An example of how to run Facebook Prophet in Metaflow.

Read full blog post here!!

Minimum daily temperatures dataset is used. This dataset describes the minimum daily temperatures over 10 years (1981-1990) in the city Melbourne, Australia. The units are in degrees Celsius and there are 3650 observations. The source of the data is credited as the Australian Bureau of Meteorology.

ProphetFlow is a Metaflow flow using Facebook Prophet to predict future values of a timeseries. A Prophet model is trained to predict next year minimum daily temperatures. Hyperparameters tuning is performed with cross-validation.

Step start
    Init configuration
    => load

Step load
    Raw data is loaded and prepared
    => hyper_tuning

Step hyper_tuning
    Hyperparameters tuning
    => cross_validation

Step cross_validation
    Perform cross-validation on given hyperparameters
    => train

Step train
    Check cross-validation results and find best parameters.
    A new Prophet model is fitted.
    => end

Step end
    Last step, process is finished

If you need, install required Python packages (Metaflow & FBProphet) as usual: pip install -r requirements.txt

Simply run training flow with: python src/ProphetFlow.py run

In predict.ipynb notebook it's explained how to use Prophet trained model to predict future values, accessing Metaflow runs artifacts.

A simple flow without hyperparameters tuning: python src/ProphetSimpleFlow.py run

To configure Metaflow to use AWS Cloud, you first need to create stack:

aws s3 cp aws/metaflow-cfn-template.yml s3://<your_bucket>
aws cloudformation create-stack --stack-name metaflow \
                                --template-url <url_of_template_in_your_bucket> \
                                --capabilities CAPABILITY_IAM

When Cloudformation stack is created, you must configure Metaflow.

metaflow configure aws

To run all steps in flow in AWS Batch:

python src/ProphetFlow.py run --with batch:image=vnardone/prophet-metaflow

An example of how to run a single step in AWS Batch (using @batch decorator):

python src/ProphetAWSStepFlow.py run