################### OLD WAY ####################
# First Way
#
# # Output f1 score of the chosen model using MCenter
# mlops.set_stat("User Defined: F1 Score", f1)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.F1_SCORE, data=f1)
# OR
# Third Way
mlops.metrics.f1_score(labels,
labels_pred,
pos_label=pos_label,
average=None)
#################### DONE NEW WAY ####################
##############################################################
#################### End: Output F1 Score ####################
##############################################################
################################################################
#################### Start: Output FBeta Score ####################
################################################################
fbeta = sklearn.metrics.fbeta_score(labels,
labels_pred,
beta=0.5,
average=None)
#################### OLD WAY ####################
# First Way
#
# # Output fbeta score of the chosen model using MCenter
# mlops.set_stat("User Defined: F-beta Score", fbeta)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.FBETA_SCORE, data=fbeta)
# OR
# Third Way
mlops.metrics.fbeta_score(labels,
labels_pred,
pos_label=pos_label,
beta=0.5,
average=None)
#################### DONE NEW WAY ####################
#################################################################
#################### End: Output FBeta Score ####################
#################################################################
####################################################################
#################### Start: Output Hamming Loss ####################
####################################################################
hamming_loss = sklearn.metrics.hamming_loss(labels, labels_pred)
#################### OLD WAY ####################
# First Way
#
# # Output hamming loss of the chosen model using MCenter
# mlops.set_stat("User Defined: Hamming Loss", hamming_loss)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.HAMMING_LOSS, data=hamming_loss)
# OR
# Third Way
mlops.metrics.hamming_loss(labels, labels_pred)
#################### DONE NEW WAY ####################
##################################################################
#################### End: Output Hamming Loss ####################
##################################################################
##################################################################
#################### Start: Output Hinge Loss ####################
##################################################################
hinge_loss = sklearn.metrics.hinge_loss(labels, labels_decision_score)
#################### OLD WAY ####################
# First Way
#
# # Output hinge loss of the chosen model using MCenter
# mlops.set_stat("User Defined: Hinge Loss", hinge_loss)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.HINGE_LOSS, data=hinge_loss)
# OR
# Third Way
mlops.metrics.hinge_loss(labels, labels_decision_score)
#################### DONE NEW WAY ####################
################################################################
#################### End: Output Hinge Loss ####################
################################################################
##############################################################################
#################### Start: Output Jaccard Similarity Score ####################
##############################################################################
jaccard_sim_score = sklearn.metrics.jaccard_similarity_score(
labels, labels_pred)
#################### OLD WAY ####################
# First Way
#
# # Output jaccard similarity score of the chosen model using MCenter
# mlops.set_stat("User Defined: Jaccard Similarity Score", jaccard_sim_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.JACCARD_SIMILARITY_SCORE,
data=jaccard_sim_score)
# OR
# Third Way
mlops.metrics.jaccard_similarity_score(labels, labels_pred)
#################### DONE NEW WAY ####################
############################################################################
#################### End: Output Jaccard Similary Score ####################
############################################################################
################################################################
#################### Start: Output Log Loss ####################
################################################################
log_loss = sklearn.metrics.log_loss(labels, labels_prob)
#################### OLD WAY ####################
# First Way
#
# # Output log loss of the chosen model using MCenter
# mlops.set_stat("User Defined: Log Loss", log_loss)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.LOG_LOSS, data=log_loss)
# OR
# Third Way
mlops.metrics.log_loss(labels, labels_prob)
#################### DONE NEW WAY ####################
##############################################################
#################### End: Output Log Loss ####################
##############################################################
########################################################################################
#################### Start: Output Matthews Correlation Coefficient ####################
########################################################################################
mcc = sklearn.metrics.matthews_corrcoef(labels, labels_pred)
#################### OLD WAY ####################
# First Way
#
# # Output mcc of the chosen model using MCenter
# mlops.set_stat("User Defined: Matthews Correlation Coefficient", mcc)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.MATTHEWS_CORRELATION_COEFFICIENT,
data=mcc)
# OR
# Third Way
mlops.metrics.matthews_corrcoef(labels, labels_pred)
#################### DONE NEW WAY ####################
######################################################################################
#################### End: Output Matthews Correlation Coefficient ####################
######################################################################################
##############################################################################
#################### Start: Output Precision Recall Curve ####################
##############################################################################
# precision_recall_curve is not supported for multiclass
if len(labels_ordered) <= 2:
precision, recall, thresholds = sklearn.metrics.precision_recall_curve(
labels, labels_decision_score, pos_label=pos_label)
classes = len(labels_ordered)
average_precision = sklearn.metrics.average_precision_score(
labels, labels_decision_score, average="macro")
graph_label_str = "{}-class Precision Recall Curve -- AP: {}".format(
classes, average_precision)
#################### OLD WAY ####################
# First Way
# from parallelm.mlops.stats.graph import Graph
#
# p_r_curve = Graph() \
# .name("User Defined: Precision Recall Curve") \
# .set_x_series(list(recall)) \
# .add_y_series(label="User Defined: {}".format(graph_label_str), data=list(precision))
#
# p_r_curve.x_title("Recall")
# p_r_curve.y_title("Precision")
# mlops.set_stat(p_r_curve)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.PRECISION_RECALL_CURVE,
[precision, recall],
legend=graph_label_str)
# OR
# Third Way
mlops.metrics.precision_recall_curve(y_true=labels,
probas_pred=labels_decision_score,
pos_label=pos_label,
average="macro")
#################### DONE NEW WAY ####################
