Sample MLOps pipeline

Train a simple genre classification model to recognize a song genre basing on it's characteristics.

Requirements

1. Conda: https://docs.conda.io/projects/conda/en/latest/user-guide/install/
2. Weights&Biases: https://docs.wandb.ai/quickstart
   1. Sign-up: https://wandb.ai/login?signup=true
   2. Install: pip install wandb
   3. Login from CLI: wandb login
3. Mlflow: pip install mlflow
   1. in case CLI command mlflow run . can't find it, install from conda forge as part of conda.yml
4. I suggest to:
   1. set-up a dedicated conda environment with e.g.: conda create --name mlops python=3.8 mlflow jupyter pandas matplotlib requests -c conda-forge
   2. Then, activate it with conda activate mlops

Sample runs

1. To perform a simple, complete run, in terminal (CLI) type mlflow run .

2. This sample pipeline consists of the following steps (see config.yaml, which stores all the Hydra options):

   - 1_download
   - 2_preprocess
   - 3_check_data
   - 4_segregate
   - 5_random_forest
   - 6_evaluate
   

   To run a specific pipeline phase, type e.g.

   mlflow run . -P hydra_options="main.execute_steps='1_download,2_preprocess'"

   That will run only first 2 of these steps.

3. To specify model (random_forest) hyper-parameters from CLI:

   1. single hyper-parameter: mlflow run . -P hydra_options="random_forest_pipeline.random_forest.max_depth=15"
   2. two values of a single hyper-parameter: mlflow run . -P hydra_options="-m random_forest_pipeline.random_forest.max_depth=13,15"
   3. two values of two different hyper-parameters: mlflow run . -P hydra_options="-m random_forest_pipeline.random_forest.max_depth=15 random_forest_pipeline.random_forest.min_samples_split: 3"
   4. a range of one single hyper-parameter: mlflow run . -P hydra_options="-m random_forest_pipeline.random_forest.max_depth=(10,20,2)" < generate a range from 10 to 20 in increments of 2

4. The same as above,, but with joblib launcher to make parallel computations: mlflow run . -P hydra_options="hydra/launcher=joblib -m random_forest_pipeline.random_forest.max_depth=13,15" To apply more intelligent grid search, see hydra sweepers: https://hydra.cc/docs/plugins/ax_sweeper/

MLflow run directly from GitHub

I've made a first release (1.0.0) of this pipeline on GitHub, so you can call it directly from there with:

mlflow run -v 1.0.0 https://github.com/wodecki/sample_mlops_pipeline

Serving the model with MLFlow

First, download Your model

Download the model to the "model" subdirectory, e.g. with: wandb artifact get sample_MLOps_pipeline/model_export:latest --root model

It downloads:

• conda.yaml
• input_example.json
• MLmodel
• model.pkl

Online processing

You can serve a model for online inference by using mlflow models serve:

mlflow models serve -m model &

Mlflow will create a REST API for us that we can interrogate by sending requests to the endpoint (which by default is http://localhost:5000/invocations). For example, we can do this from python like this:

predictions = requests.post("http://localhost:5000/invocations", json=data)

Where data is a file with samples to predict in a json format.

Batch processing

Just run:

mlflow models predict -t csv -i input.csv -m model

Where input.csv contains samples to predict.