Skip to content

Latest commit

 

History

History
242 lines (191 loc) · 8.34 KB

README.md

File metadata and controls

242 lines (191 loc) · 8.34 KB

mlflow-examples - e2e-ml-pipeline

End-to-end ML pipeline: training, model registry, deployment and prediction with MLflow scoring server.

Basis of subsequent CI/CD pipeline.

Overview

  • train.py - Run several sklearn training runs with different hyperparameters.
  • register_model.py - Find the best run and register it as model version models:/e2e-ml-pipeline/production.
  • batch_score.py - Batch score with sklearn and pyfunc flavors.
  • deploy_server.py - Launch a scoring server and submit scoring payload to http://localhost:5001/invocations endpoint. The scoring server can be either local web server or local SageMaker container.
  • test.py - Test above steps - train, register, batch score and deploy.

Architecture Diagrams

Batch Scoring

Real-time Scoring

Setup

See conda.yaml.

conda env create --file conda.yaml
conda activate mlflow-examples-e2e-ml-pipeline

Run

Key default values - see common.py.

Name Value
Experiment name e2e-ml-pipeline
Model name e2e-ml-pipeline
Model URI models:/e2e-ml-pipeline/production
Port 5001
Docker image name sm-e2e-ml-pipeline

Train

Program: train.py.

Run several sklearn training runs with different hyperparameters.

Options

Name Required Default Description
experiment_name no e2e-ml-pipeline Name of experiment
data_path no ../../data/wine-quality-white.csv Data file

Run

python train.py 

Arguments:
  experiment_name: e2e-ml-pipeline
  data_path: ../../data/wine-quality-white.csv
Experiment ID: 5
Params: (1, 2, 4, 16)
0.820  1 ae9b699cfc5e42f9bca2bd83f95f9524 5
0.785  2 5ae2374d1b524a4e8051c4fa1d7dbcc8 5
0.759  4 4c3da779aa8d48ff97f0c0ed9cafe47c 5
0.867 16 319b94e45b63483cbc8225d0d3a8d6bc 5
Best run: 0.759 4c3da779aa8d48ff97f0c0ed9cafe47c

Register Model

Program: register_model.py.

Find the best run and register it as model version models:/e2e-ml-pipeline/production.

Options

Name Required Default Description
experiment_name no e2e-ml-pipeline Name of experiment
data_path no ../../data/wine-quality-white.csv Data file
model_name no e2e-ml-pipeline Registered model

Run

python register_model.py 

Arguments:
  experiment_name: e2e-ml-pipeline
  data_path: ../../data/wine-quality-white.csv
  model_name: e2e-ml-pipeline
Best run: 4c3da779aa8d48ff97f0c0ed9cafe47c 0.7592585886611769
Found model e2e-ml-pipeline
Found 0 versions for model e2e-ml-pipeline
Reg Model: <class 'mlflow.entities.model_registry.registered_model.RegisteredModel'> {'_name': 'e2e-ml-pipeline', '_creation_time': 1584292037095, '_last_updated_timestamp': 1584302072077, '_description': '', '_latest_version': [<ModelVersion: creation_timestamp=1584302072077, current_stage='None', description='', last_updated_timestamp=1584302072077, name='e2e-ml-pipeline', run_id='4c3da779aa8d48ff97f0c0ed9cafe47c', source='file:///Users/ander/work/mlflow/server/local_mlrun/mlruns/5/4c3da779aa8d48ff97f0c0ed9cafe47c/artifacts/sklearn-model', status='READY', status_message='', user_id='', version='35'>]}
Version: id=36 status=READY state=None
Waited 0.01 seconds
Version: id=36 status=READY state=None
Version: id=36 status=READY state=Production
predictions: [6.24342105 6.24342105 6.68112798 ... 6.68112798 5.94352941 5.35624284]

Batch Score

Program: batch_score.py.

Score the best model with sklearn and pyfunc flavors.

Options

Name Required Default Description
model_uri no models:/e2e-ml-pipeline/production Registered model URI
data_path no ../../data/wine-quality-white.csv Data file

Run

python batch_score.py

sklearn score
predictions: [6.24931129 6.74246575 6.74246575 ... 6.24931129 6.74246575 6.24931129]

pyfunc score
predictions: [6.24931129 6.74246575 6.74246575 ... 6.24931129 6.74246575 6.24931129]

Deploy Scoring Server

Program: deploy_server.py.