############################################################################
#################### End: Output Precision Recall Curve ####################
############################################################################
#######################################################################
#################### Start: Output Precision Score ####################
#######################################################################
precision_score = sklearn.metrics.precision_score(labels,
labels_pred,
pos_label=pos_label,
average=None)
#################### OLD WAY ####################
# First Way
#
# # Output precision score of the chosen model using MCenter
# mlops.set_stat("User Defined: Precision Score", precision_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.PRECISION_SCORE, data=precision_score)
# OR
# Third Way
mlops.metrics.precision_score(labels,
labels_pred,
pos_label=pos_label,
average=None)
#################### DONE NEW WAY ####################
############################################################################
#################### End: Output Precision Score ###########################
############################################################################
####################################################################
#################### Start: Output Recall Score ####################
####################################################################
recall_score = sklearn.metrics.recall_score(labels,
labels_pred,
pos_label=pos_label,
average=None)
#################### OLD WAY ####################
# First Way
#
# # Output recall score of the chosen model using MCenter
# mlops.set_stat("User Defined: Recall Score", recall_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.RECALL_SCORE, data=recall_score)
# OR
# Third Way
mlops.metrics.recall_score(labels,
labels_pred,
pos_label=pos_label,
average=None)
#################### DONE NEW WAY ####################
#########################################################################
#################### End: Output Recall Score ###########################
#########################################################################
#####################################################################
#################### Start: Output ROC AUC Score ####################
#####################################################################
# roc_auc_score is not supported for multiclass
if len(labels_ordered) <= 2:
roc_auc_score = sklearn.metrics.roc_auc_score(labels,
labels_decision_score)
#################### OLD WAY ####################
# First Way
#
# # Output roc auc score of the chosen model using MCenter
# mlops.set_stat("User Defined: ROC AUC Score", roc_auc_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.ROC_AUC_SCORE, data=roc_auc_score)
# OR
# Third Way
mlops.metrics.roc_auc_score(labels, labels_decision_score)
#################### DONE NEW WAY ####################
###################################################################
#################### End: Output ROC AUC Score ####################
###################################################################
#################################################################
#################### Start: Output ROC Curve ####################
#################################################################
# roc_auc_score is not supported for multiclass
if len(labels_ordered) <= 2:
fpr, tpr, thresholds = sklearn.metrics.roc_curve(labels,
labels_decision_score,
pos_label=pos_label)
roc_auc_score = sklearn.metrics.roc_auc_score(labels,
labels_decision_score)
graph_label_str = "ROC Curve, AUC: {}".format(roc_auc_score)
#################### OLD WAY ####################
# First Way
# from parallelm.mlops.stats.graph import Graph
#
# roc_curve = Graph() \
# .name("User Defined: ROC Curve") \
# .set_x_series(list(fpr)) \
# .add_y_series(label="User Defined: {}".format(graph_label_str), data=list(tpr))
#
# roc_curve.x_title("False Positive Rate")
# roc_curve.y_title("True Positive Rate")
#
# mlops.set_stat(roc_curve)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
mlops.set_stat(ClassificationMetrics.ROC_CURVE, [tpr, fpr],
legend=graph_label_str)
# OR
# Third Way
mlops.metrics.roc_curve(y_true=labels,
y_score=labels_decision_score,
pos_label=pos_label)
#################### DONE NEW WAY ####################
###############################################################
#################### End: Output ROC Curve ####################
###############################################################
#####################################################################
#################### Start: Output Zero One Loss ####################
#####################################################################
zol = sklearn.metrics.zero_one_loss(labels, labels_pred)
#################### OLD WAY ####################
# First Way
#
# # Output zol of the chosen model using MCenter
# mlops.set_stat("User Defined: Zero One Loss", zol)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClassificationMetrics.ZERO_ONE_LOSS, data=zol)
# OR
# Third Way
mlops.metrics.zero_one_loss(labels, labels_pred)
#################### DONE NEW WAY ####################
###################################################################
#################### End: Output Zero One Loss ####################
###################################################################
# Save the model
import pickle
model_file = open(pm_options.output_model, 'wb')
pickle.dump(final_model, model_file)
model_file.close()
# Terminate MLOPs
mlops.done()
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: # Sample: [{}]".format(
pm_options.num_samples))
print("PM: # Features: [{}]".format(
pm_options.num_features))
print("PM: # Classes: [{}]".format(
pm_options.num_cluster))
print("PM: Init: [{}]".format(pm_options.init))
print("PM: N Init: [{}]".format(pm_options.n_init))
print("PM: Tolerance: [{}]".format(pm_options.tol))
print("PM: Maximum Iterations: [{}]".format(pm_options.max_iter))
print("PM: Pre-Compute Distances: [{}]".format(
pm_options.precompute_distances))
print("PM: Algorithm: [{}]".format(pm_options.algorithm))
print("PM: Output model: [{}]".format(
pm_options.output_model))
# Initialize MLOps Library
mlops.init()
n_samples = int(pm_options.num_samples)
n_features = int(pm_options.num_features)
n_clusters = int(pm_options.num_cluster)
init = str(pm_options.init)
n_init = int(pm_options.n_init)
max_iter = int(pm_options.max_iter)
tol = float(pm_options.tol)
precompute_distances = str(pm_options.precompute_distances)
algorithm = str(pm_options.algorithm)
verbose = 0
n_jobs = 1
# Create synthetic data using scikit learn
X, y = make_classification(n_samples=n_samples,
n_features=n_features,
n_informative=10,
n_redundant=1,
n_classes=n_clusters,
n_clusters_per_class=1,
random_state=42)
# Separate into features and labels
features = X
labels_true = y
# Add noise to the data
noisy_features = np.random.uniform(0, 10) * \
np.random.normal(0, 1,
(n_samples, n_features))
features = features + noisy_features
kmeans_model = KMeans(n_clusters=n_clusters,
init=init,
n_init=n_init,
max_iter=max_iter,
tol=tol,
precompute_distances=precompute_distances,
verbose=verbose,
random_state=None,
copy_x=True,
n_jobs=n_jobs,
algorithm=algorithm).fit(features, labels_true)
mlops.set_stat("User Defined: Training Inertia", kmeans_model.inertia_)
