rohan2810 / mlflow

mlflow

Hyperparameter search

mlflow run . --env-manager=local -P file_name='hyperparameter_search.py'

From Github

mlflow run https://github.com/rohan2810/mlflow.git --env-manager=local -P file_name='hyperparameter_search.py'

model training

use alpha and l1_ratio from hyperparameter search

mlflow run . --env-manager=local -P file_name='train.py' -P alpha=0.05 -P l1_ratio=0.1

From Github

mlflow run https://github.com/rohan2810/mlflow.git --env-manager=local -P file_name='train.py' -P alpha=0.05 -P l1_ratio=0.1

model testing

mlflow run . --env-manager=local -P file_name='test.py' -P model_name='Model A' -P model_version=1

From Github

mlflow run https://github.com/rohan2810/mlflow.git --env-manager=local -P file_name='test.py' -P model_name='Model A' -P model_version=1

Notes:

MLFlow:

experiment tracking, project management, model management/registry

Installation:

mlflow: pip3 install mlflow

UI (localhost:5000): run mlflow ui on a seperate terminal (same project directory)

Experiment tracking:

After data pre-processing step/before running the model(training, testing, hyperparamter), add mlflow.set_tracking_uri("http://127.0.0.1:5000") to log everything to localhost.

Logging parameters: mlflow.log_param("var", var)

Logging metrics: mlflow.log_metric("var", var)

Logging artifacts: mlflow..log_model

localhost:5000 lists individual runs with params, metrics, and artifacts and models

Project management:

Can package MLflow projects for reusability

Conda.yml file to list dependencies

MlFlow file for parameters and command

Eg. To find best hyperparameters:

1. From local project: mlflow run . --env-manager=local -P file_name='hyperparameter_search.py'

   '--env-manager=local' forces to use local env instead of creating a new conda env

2. From github: mlflow run https://github.com/rohan2810/mlflow.git --env-manager=local -P file_name='hyperparameter_search.py'

Model Registry:

Can register models and reuse using APIs and CLI

Versions, Stage (None, staging, production, archived)

Models can be loaded using model = mlflow.sklearn.load_model(model_uri=f"models:/{model_name}/{model_version}")

Eg. https://github.com/rohan2810/mlflow/blob/main/test.py#L44