Model Serving with KFServing and Scikit-learn - Iris Flower Classification
Model Serving with KFServing and Scikit-Learn - Iris Flower Classification
INPUT –> MODEL –> PREDICTION
This notebook requires KFServing to be installed
NOTE: It is assumed that a model called irisflowerclassifier is already available in Hopsworks. An example of training a model for the Iris flower classification problem is available in
Jupyter/end_to_end_pipelines/sklearn/end_to_end_sklearn.ipynb
Model Serving on Hopsworks
The hops
python library
hops
is a helper library for Hops that facilitates development by hiding the complexity of running applications and iteracting with services.
Have a feature request or encountered an issue? Please let us know on github.
Serve the Iris Flower classifier
Query Model Registry for best Iris Classifier Model
import hsml
conn = hsml.connection()
mr = conn.get_model_registry()
Connected. Call `.close()` to terminate connection gracefully.
MODEL_NAME="irisflowerclassifier"
EVALUATION_METRIC="accuracy"
best_model = mr.get_best_model(MODEL_NAME, EVALUATION_METRIC, "max")
print('Model name: ' + best_model.name)
print('Model version: ' + str(best_model.version))
print(best_model.training_metrics)
Model name: irisflowerclassifier
Model version: 1
{'accuracy': '0.98'}
Create Model Serving of Exported Model
from hops import serving
# Create serving instance
SERVING_NAME = MODEL_NAME
response = serving.create_or_update(SERVING_NAME, # define a name for the serving instance
model_path=best_model.model_path, # set the path of the model to be deployed
model_server="PYTHON", # set the model server to run the model
kfserving=True, # whether to serve the model using KFServing or the default tool in the current Hopsworks version
# optional arguments
model_version=best_model.version, # set the version of the model to be deployed
topic_name="CREATE", # (optional) set the topic name or CREATE to create a new topic for inference logging
inference_logging="ALL", # with KFServing, select the type of inference data to log into Kafka, e.g MODEL_INPUTS, PREDICTIONS or ALL
instances=1, # with KFServing, set 0 instances to leverage scale-to-zero capabilities
)
2022-01-17 15:14:07,102 INFO: Serving irisflowerclassifier successfully created
Once the serving instance is created, it will be shown in the “Model Serving” tab in the Hopsworks UI. You can view detailed information like server-logs and which Kafka Topic it is logging inference requests to.
You can also use the Python module to query the Hopsworks REST API about information on the existing servings using methods like:
get_all()
get_id(serving_name)
get_model_path(serving_name)
get_model_version(serving_name)
get_artifact_version(serving_name)
get_kafka_topic(serving_name)
...
print("Info: \tid: {},\n \
model_path: {},\n \
model_version: {},\n \
artifact_version: {},\n \
model_server: {},\n \
serving_tool: {}".format(
serving.get_id(SERVING_NAME),
serving.get_model_path(SERVING_NAME),
serving.get_model_version(SERVING_NAME),
serving.get_artifact_version(SERVING_NAME),
serving.get_model_server(SERVING_NAME),
serving.get_serving_tool(SERVING_NAME)))
Info: id: 2523,
model_path: /Projects/demo_ml_meb10000/Models/irisflowerclassifier,
model_version: 1,
artifact_version: 0,
model_server: PYTHON,
serving_tool: KFSERVING
for s in serving.get_all():
print(s.name)
irisflowerclassifier
Classify flowers with the Iris Flower classifier
Start Model Serving Server
if serving.get_status(SERVING_NAME) == 'Stopped':
serving.start(SERVING_NAME)
Starting serving with name: irisflowerclassifier...
Serving with name: irisflowerclassifier successfully started
import time
while serving.get_status(SERVING_NAME) != "Running":
time.sleep(5) # Let the serving startup correctly
time.sleep(10)
Check the Server Logs
You can access the server logs using Kibana by clicking on the ‘Show logs’ button in the action bar, and filter them using fields such as serving component (i.e., predictor or transformer) or container name among other things.
Send Prediction Requests to the Served Model using Hopsworks REST API
import json
import random
NUM_FEATURES = 4
for i in range(20):
data = {"instances" : [[random.uniform(1, 8) for i in range(NUM_FEATURES)]]}
response = serving.make_inference_request(SERVING_NAME, data)
print(response)
{'predictions': [2]}
{'predictions': [1]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [0]}
{'predictions': [0]}
{'predictions': [2]}
{'predictions': [1]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [0]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [2]}
{'predictions': [0]}
Monitor Prediction Requests and Responses using Kafka
from hops import kafka
from confluent_kafka import Producer, Consumer, KafkaError
Setup Kafka consumer and subscribe to the topic containing the prediction logs
TOPIC_NAME = serving.get_kafka_topic(SERVING_NAME)
config = kafka.get_kafka_default_config()
config['default.topic.config'] = {'auto.offset.reset': 'earliest'}
consumer = Consumer(config)
topics = [TOPIC_NAME]
consumer.subscribe(topics)
Read the Kafka Avro schema from Hopsworks and setup an Avro reader
json_schema = kafka.get_schema(TOPIC_NAME)
avro_schema = kafka.convert_json_schema_to_avro(json_schema)
Read messages from the Kafka topic, parse them with the Avro schema and print the results
PRINT_INSTANCES=False
PRINT_PREDICTIONS=True
for i in range(0, 10):
msg = consumer.poll(timeout=5.0)
if msg is not None:
value = msg.value()
try:
event_dict = kafka.parse_avro_msg(value, avro_schema)
payload = json.loads(event_dict["payload"])
if (event_dict['messageType'] == "request" and not PRINT_INSTANCES) or \
(event_dict['messageType'] == "response" and not PRINT_PREDICTIONS):
continue
print("INFO -> servingId: {}, modelName: {}, modelVersion: {},"\
"requestTimestamp: {}, inferenceId:{}, messageType:{}".format(
event_dict["servingId"],
event_dict["modelName"],
event_dict["modelVersion"],
event_dict["requestTimestamp"],
event_dict["inferenceId"],
event_dict["messageType"]))
if event_dict['messageType'] == "request":
print("Instances -> {}\n".format(payload['instances']))
if event_dict['messageType'] == "response":
print("Predictions -> {}\n".format(payload['predictions']))
except Exception as e:
print("A message was read but there was an error parsing it")
print(e)
else:
print("timeout.. no more messages to read from topic")
INFO -> servingId: 2523, modelName: irisflowerclassifier, modelVersion: 1,requestTimestamp: 1641839298, inferenceId:9fff76ad-55e1-4c52-b0e9-7019ce79a249, messageType:response
Predictions -> [2]
INFO -> servingId: 2523, modelName: irisflowerclassifier, modelVersion: 1,requestTimestamp: 1641839298, inferenceId:03c8d155-cbb9-4907-bfc2-630d3777e56f, messageType:response
Predictions -> [1]
INFO -> servingId: 2523, modelName: irisflowerclassifier, modelVersion: 1,requestTimestamp: 1641839298, inferenceId:cd45feb2-d1c1-43b6-a612-f0dffeaf0e01, messageType:response
Predictions -> [2]
INFO -> servingId: 2523, modelName: irisflowerclassifier, modelVersion: 1,requestTimestamp: 1641839298, inferenceId:4b14a25a-fed1-45cc-a994-a1b37ba55543, messageType:response
Predictions -> [2]
INFO -> servingId: 2523, modelName: irisflowerclassifier, modelVersion: 1,requestTimestamp: 1641839299, inferenceId:dbe84132-b0ce-4196-bd5c-da72f1607f45, messageType:response
Predictions -> [0]