mlops.set_stat("User Defined: Training Iteration", kmeans_model.n_iter_)
value, counts = np.unique(labels_true, return_counts=True)
label_distribution = np.asarray((value, counts)).T
# Output actual label distribution as a BarGraph using MCenter
bar_true = BarGraph().name("User Defined: Actual Label Distribution") \
.cols((label_distribution[:, 0]).astype(str).tolist()) \
.data((label_distribution[:, 1]).tolist())
mlops.set_stat(bar_true)
# prediction labels
labels_pred = kmeans_model.predict(features)
value_pred, counts_pred = np.unique(labels_pred, return_counts=True)
label_distribution_pred = np.asarray((value_pred, counts_pred)).T
# Output prediction label distribution as a BarGraph using MCenter
bar_pred = BarGraph().name("User Defined: Prediction Label Distribution") \
.cols((label_distribution_pred[:, 0]).astype(str).tolist()) \
.data((label_distribution_pred[:, 1]).tolist())
mlops.set_stat(bar_pred)
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data
mlops.set_data_distribution_stat(features)
###########################################################################
#################### Start: Adjusted Mutual Info Score ####################
###########################################################################
adjusted_mutual_info_score = sklearn.metrics \
.adjusted_mutual_info_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Adjusted Mutual Info Score", adjusted_mutual_info_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.ADJUSTED_MUTUAL_INFO_SCORE,
adjusted_mutual_info_score)
# OR
# Third Way
mlops.metrics.adjusted_mutual_info_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### DONE NEW WAY ####################
#########################################################################
#################### End: Adjusted Mutual Info Score ####################
#########################################################################
####################################################################
#################### Start: Adjusted Rand Score ####################
####################################################################
adjusted_rand_score = sklearn.metrics \
.adjusted_rand_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Adjusted Rand Score", adjusted_rand_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.ADJUSTED_RAND_SCORE, adjusted_rand_score)
# OR
# Third Way
mlops.metrics.adjusted_rand_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### DONE NEW WAY ####################
##################################################################
#################### End: Adjusted Rand Score ####################
##################################################################
#######################################################################
#################### Start: Calinski Harabaz Score ####################
#######################################################################
calinski_harabaz_score = sklearn.metrics \
.calinski_harabaz_score(X=features, labels=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Calinski Harabaz Score", calinski_harabaz_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.CALINSKI_HARABAZ_SCORE,
calinski_harabaz_score)
# OR
# Third Way
mlops.metrics.calinski_harabaz_score(X=features, labels=labels_pred)
#################### DONE NEW WAY ####################
#####################################################################
#################### End: Calinski Harabaz Score ####################
#####################################################################
###################################################################
#################### Start: Completeness Score ####################
###################################################################
completeness_score = sklearn.metrics \
.completeness_score(labels_true=labels_true, labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Completeness Score", completeness_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.COMPLETENESS_SCORE, completeness_score)
# OR
# Third Way
mlops.metrics.completeness_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### DONE NEW WAY ####################
#################################################################
#################### End: Completeness Score ####################
#################################################################
###################################################################
#################### Start: Contingency Matrix ####################
###################################################################
contingency_matrix = sklearn.metrics.cluster \
.contingency_matrix(labels_true, labels_pred)
# list of sorted labels. i.e. [0, 1, 2, ..]
pred_labels_list = sorted(set(labels_pred))
true_labels_list = sorted(set(labels_true))
#################### OLD WAY ####################
# First Way
# from parallelm.mlops.stats.table import Table
#
# cm_cols_ordered_string = [str(i) for i in pred_labels_list]
# cm_rows_ordered_string = [str(i) for i in true_labels_list]
# cm_matrix = Table().name("User Defined: Contingency Matrix").cols(cm_cols_ordered_string)
#
# for index in range(len(contingency_matrix)):
# cm_matrix.add_row(cm_rows_ordered_string[index], list(contingency_matrix[index]))
#
# mlops.set_stat(cm_matrix)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.CONTINGENCY_MATRIX,
data=contingency_matrix,
true_labels=true_labels_list,
pred_labels=pred_labels_list)
# OR
# Third Way
mlops.metrics.cluster.contingency_matrix(labels_true, labels_pred)
#################### DONE NEW WAY ####################
#################################################################
#################### End: Contingency Matrix ####################
#################################################################
######################################################################
#################### Start: Fowlkes Mallows Score ####################
######################################################################
fowlkes_mallows_score = \
sklearn.metrics.fowlkes_mallows_score(labels_true=labels_true,
labels_pred=labels_pred,
sparse=False)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Fowlkes Mallows Score", fowlkes_mallows_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.FOWLKES_MALLOWS_SCORE,
fowlkes_mallows_score)
# OR
# Third Way
mlops.metrics.fowlkes_mallows_score(labels_true=labels_true,
labels_pred=labels_pred,
sparse=False)
#################### DONE NEW WAY ####################
####################################################################
#################### End: Fowlkes Mallows Score ####################
####################################################################
#####################################################################################
#################### Start: Homogeneity, Completeness, V Measure ####################
#####################################################################################
homogeneity, completeness, v_measure = sklearn.metrics \
.homogeneity_completeness_v_measure(labels_true=labels_true, labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# multiline_object = MultiLineGraph() \
# .name("User Defined: Homogeneity - Completeness - V Measure") \
# .labels(["Homogeneity", "Completeness", "V Measure"])
#
# multiline_object.data([homogeneity, completeness, v_measure])
#
# mlops.set_stat(multiline_object)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.HOMOGENEITY_COMPLETENESS_V_MEASURE,
data=[homogeneity, completeness, v_measure])
