dffml 0.1.1

Is DFFML Right For Me?

If you answer yes to any of these questions DFFML can make your life easier.

• Dataset Generation

  • Need to generate a dataset

  • Need to run asynchronous operations in order to gather dataset (http requests, interaction with command line utilities, etc.)

• Models

  • Want to quickly prototype how machine learning could be used on a dataset without writing a model

  • Need to write a finely tuned model by interacting with low level APIs of popular machine learning frameworks.

• Storage

  • Need a way to use datasets which could be stored in different locations or formats.

About

DFFML facilitates data generation, model creation, and use of models via services. See Architecture to learn how it works.

• Facilitates data collection, model creation, and use of models via services.

• Provides plumbing to facilitate the collection of feature data to create datasets.

• Allows developers to define their ML models via a standardized API.

  • This let’s users try different libraries / models to compare performance.

• Plugin based

  • Features which gather feature data (Number of Git Authors, etc.)

  • Models which expose ML models via the standard API (Tensorflow, Scikit, etc.)

  • Sources which load and store feature data (CSV, JSON, MySQL, etc.)

The plumbing DFFML provides enables users to swap out models and features, in order to quickly prototype.

Installation

DFFML currently should work with Python 3.6. However, only Python 3.7 is officially supported. This is because there are a lot of nice helper methods Python 3.7 implemented that we intend to use instead of re-implementing.

python3.7 -m pip install -U dffml

You can also install the Features for Git Version Control, and Models for Tensorflow Library all at once.

• DFFML Features for Git Version Control

• DFFML Models for Tensorflow Library

If you want a quick how to on the iris dataset head to the DFFML Models for Tensorflow Library repo.

python3.7 -m pip install -U dffml[git,tensorflow]

Docker Build

This is a good option if you don’t want to deal with installing Python 3.7.

docker build -t dffml .

You can then alias dffml to run the docker container.

alias dffml="docker run --rm -ti -v $HOME/.local:/home/$USER/.local/ -v $PWD:/workdir -w /workdir -e UID=$(id -u) -e USER=$USER dffml"

This creates an alias that takes your current working directory and mounts it into /workdir as well as your $HOME/.local to the same in the container.

With the alias, you can run dffml commands as you would if installed via pip.

dffml list

Keep in mind that if you’re working on files they can only be ones in your current working directory, and if you want to access network resources and they are on your host, you’ll have to talk to 172.17.0.1 (docker0 inet address) instead of localhost or 127.0.0.1.

The purpose of mounting $HOME/.local is so that if you want to pip install anything, you can, and it will persist between invocations due to that being on the host.

If you wan to run pip you can put it after dffml.

dffml pip install example

Hacking

Then install in development mode to the virtualenv and development dependencies.

git clone git@github.com:intel/dffml
cd dffml
pip install --user -e .[git,tensorflow]

Usage

See DFFML Models for Tensorflow Library repo until documentation here is updated with a generic example.

Testing

python3.7 setup.py test

Architecture

When applying Machine Learning to a new problem developers must first collect data for models to train on. DFFML facilitates the collection of feature data to create datasets for models to learn on.

DFFML’s architecture can be thought of similarly to a search engine. Each Feature a developer defines searches for data associated with the unique key its provided with. Once the data is found it is added to a Repo (repository) associated with that unique key. A Feature’s search for data is dubbed evaluation. A Repo holds the results of each Feature’s evaluation. Results are stored under their respective Feature names.

To define machine learning a model within DFFML, users create a Model. Models are responsible for training, assessing accuracy, and making predictions. After evaluation a Repo can be used by a Model for any of those tasks. Defining a machine learning model as a Model allows users to quickly compare accuracy of various models on their gathered dataset.

Once the best most accurate model is known, users can easily integrate use of the model into existing applications via the Python API, or a Service. Services provide applications with ways to access the DFFML API over various protocols and deployment scenarios.

Repo

A repo is a repository of information. It is instantiated with a source URL which represents or points to where more information on it can be found.

Every repo has (or wants) a classification. Those which already have classifications can be used to train Models. The classification of the repo is what Education will ask it’s models to make predictions on.

Feature

Features are given a repo, containing at the minimum a source URL for it, and produce a list of results which represent the evaluation of that feature.

Not all methods are applicable to all repos. As such, all Features implement the applicable method.

Feature is the abstract base class for all features. New features must be derived from this class and implement the fetch, parse, and calc methods. These methods are always called in order by the evaluator. However, they are executed in parallel with the same stages of other features.

A feature is provided with a repo and is expected to fetch any data it needs to calculate itself when fetch is called. All data fetched should be stored in tempdir() if it must reside on disk.

Once the appropriate data is fetched the parse method is responsible for storing the parts of that data which will be used to calculate in the subclass

from dffml.feature import Feature

class StringByFT(Feature):

    async def fetch(self):
        self.__value = '42'

    async def parse(self):
        self.__value = int(self.__value)

The calc method then uses variables set in parse to calculate the feature.

async def calc(self):
    return self.__value * 42

entry_points={
    'dffml.feature': [
        'string_by_42 = mypackage.string_by_42:StringByFT',
    ],
},

Source

Repos come from a source. Sources may contain more information on a repo than just it’s source URL. Sources are responsible for providing the repos they contain and updating those repos upon request.

Model

Models are feed classified repos from which they learn from during their training phase. After training they can be used to make a prediction about the classification of a repo.

License

dffml is distributed under the MIT License, see LICENSE.

Legal

This software is subject to the U.S. Export Administration Regulations and other U.S. law, and may not be exported or re-exported to certain countries (Cuba, Iran, Crimea Region of Ukraine, North Korea, Sudan, and Syria) or to persons or entities prohibited from receiving U.S. exports (including Denied Parties, Specially Designated Nationals, and entities on the Bureau of Export Administration Entity List or involved with missile technology or nuclear, chemical or biological weapons).