Steps:

  • Fetches the best model from the model registry.
  • Launches a scoring server with this model
  • Submits scoring payload to http://localhost:5001/invocations endpoint.

Options

Name Required Default Description
port no 5001 Port
docker_image no none Docker image name
model_uri no models:/e2e-ml-pipeline/production Registered model URI
launch_container no false Default is to launch local server, if true launch local SageMaker container
data_path no ../../data/wine-quality-white.csv Data file

Run local webserver

Executes the following mlflow command:

  • mlflow models serve --port 5001 --model-uri models:/e2e-ml-pipeline/production
python -u deploy_server.py 

Arguments:
  port: 5001
  docker_image: sm-e2e-ml-pipeline
  model_uri: models:/e2e-ml-pipeline/production
  launch_container: False
  data_path: ../../data/wine-quality-white.csv

Command: mlflow models serve --port 5001 --model-uri models:/e2e-ml-pipeline/production
Process ID: 27572
Calling scoring server: 4/10000
. . . 
Calling scoring server: 8/10000
Ran pip subprocess with arguments:
['/Users/andre/miniconda/envs/mlflow-9373994feef365bde89b7072d1498bdbc226ba90/bin/python', '-m', 'pip', 'install', '-U', '-r', '/Users/ander/work/mlflow/server/local_mlrun/mlruns/2/868ce3ccb7c448fe9939c5b21c29ff0a/artifacts/sklearn-model/condaenv.kdzq1ifu.requirements.txt']
Processing /Users/ander/Library/Caches/pip/wheels/6d/72/87/348958818bec20c3a64243396065e34600ada290199f96abfa/mlflow-1.7.0-py3-none-any.whl
Collecting cloudpickle==1.2.2
  Using cached cloudpickle-1.2.2-py2.py3-none-any.whl (25 kB)
. . . 
2020/03/15 22:44:36 INFO mlflow.models.cli: Selected backend for flavor 'python_function'
Done waiting for 4.36 seconds
Done waiting - OK - successful deploy - predictions: [5.335031847133758, 5.050955414012739, 5.726950354609929]
Done waiting - killing process 27572

Run local SagemMaker container

Executes the following mlflow commands:

  • mlflow sagemaker build-and-push-container --build --no-push --container sm-e2e-ml-pipeline
  • mlflow sagemaker run-local -m models:/e2e-ml-pipeline/production -p 5001 --image sm-e2e-ml-pipeline
python -u deploy_server.py --launch_container

Arguments:
  port: 5001
  docker_image: sm-e2e-ml-pipeline
  model_uri: models:/e2e-ml-pipeline/production
  launch_container: False
  data_path: ../../data/wine-quality-white.csv
Starting command: mlflow sagemaker build-and-push-container --build --no-push --container sm-e2e-ml-pipeline
. . .
Successfully built a03004c4fa63
Successfully tagged sm-e2e-ml-pipeline:latest
Done waiting for command: mlflow sagemaker build-and-push-container --build --no-push --container sm-e2e-ml-pipeline
Starting command: mlflow sagemaker run-local -m models:/e2e-ml-pipeline/production -p 5001 --image sm-e2e-ml-pipeline
2020/03/15 22:48:34 INFO mlflow.models.docker_utils: Building docker image with name sm-e2e-ml-pipeline
/var/folders/_9/tbkxzw0116v2cp_zq4f1_1cm0000gp/T/tmp0q3cm6g9/
/var/folders/_9/tbkxzw0116v2cp_zq4f1_1cm0000gp/T/tmp0q3cm6g9//Dockerfile
Sending build context to Docker daemon  3.072kB
Process ID: 27632
Calling scoring server: 0/10000
. . .
Calling scoring server: 23/10000
. . .
Done waiting for 65.36 seconds
Done waiting - OK - Successful deploy - predictions: [5.335031847133758, 5.050955414012739, 5.726950354609929]
Done waiting - killing process 27632

Tests

Program: test.py.

Test above steps in sequential order - train, register, batch score and deploy.

By default launches a local server. If you wish to launch the docker container: export LAUNCH_CONTAINER=true.

py.test -v test.py

test.py::test_train PASSED 
test.py::test_register_model PASSED 
test.py::test_batch_score PASSED 
test.py::test_deploy_server PASSED 
============================== 4 passed in 6.17s ===============================