# OR
# Third Way
mlops.metrics \
.homogeneity_completeness_v_measure(labels_true=labels_true, labels_pred=labels_pred)
#################### DONE NEW WAY ####################
###################################################################################
#################### End: Homogeneity, Completeness, V Measure ####################
###################################################################################
##################################################################
#################### Start: Homogeneity Score ####################
##################################################################
homogeneity_score = sklearn.metrics \
.homogeneity_score(labels_true=labels_true, labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Homogeneity Score", homogeneity_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.HOMOGENEITY_SCORE, homogeneity_score)
# OR
# Third Way
mlops.metrics \
.homogeneity_score(labels_true=labels_true, labels_pred=labels_pred)
#################### DONE NEW WAY ####################
################################################################
#################### End: Homogeneity Score ####################
################################################################
##################################################################
#################### Start: Mutual Info Score ####################
##################################################################
mutual_info_score = sklearn.metrics \
.mutual_info_score(labels_true=labels_true, labels_pred=labels_pred, contingency=None)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Mutual Info Score", mutual_info_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.MUTUAL_INFO_SCORE, mutual_info_score)
# OR
# Third Way
mlops.metrics \
.mutual_info_score(labels_true=labels_true, labels_pred=labels_pred, contingency=None)
#################### DONE NEW WAY ####################
################################################################
#################### End: Mutual Info Score ####################
################################################################
#############################################################################
#################### Start: Normalized Mutual Info Score ####################
#############################################################################
normalized_mutual_info_score = sklearn.metrics \
.normalized_mutual_info_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Normalized Mutual Info Score", normalized_mutual_info_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.NORMALIZED_MUTUAL_INFO_SCORE,
normalized_mutual_info_score)
# OR
# Third Way
mlops.metrics \
.normalized_mutual_info_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### DONE NEW WAY ####################
###########################################################################
#################### End: Normalized Mutual Info Score ####################
###########################################################################
#################################################################
#################### Start: Silhouette Score ####################
#################################################################
silhouette_score = sklearn.metrics \
.silhouette_score(X=features, labels=labels_pred, metric="euclidean", sample_size=None, random_state=None)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: Silhouette Score", silhouette_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.SILHOUETTE_SCORE, silhouette_score)
# OR
# Third Way
mlops.metrics \
.silhouette_score(X=features, labels=labels_pred, metric="euclidean", sample_size=None, random_state=None)
#################### DONE NEW WAY ####################
###############################################################
#################### End: Silhouette Score ####################
###############################################################
################################################################
#################### Start: V Measure Score ####################
################################################################
v_measure_score = sklearn.metrics.v_measure_score(labels_true=labels_true,
labels_pred=labels_pred)
#################### OLD WAY ####################
# First Way
# mlops.set_stat("User Defined: V Measure Score", v_measure_score)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
mlops.set_stat(ClusteringMetrics.V_MEASURE_SCORE, v_measure_score)
# OR
# Third Way
mlops.metrics \
.v_measure_score(labels_true=labels_true, labels_pred=labels_pred)
#################### DONE NEW WAY ####################
##############################################################
#################### End: V Measure Score ####################
##############################################################
# Save the model
import pickle
model_file = open(pm_options.output_model, 'wb')
pickle.dump(kmeans_model, model_file)
model_file.close()
# Terminate MLOPs
mlops.done()
def main():
pm_options = parse_args()
print("PM: Configuration:")
print("PM: # Sample: [{}]".format(
pm_options.num_samples))
print("PM: # Features: [{}]".format(
pm_options.num_features))
print("PM: Kernel: [{}]".format(pm_options.kernel))
print("PM: Degree: [{}]".format(pm_options.degree))
print("PM: Gamma: [{}]".format(pm_options.gamma))
print("PM: Tolerance: [{}]".format(pm_options.tol))
print("PM: Maximum iterations: [{}]".format(pm_options.max_iter))
print("PM: Output model: [{}]".format(
pm_options.output_model))
# Initialize MLOps Library
mlops.init()
num_samples = int(pm_options.num_samples)
num_features = int(pm_options.num_features)
# Create synthetic data using scikit learn
X, y = make_regression(n_samples=num_samples,
n_features=num_features,
n_informative=2,
random_state=42)
# for making labels all positive
y = y + -1 * np.min(y)
# Separate into features and labels
features = X
labels = y
# Add noise to the data
noisy_features = np.random.uniform(0, 10) * \
np.random.normal(0, 1,
(num_samples, num_features))
features = features + noisy_features
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data
mlops.set_data_distribution_stat(features)
hist, bin_edges = np.histogram(labels)
# Output label distribution as a BarGraph using MCenter
bar = BarGraph().name("User Defined: Label Distribution") \
.cols((bin_edges).astype(str).tolist()) \
.data((hist).tolist()) \
.as_continuous()
mlops.set_stat(bar)
# Create a model that should be deployed into production
final_model = SVR(kernel=pm_options.kernel,
degree=int(pm_options.degree),
gamma=str(pm_options.gamma),
tol=float(pm_options.tol),
max_iter=int(pm_options.max_iter))
final_model.fit(features, labels)
labels_pred = final_model.predict(features)
hist_pred, bin_edges_pred = np.histogram(labels_pred)
# Output prediction label distribution as a BarGraph using MCenter
pred_label_bar = BarGraph().name("User Defined: Prediction Label Distribution") \
.cols((bin_edges_pred).astype(str).tolist()) \
.data((hist_pred).tolist()) \
.as_continuous()
mlops.set_stat(pred_label_bar)
y_true = labels
y_pred = labels_pred
# Regression Metrics
##########################################################################
