Here you find a Skeleton project for building a simple model in a python script or notebook and log the results on MLFlow.
There are two ways to do it:
In Jupyter Notebooks: We train a simple model in the jupyter notebook, where we select only some features and do minimal cleaning. The hyperparameters of feature engineering and modeling will be logged with MLflow
With Python scripts: The main script will go through exactly the same process as the jupyter notebook and also log the hyperparameters with MLflow
Data used is the coffee quality dataset.
Use the requirements file in this repo to create a new environment.
make setup
#or
pyenv local 3.11.3
python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements_dev.txt
The requirements.txt
file contains the libraries needed for deployment.. of model or dashboard .. thus no jupyter or other libs used during development.
The MLFLOW URI should not be stored on git, you have two options, to save it locally in the .mlflow_uri
file:
echo http://127.0.0.1:5000/ > .mlflow_uri
This will create a local file where the uri is stored which will not be added on github (.mlflow_uri
is in the .gitignore
file). Alternatively you can export it as an environment variable with
export MLFLOW_URI=http://127.0.0.1:5000/
This links to your local mlflow, if you want to use a different one, then change the set uri.
The code in the config.py will try to read it locally and if the file doesn't exist will look in the env var.. IF that is not set the URI will be empty in your code.
You can do it via the GUI or via command line if you use the local mlflow:
mlflow experiments create --experiment-name 0-template-ds-modeling
Check your local mlflow
mlflow ui
and open the link http://127.0.0.1:5000
This will throw an error if the experiment already exists. Save the experiment name in the config file.
In order to train the model and store test data in the data folder and the model in models run:
#activate env
source .venv/bin/activate
python -m modeling.train
In order to test that predict works on a test set you created run:
python modeling/predict.py models/linear data/X_test.csv data/y_test.csv
MLFlow is a tool for tracking ML experiments. You can run it locally or remotely. It stores all the information about experiments in a database. And you can see the overview via the GUI or access it via APIs. Sending data to mlflow is done via APIs. And with mlflow you can also store models on S3 where you version them and tag them as production for serving them in production.
You can group model trainings in experiments. The granularity of what an experiment is up to your usecase. Recommended is to have an experiment per data product, as for all the runs in an experiment you can compare the results.
In order to send data about your model you need to set the connection information, via the tracking uri and also the experiment name (otherwise the default one is used). One run represents a model, and all the rest is metadata. For example if you want to save train MSE, test MSE and validation MSE you need to name them as 3 different metrics. If you are doing CV you can set the tracking as nested.
There is no constraint between runs to have the same metadata tracked. I.e. for one run you can track different tags, different metrics, and different parameters (in cv some parameters might not exist for some runs so this .. makes sense to be flexible).