There is a complete ForML project available under examples/tutorial/titanic/. It is the famous Titanic Challenge. We will use it here to demonstrate the typical ForML usecases.

Before you start, please make sure to install ForML as per the installation instructions and ideally also familiarize yourself with the ForML concepts.


Datasource Preparation

ForML uses feeds to supply data into your projects. We need to register the Titanic dataset to our platform:

  1. Let’s start with fetching the Titanic dataset from Kaggle (assuming you have the kaggle API CLI tool installed and configured):

    $ kaggle competitions download -p /tmp -f train.csv titanic
  2. Convert the dataset to sqlite DB:

    import pandas as pd
    import sqlite3
    pd.read_csv('/tmp/train.csv').to_sql('passenger', sqlite3.connect('/tmp/tutorial.db'), index=False)

Platform Setup

Assuming you have no existing feeds configured in your system yet, let’s create one and register the dataset within:

Create a python file under ~/.forml/tutorial.py with the following content:

from forml.io import feed
from forml.lib.reader import sqlite
from openschema.kaggle import titanic

class Feed(feed.Provider):
    """Tutorial feed."""

    class Reader(sqlite.Reader):
        """Using a SQLite reader."""

    def sources(self):
        """This feed can serve just one and only dataset - the titanic passenger table mapped to
           the titanic.Passenger schema."""

        return {titanic.Passenger: 'passenger'}

Now let’s specify the actual ForML platform configuration. Add the following content to your ~/.forml/config.toml:

default = "compute"

provider = "dask"
scheduler = "multiprocessing"

provider = "graphviz"
format = "png"

default = "tutorial"

provider = "filesystem"
path = "/tmp/forml-tutorial"

default = ["tutorial"]

provider = "tutorial:Feed"
database = "/tmp/tutorial.db"

default = "print"

provider = "stdout"

Project Operations

We will exercise the standard lifecycle actions.

Development Lifecycle Actions

  1. Change directory to the root of the titanic project working copy.

  2. Let’s first run all the operator unit tests to confirm the project is in good shape:

    $ python3 setup.py test
    running test
    Test of Invalid Params ... ok
    Test of Not Trained ... ok
    Test of Valid Imputation ... ok
    Test of Invalid Params ... ok
    Test of Invalid Source ... ok
    Test of Valid Parsing ... ok
    Ran 6 tests in 0.591s
  3. Try running the train mode on the graphviz runner (called visual in our config ) to see the train task graph:

    $ python3 setup.py train --runner visual
  1. Run the eval mode on the (default) dask runner (called compute in our config) to get the cross-validation score:

    $ python3 setup.py eval
  2. Create the project package artifact and upload it to the (default) filesystem registry (assuming the same linage doesn’t already exist - otherwise increment the project version in the setup.py):

    $ python3 setup.py bdist_4ml upload

    This should publish the project into your local filesystem forml registry making it available for the production lifecycle. This becomes the first published lineage of this project versioned (according to the version from :ref:setup.py <project-setup> as 0.1.dev0)

Production Lifecycle Actions

Production lifecycles doesn’t need the project working copy so feel free to change the directory to another location before executing the commands.

  1. List the local registry confirming the project has been published its first lineage:

    $ forml list
    $ forml list forml-example-titanic
    $ forml list forml-example-titanic 0.1.dev0

    The output shows the project artifact is available in the registry as a lineage 0.1.dev0 not having any generation yet (the last command not producing any output).

  2. Train the project to create the first generation of its models and list the registry to confirm it got persisted:

    $ forml train forml-example-titanic
    $ forml list forml-example-titanic 0.1.dev0

    Now we have our first generation of the titanic models available in the registry.

  3. Apply the trained generation of the project to get the predictions:

    $ forml apply forml-example-titanic
    [[0.59180614 0.40819386]
    [0.60498469 0.39501531]
    [0.61020908 0.38979092]
    [0.64064548 0.35935452]]
  4. Run the apply mode alternatively on the graphviz runner to explore its task graph:

    $ forml -R visual apply forml-example-titanic

Working with Jupyter Notebooks

See the tutorial notebook stored in the demo project under examples/tutorial/titanic/notebooks/tutorial.ipynb for a step-by-step examples of working with ForML project in Jupyter.

Further details on the interactive style of work with ForML in general can be found in the Interactive Mode sections.