#################### Start: Output Explained Variance ####################
##########################################################################
evs = sklearn.metrics.explained_variance_score(y_true, y_pred)
#################### OLD WAY ####################
# First Way
mlops.set_stat("User Defined: Explained Variance ", evs)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
# mlops.set_stat(RegressionMetrics.EXPLAINED_VARIANCE_SCORE, evs)
# OR
# Third Way
# mlops.metrics.explained_variance_score(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
########################################################################
#################### End: Output Explained Variance ####################
########################################################################
######################################################################
#################### Start: Output Mean Abs Error ####################
######################################################################
mae = sklearn.metrics.mean_absolute_error(y_true, y_pred)
#################### OLD WAY ####################
# First Way
mlops.set_stat("User Defined: Mean Abs Error", mae)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
# mlops.set_stat(RegressionMetrics.MEAN_ABSOLUTE_ERROR, mae)
# OR
# Third Way
# mlops.metrics.mean_absolute_error(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
####################################################################
#################### End: Output Mean Abs Error ####################
####################################################################
##########################################################################
#################### Start: Output Mean Squared Error ####################
##########################################################################
mse = sklearn.metrics.mean_squared_error(y_true, y_pred)
#################### OLD WAY ####################
# First Way
mlops.set_stat("User Defined: Mean Squared Error", mse)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
# mlops.set_stat(RegressionMetrics.MEAN_SQUARED_ERROR, mse)
# OR
# Third Way
# mlops.metrics.mean_squared_error(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
########################################################################
#################### End: Output Mean Squared Error ####################
########################################################################
##############################################################################
#################### Start: Output Mean Squared Log Error ####################
##############################################################################
msle = sklearn.metrics.mean_squared_log_error(y_true, y_pred)
#################### OLD WAY ####################
# First Way
mlops.set_stat("User Defined: Mean Squared Log Error", msle)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
# mlops.set_stat(RegressionMetrics.MEAN_SQUARED_LOG_ERROR, msle)
# OR
# Third Way
# mlops.metrics.mean_squared_log_error(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
############################################################################
#################### End: Output Mean Squared Log Error ####################
############################################################################
########################################################################
#################### Start: Output Median Abs Error ####################
########################################################################
median_ae = sklearn.metrics.median_absolute_error(y_true, y_pred)
#################### OLD WAY ####################
# First Way
mlops.set_stat("User Defined: Median Abs Error", median_ae)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
# mlops.set_stat(RegressionMetrics.MEDIAN_ABSOLUTE_ERROR, median_ae)
# OR
# Third Way
# mlops.metrics.median_absolute_error(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
######################################################################
#################### End: Output Median Abs Error ####################
######################################################################
################################################################
#################### Start: Output R2 Score ####################
################################################################
r2_s = sklearn.metrics.r2_score(y_true, y_pred)
#################### OLD WAY ####################
# First Way
mlops.set_stat("User Defined: R2 Score", r2_s)
#################### DONE OLD WAY ####################
#################### NEW WAY ####################
# Second Way
# mlops.set_stat(RegressionMetrics.R2_SCORE, r2_s)
# OR
# Third Way
# mlops.metrics.r2_score(y_true=labels, y_pred=labels_pred)
#################### DONE NEW WAY ####################
##############################################################
#################### End: Output R2 Score ####################
##############################################################
# Save the model
import pickle
model_file = open(pm_options.output_model, 'wb')
pickle.dump(final_model, model_file)
model_file.close()
# Terminate MLOPs
mlops.done()
def kmeans_train(pm_options, spark):
"""
Kmeans Training function
:param pm_options:
:param spark:
:return:
"""
# Import Data
##################################
input_data = (spark.read.format("csv")
.option("header", pm_options.with_headers)
.option("ignoreLeadingWhiteSpace", "true")
.option("ignoreTrailingWhiteSpace", "true")
.option("inferschema", "true")
.load(pm_options.data_file)).repartition(10)
# If Data doesn't have headers Create column names c0-cn
column_names_all = input_data.columns
if not pm_options.with_headers == "true":
for col_index in range(0, len(column_names_all)):
input_data = input_data.withColumnRenamed(column_names_all[col_index],
'c' + str(col_index))
input_data = input_data.cache()
# Set both train and tesst data to the entire dataset
input_train = input_data
input_test = input_data
# SparkML pipeline
##################################
# Create column names for vector assembler. Handle exclude columns for vector assembler
exclude_cols = [] # No columns to exclude - kmeans of all columns
column_names = input_train.columns
input_col_names = []
for elmts in column_names:
ind = True
for excludes in exclude_cols:
if elmts == excludes:
ind = False
if ind:
input_col_names.append(elmts)
print(input_col_names)
# Set hyper parameters search parameters
k_range = pm_options.KRange.split(',')
db_index_max = np.finfo(np.float64).max
k_max = k_range[0]
db_index_array = np.zeros(len(k_range))
for index_hs in range (0,len(k_range)):
vector_assembler = VectorAssembler(
inputCols=input_col_names,
outputCol="features")
kmeans_pipe = KMeans(
k=int(k_range[index_hs]),
initMode="k-means||",
initSteps=5,
tol=1e-4,
maxIter=100,
featuresCol="features")
full_pipe = [vector_assembler, kmeans_pipe]
model_kmeans = Pipeline(stages=full_pipe).fit(input_train)
# Test validation and statistics collection
############################################################
predicted_df = model_kmeans.transform(input_test)
print("model_kmeans.stages(1) = ", model_kmeans.stages[1])
sum_errors = model_kmeans.stages[1].computeCost(predicted_df)
print("Sum of Errors for Kmeans = " + str(sum_errors))
kmeans_centers = model_kmeans.stages[1].clusterCenters()
print("Kmeans Centers: ")
for center in kmeans_centers:
print(center)
# calculating stats
############################################################
# Calculating Inter cluster distance
inter_cluster_distance = np.zeros((len(kmeans_centers), len(kmeans_centers)))
for centerIndex1 in range(0, len(kmeans_centers)):
for centerIndex2 in range(0, len(kmeans_centers)):
inter_cluster_distance[centerIndex1, centerIndex2] = \
eq_dist(kmeans_centers[centerIndex1], kmeans_centers[centerIndex2])
print("inter_cluster_distance = ", inter_cluster_distance)
# Calculating Intra cluster distances and the bars for the cluster distribution
intra_cluster_distance = np.zeros(len(kmeans_centers))
cluster_dist = np.zeros(len(kmeans_centers))
for centerIndex1 in range(0, len(kmeans_centers)):
filtered_df = predicted_df.filter(predicted_df["prediction"] == centerIndex1)
cluster_dist[centerIndex1] = filtered_df.count()
if cluster_dist[centerIndex1] == 0:
intra_cluster_distance[centerIndex1] = 0
else:
filtered_df = \
filtered_df.withColumn('distance',
udf(eq_dist, FloatType())(col("features"),
array([lit(v) for v in kmeans_centers[centerIndex1]])))
intra_cluster_distance[centerIndex1] = \
filtered_df.agg(sum("distance")).first()[0] / cluster_dist[centerIndex1]
# calculating Davies-Boulding Index
############################################################
# R[i,j] = (S[i] + S[j])/M[i,j]
# D[i] = max(R[i,j]) for i !=j
# DB = (1/K) * sum(D[i])
r_index = np.zeros((len(kmeans_centers), len(kmeans_centers)))
for centerIndex1 in range(0, len(kmeans_centers)):
for centerIndex2 in range(0, len(kmeans_centers)):
r_index[centerIndex1, centerIndex2] = 0
if not inter_cluster_distance[centerIndex1, centerIndex2] == 0:
r_index[centerIndex1, centerIndex2] = \
(intra_cluster_distance[centerIndex1] + intra_cluster_distance[centerIndex2]) \
/ inter_cluster_distance[centerIndex1, centerIndex2]
d_index = np.max(r_index, axis=0)
db_index = np.sum(d_index, axis=0) / len(kmeans_centers)
db_index_array[index_hs] = db_index
# Check Hyper Parameter Search max
if (db_index < db_index_max):
db_index_max = db_index
k_max = k_range[index_hs]
model_kmeans_max = model_kmeans
sum_errors_max = sum_errors
kmeans_centers_max = kmeans_centers
inter_cluster_distance_max = inter_cluster_distance
intra_cluster_distance_max = intra_cluster_distance
cluster_dist_max = cluster_dist
# PM stats
############################################################
print("Optimal K = " + str(k_max))
pm.set_stat("Optimal number of clusters", k_max, st.TIME_SERIES)
print("Sum of Errors for Kmeans = " + str(sum_errors_max))
pm.set_stat("Sum of Errors for Kmeans", sum_errors_max, st.TIME_SERIES)
print("Davies-Bouldin index = " + str(db_index_max))
pm.set_stat("Davies-Bouldin index", db_index_max, st.TIME_SERIES)
# Tables
tbl_col_name = []
for j in range(0, len(k_range)):
tbl_col_name.append(str(k_range[j]))
tbl = Table().name("Davies-Bouldin index for hyper parameter Search").cols(tbl_col_name)
tbl.add_row("Davies-Bouldin index:", ["%.2f" % x for x in db_index_array])
pm.set_stat(tbl)
tbl_col_name = []
for j in range(0, len(kmeans_centers_max)):
tbl_col_name.append(str(j))
tbl = Table().name("Inter cluster distance").cols(tbl_col_name)
for j in range(0, len(kmeans_centers_max)):
tbl.add_row(str(j) + ":", ["%.2f" % x for x in inter_cluster_distance_max[j, :]])
pm.set_stat(tbl)
tbl = Table().name("Intra cluster avg. distance").cols(tbl_col_name)
tbl.add_row("Distances:", ["%.2f" % x for x in intra_cluster_distance_max])
pm.set_stat(tbl)
if (len(kmeans_centers_max) < 6) & (len(kmeans_centers_max[0]) < 12):
tbl_col_name1 = []
for j in range(0, len(kmeans_centers_max[0])):
tbl_col_name1.append(str(j))
tbl = Table().name("Centers (for K<6, Attr<12)").cols(tbl_col_name1)
for j in range(0, len(kmeans_centers_max)):
tbl.add_row("center" + str(j) + ":", ["%.2f" % x for x in kmeans_centers_max[j]])
pm.set_stat(tbl)
# BarGraph
bar = BarGraph().name("Cluster Destribution").cols(tbl_col_name).data(cluster_dist_max.tolist())
pm.set_stat(bar)
return model_kmeans_max
def main():
pm_options = parse_args()
mlops.init()
# Load the model
if pm_options.input_model is not None:
try:
filename = pm_options.input_model
file_obj = open(filename, 'rb')
mlops.set_stat("model_file", 1)
except Exception as e:
print("Model not found")
print("Got exception: {}".format(e))
mlops.set_stat("model_file", 0)
mlops.done()
return 0
classifier = pickle.load(file_obj)
# Load the data
test_dataset = pd.read_csv(pm_options.input_file)
mlops.set_data_distribution_stat(test_dataset)
# Extract numpy array
test_features = test_dataset.values
# Predict labels
result = classifier.predict(test_features)
# Predict probability
class_probability = classifier.predict_proba(test_features)
maximum_prob = np.max(class_probability, axis=1)
# Tag samples that are below a certain probability and write to a file
confidence = 0.8
low_prob_samples = test_features[np.where(maximum_prob < confidence)]
low_prob_predictions = result[np.where(maximum_prob < confidence)]
unique_elements_low, counts_elements_low = np.unique(low_prob_predictions,
return_counts=True)
unique_elements_low = [str(i) for i in unique_elements_low]
print("Low confidence predictions: \n {0} \n with frequency {1}".format(
unique_elements_low, counts_elements_low))
########## Start of ParallelM instrumentation ##############
# BarGraph showing distribution of low confidence labels
bar = BarGraph().name("Low confidence label distribution").cols(
unique_elements_low).data(counts_elements_low.tolist())
mlops.set_stat(bar)
########## End of ParallelM instrumentation ################
# Samples with high probability
high_prob_samples = test_features[np.where(maximum_prob >= confidence)]
high_prob_predictions = result[np.where(maximum_prob >= confidence)]
unique_elements_high, counts_elements_high = np.unique(
high_prob_predictions, return_counts=True)
unique_elements_high = [str(i) for i in unique_elements_high]
print("High confidence predictions: \n {0} \n with frequency {1}".format(
unique_elements_high, counts_elements_high))
########## Start of ParallelM instrumentation ##############
# BarGraph showing distribution of high confidence labels
bar = BarGraph().name("High confidence label distribution").cols(
unique_elements_high).data(counts_elements_high.tolist())
mlops.set_stat(bar)
########## End of ParallelM instrumentation ################
mlops.done()
def main(args):
# Parse arguments
parser = argparse.ArgumentParser()
add_parameters(parser)
args = parser.parse_args()
# Initialize MLOps Library
mlops.init()
# Create synthetic data (Gaussian Distribution, Poisson Distribution and Beta Distribution)
num_samples = 50
num_features = 20
np.random.seed(0)
g = np.random.normal(0, 1, (num_samples, num_features))
p = np.random.poisson(0.7, (num_samples, num_features))
b = np.random.beta(2, 2, (num_samples, num_features))
test_data = np.concatenate((g, p, b), axis=0)
np.random.seed()
features = test_data[np.random.choice(test_data.shape[0],
num_samples,
replace=False)]
# Start tensorflow session
sess = tf.InteractiveSession()
tag_set = ["serve"]
if args.model_dir is not None:
try:
print("args.model_dir = ", args.model_dir)
tf.saved_model.loader.load(sess, tag_set, args.model_dir)
except Exception as e:
print("Model not found")
print("Got exception: " + str(e))
return 0
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data and compare it automatically with the ones
# reported during training to generate the similarity score.
mlops.set_data_distribution_stat(data=features)
# Output the number of samples being processed using MCenter
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT, len(features))
graph = tf.get_default_graph()
x = graph.get_tensor_by_name("features:0")
y_pred = graph.get_tensor_by_name("predictions:0")
predictions = sess.run(y_pred, {x: features})
print('predictions', np.array(predictions))
# Ouput prediction distribution as a BarGraph using MCenter
predict_int = np.argmax(predictions, axis=1)
unique, counts = np.unique(predict_int, return_counts=True)
counts = list(map(int, counts))
x_series = list(map(str, unique))
mlt = BarGraph().name("Prediction Distribution").cols(x_series).data(
list(counts))
mlops.set_stat(mlt)
# Show average prediction probability value for each prediction
num_labels = len(np.unique(predict_int))
probability = np.zeros((num_labels, ))
for a in range(0, num_labels):
temp = predictions[np.argmax(predictions, axis=1) == a, :]
print(temp)
probability[a] = np.mean(temp[:, a])
print("probability", list(np.squeeze(probability)))
# Plot average probability in each class using MCenter
bg = BarGraph().name("Probability of Each Label").cols(x_series).data(
list(np.squeeze(probability)))
mlops.set_stat(bg)
def _prep_and_infer(self, df_dataset):
# Get number of features
self.num_features = df_dataset.shape[1]
# Get number of samples
self.num_samples = df_dataset.shape[0]
#get input model
self.input_model = self._params["input-model"]
self._logger.info("PM: Configuration:")
self._logger.info("PM: # Sample: [{}]".format(
self.num_samples))
self._logger.info("PM: # Features: [{}]".format(
self.num_features))
self._logger.info("PM: # Input-Model: [{}]".format(
self.input_model))
# Initialize MLOps Library
mlops.init()
# Load the model
if self.input_model is not None:
try:
filename = self._params["input-model"]
model_file_obj = open(filename, 'rb')
mlops.set_stat("# Model Files Used", 1)
except Exception as e:
#self._logger.error("Model Not Found")
self._logger.error("Got Exception: {}".format(e))
mlops.set_stat("# Model Files Used", 0)
mlops.done()
return 0
final_model = pickle.load(model_file_obj)
features = df_dataset
# Output Health Statistics to MCenter
# MLOps API to report the distribution statistics of each feature in the data
# and compare it automatically with the ones
mlops.set_data_distribution_stat(features)
# Output the number of samples being processed using MCenter
mlops.set_stat(PredefinedStats.PREDICTIONS_COUNT, len(features),
st.TIME_SERIES)
# Accuracy for the chosen model
pred_labels = final_model.predict(features)
pred_probs = final_model.predict_proba(features)
self._logger.info("Pred Labels: {}".format(
pred_labels)) # Remove printout can be huge
self._logger.info("Pred Probabilities: {}".format(
pred_probs)) # Remove printout can be huge
# Pred Label distribution
pred_value, pred_counts = np.unique(pred_labels, return_counts=True)
pred_label_distribution = np.asarray((pred_value, pred_counts)).T
# pred_column_names = pred_value.astype(str).tolist()
self._logger.info(
"Pred Label distributions: \n {}".format(pred_label_distribution))
# Output Pred label distribution as a BarGraph using MCenter
pred_bar = BarGraph().name("Pred Label Distribution").cols(
(pred_label_distribution[:, 0]).astype(str).tolist()).data(
(pred_label_distribution[:, 1]).tolist())
mlops.set_stat(pred_bar)
# Pred Label confidence per label
label_number = len(pred_counts)
average_confidence = np.zeros(label_number)
max_pred_probs = pred_probs.max(axis=1)
for i in range(0, label_number):
index_class = np.where(pred_labels == i)[0]
self._logger.info("np.sum(confidence[index_class]) {}".format(
np.sum(max_pred_probs[index_class])))
self._logger.info("counts_elements[i] {}".format(pred_counts[i]))
if pred_counts[i] > 0:
average_confidence[i] = np.sum(
max_pred_probs[index_class]) / (float(pred_counts[i]))
else:
average_confidence[i] = 0
# BarGraph showing confidence per class
pred_values1 = [str(i) for i in pred_value]
bar = BarGraph().name("Average Confidence Per Class").cols(
pred_values1).data(average_confidence.tolist())
mlops.set_stat(bar)
# Terminate MLOPs
mlops.done()
df_result = pd.concat([
df_dataset,
pd.DataFrame({'predict': pred_labels}),
pd.DataFrame({
'probs-0': pred_probs[:, 0],
'probs-1': pred_probs[:, 1]
})
],
axis=1)
df_result.insert(0,
'idx', [x for x in range(1, df_result.shape[0] + 1)],
allow_duplicates=False)
return df_result
def main():
# Parse arguments
parser = argparse.ArgumentParser()
add_parameters(parser)
args = parser.parse_args()
print("PM: Configuration:")
print("PM: Step size: [{}]".format(args.step_size))
print("PM: Iterations: [{}]".format(args.iterations))
print("PM: Model version: [{}]".format(args.model_version))
print("PM: Stats interval: [{}]".format(args.stats_interval))
print("PM: Save dir: [{}]".format(args.save_dir))
# Initialize MLOps Library
mlops.init()
# print the number of iteration used by optimization algorithm
print('Training for %i iterations' % args.iterations)
# Create sythetic data using scikit learn
num_samples = 50
num_features = 20
features, labels = make_classification(n_samples=50,
n_features=20,
n_informative=2,
n_redundant=1,
n_classes=3,
n_clusters_per_class=1,
random_state=42)
# Add noise to the data
noisy_features = np.random.uniform(0, 5) * np.random.normal(
0, 1, (num_samples, num_features))
features = features + noisy_features
num_features = (features.shape[1])
num_labels = len(np.unique(labels))
# One-hot encode labels for all data
onehot_labels = np.eye(num_labels)[labels]
# Label distribution in training
value, counts = np.unique(labels, return_counts=True)
label_distribution = np.asarray((value, counts)).T
column_names = value.astype(str).tolist()
print("Label distributions: \n {0}".format(label_distribution))
# Output label distribution as a BarGraph using MCenter
bar = BarGraph().name("Label Distribution").cols(
(label_distribution[:, 0]).astype(str).tolist()).data(
(label_distribution[:, 1]).tolist())
mlops.set_stat(bar)
# Output Health Statistics to MCenter
# Report features whose distribution should be compared during inference
mlops.set_data_distribution_stat(features)
# Algorithm parameters parsed from arguments
learning_rate = args.step_size
training_epochs = args.iterations
display_step = args.stats_interval
# tf Graph Input
x = tf.placeholder(tf.float32, [None, num_features], name="features")
y = tf.placeholder(tf.float32, [None, num_labels], name="labels")
# Set model weights
W = tf.Variable(tf.zeros([num_features, num_labels]))
b = tf.Variable(tf.zeros([num_labels]))
# Store values for saving model
serialized_tf_example = tf.placeholder(tf.string, name='tf_example')
# Construct model
pred = tf.nn.softmax(tf.matmul(x, W) + b, name="predictions") # Softmax
# Minimize error using cross entropy
cost = tf.reduce_mean(-tf.reduce_sum(y *
tf.log(pred), reduction_indices=1))
# Gradient Descent
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
# Evaluation
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(pred, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
# Start timer
training_start_time = time.time()
# Initialize the variables in a tf session
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
iteration_array = []
cost_array = []
accuracy_array = []
# Training cycle
for epoch in range(training_epochs):
avg_cost = 0
temp, c, a = sess.run([optimizer, cost, accuracy],
feed_dict={
x: features,
y: onehot_labels
})
# Compute average loss
avg_cost += c / num_samples
# Display logs per epoch step
if (epoch + 1) % display_step == 0:
iteration_array.append(epoch)
cost_array.append(avg_cost)
accuracy_array.append(np.float(a))
print("accuracy", a)
print("Epoch:", '%04d' % (epoch + 1), "cost=",
"{:.9f}".format(avg_cost))
# Plot the cost function using MCenter
gg = Graph().name("Cost function across epochs").set_x_series(
iteration_array).add_y_series(label="Cost Function Across Iterations",
data=cost_array)
gg.x_title("Average Cost")
gg.y_title('Iterations')
mlops.set_stat(gg)
# Plot the accuracy function using MCenter
gg1 = Graph().name("Accuracy across epochs").set_x_series(
iteration_array).add_y_series(label="Accuracy Across Iterations",
data=accuracy_array)
gg1.x_title("Accuracy")
gg1.y_title('Iterations')
mlops.set_stat(gg1)
# Plot accuracy and cost across epochs using MCenter
mg = MultiGraph().name("Cost and Accuracy Progress Across Epochs")
mg.add_series(x=iteration_array,
label="Cost Function Across Iterations",
y=cost_array)
mg.add_series(x=iteration_array,
label="Accuracy across epochs",
y=accuracy_array)
mlops.set_stat(mg)
# Plot final cost and accuracy in this session using MCenter
mlt = MultiLineGraph().name("Final Accuracy and Cost").labels(
["Cost", "Accuracy"])
mlt.data([cost_array[-1], accuracy_array[-1]])
mlops.set_stat(mlt)
# Save the model
export_path = args.save_dir
print('Exporting trained model to', export_path)
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
values, indices = tf.nn.top_k(y, num_labels)
table = tf.contrib.lookup.index_to_string_table_from_tensor(
tf.constant([str(i) for i in range(num_labels)]))
prediction_classes = table.lookup(tf.to_int64(indices))
# Build the signature_def_map.
classification_inputs = tf.saved_model.utils.build_tensor_info(
serialized_tf_example)
classification_outputs_classes = tf.saved_model.utils.build_tensor_info(
prediction_classes)
classification_outputs_scores = tf.saved_model.utils.build_tensor_info(
values)
classification_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={
tf.saved_model.signature_constants.CLASSIFY_INPUTS:
classification_inputs
},
outputs={
tf.saved_model.signature_constants.CLASSIFY_OUTPUT_CLASSES:
classification_outputs_classes,
tf.saved_model.signature_constants.CLASSIFY_OUTPUT_SCORES:
classification_outputs_scores
},
method_name=tf.saved_model.signature_constants.CLASSIFY_METHOD_NAME
))
tensor_info_x = tf.saved_model.utils.build_tensor_info(x)
tensor_info_y = tf.saved_model.utils.build_tensor_info(y)
prediction_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={'inputs': tensor_info_x},
outputs={'outputs': tensor_info_y},
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
)
legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
'predict_images':
prediction_signature,
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
classification_signature,
},
legacy_init_op=legacy_init_op)
builder.save(as_text=args.use_text)