BigML Python Bindings

BigML makes machine learning easy by taking care of the details required to add data-driven decisions and predictive power to your company. Unlike other machine learning services, BigML creates beautiful predictive models that can be easily understood and interacted with.

These BigML Python bindings allow you to interact with BigML.io, the API for BigML. You can use it to easily create, retrieve, list, update, and delete BigML resources (i.e., sources, datasets, models and, predictions). For additional information, see the full documentation for the Python bindings on Read the Docs.

This module is licensed under the Apache License, Version 2.0.

Support

Please report problems and bugs to our BigML.io issue tracker.

Discussions about the different bindings take place in the general BigML mailing list. Or join us in our Campfire chatroom.

Requirements

Only Python 3 versions are currently supported by these bindings. Support for Python 2.7.X ended in version 4.32.3.

The basic third-party dependencies are the requests, unidecode, requests-toolbelt, bigml-chronos, msgpack, numpy and scipy libraries. These libraries are automatically installed during the basic setup. Support for Google App Engine has been added as of version 3.0.0, using the urlfetch package instead of requests.

The bindings will also use simplejson if you happen to have it installed, but that is optional: we fall back to Python's built-in JSON libraries is simplejson is not found.

The bindings provide support to use the BigML platform to create, update, get and delete resources, but also to produce local predictions using the models created in BigML. Most of them will be actionable with the basic installation, but some additional dependencies are needed to use local Topic Models and Image Processing models. Please, refer to the Installation section for details.

OS Requirements

The basic installation of the bindings is compatible and can be used on Linux and Windows based Operating Systems. However, the extra options that allow working with image processing models ([images] and [full]) are only supported and tested on Linux-based Operating Systems. For image models, Windows OS is not recommended and cannot be supported out of the box, because the specific compiler versions or dlls required are unavailable in general.

Installation

To install the basic latest stable release with pip, please use:

$ pip install bigml

Support for local Topic Distributions (Topic Models' predictions) and local predictions for datasets that include Images will only be available as extras, because the libraries used for that are not usually available in all Operative Systems. If you need to support those, please check the Installation Extras section.

Installation Extras

Local Topic Distributions support can be installed using:

pip install bigml[topics]

Images local predictions support can be installed using:

pip install bigml[images]

The full set of features can be installed using:

pip install bigml[full]

WARNING: Mind that installing these extras can require some extra work, as explained in the Requirements section.

You can also install the development version of the bindings directly from the Git repository

$ pip install -e git://github.com/bigmlcom/python.git#egg=bigml_python

Running the Tests

The tests will be run using pytest. You'll need to set up your authentication via environment variables, as explained in the authentication section. Also some of the tests need other environment variables like BIGML_ORGANIZATION to test calls when used by Organization members and BIGML_EXTERNAL_CONN_HOST, BIGML_EXTERNAL_CONN_PORT, BIGML_EXTERNAL_CONN_DB, BIGML_EXTERNAL_CONN_USER, BIGML_EXTERNAL_CONN_PWD and BIGML_EXTERNAL_CONN_SOURCE in order to test external data connectors.

With that in place, you can run the test suite simply by issuing

$ pytest

Additionally, Tox can be used to automatically run the test suite in virtual environments for all supported Python versions. To install Tox:

$ pip install tox

Then run the tests from the top-level project directory:

$ tox

Importing the module

To import the module:

import bigml.api

Alternatively you can just import the BigML class:

from bigml.api import BigML

Authentication

All the requests to BigML.io must be authenticated using your username and API key and are always transmitted over HTTPS.

This module will look for your username and API key in the environment variables BIGML_USERNAME and BIGML_API_KEY respectively.

Unix and MacOS

You can add the following lines to your .bashrc or .bash_profile to set those variables automatically when you log in:

export BIGML_USERNAME=myusername
export BIGML_API_KEY=ae579e7e53fb9abd646a6ff8aa99d4afe83ac291

refer to the next chapters to know how to do that in other operating systems.

With that environment set up, connecting to BigML is a breeze:

from bigml.api import BigML
api = BigML()

Otherwise, you can initialize directly when instantiating the BigML class as follows:

api = BigML('myusername', 'ae579e7e53fb9abd646a6ff8aa99d4afe83ac291')

These credentials will allow you to manage any resource in your user environment.

In BigML a user can also work for an organization. In this case, the organization administrator should previously assign permissions for the user to access one or several particular projects in the organization. Once permissions are granted, the user can work with resources in a project according to his permission level by creating a special constructor for each project. The connection constructor in this case should include the project ID:

api = BigML('myusername', 'ae579e7e53fb9abd646a6ff8aa99d4afe83ac291',
            project='project/53739b98d994972da7001d4a')

If the project used in a connection object does not belong to an existing organization but is one of the projects under the user's account, all the resources created or updated with that connection will also be assigned to the specified project.

When the resource to be managed is a project itself, the connection needs to include the corresponding``organization ID``:

api = BigML('myusername', 'ae579e7e53fb9abd646a6ff8aa99d4afe83ac291',
            organization='organization/53739b98d994972da7025d4a')

Authentication on Windows

The credentials should be permanently stored in your system using

setx BIGML_USERNAME myusername
setx BIGML_API_KEY ae579e7e53fb9abd646a6ff8aa99d4afe83ac291

Note that setx will not change the environment variables of your actual console, so you will need to open a new one to start using them.

Authentication on Jupyter Notebook

You can set the environment variables using the %env command in your cells:

%env BIGML_USERNAME=myusername
%env BIGML_API_KEY=ae579e7e53fb9abd646a6ff8aa99d4afe83ac291

Alternative domains

The main public domain for the API service is bigml.io, but there are some alternative domains, either for Virtual Private Cloud setups or the australian subdomain (au.bigml.io). You can change the remote server domain to the VPC particular one by either setting the BIGML_DOMAIN environment variable to your VPC subdomain:

export BIGML_DOMAIN=my_VPC.bigml.io

or setting it when instantiating your connection:

api = BigML(domain="my_VPC.bigml.io")

The corresponding SSL REST calls will be directed to your private domain henceforth.

You can also set up your connection to use a particular PredictServer only for predictions. In order to do so, you'll need to specify a Domain object, where you can set up the general domain name as well as the particular prediction domain name.

from bigml.domain import Domain
from bigml.api import BigML

domain_info = Domain(prediction_domain="my_prediction_server.bigml.com",
                     prediction_protocol="http")

api = BigML(domain=domain_info)

Finally, you can combine all the options and change both the general domain server, and the prediction domain server.

from bigml.domain import Domain
from bigml.api import BigML
domain_info = Domain(domain="my_VPC.bigml.io",
                     prediction_domain="my_prediction_server.bigml.com",
                     prediction_protocol="https")

api = BigML(domain=domain_info)

Some arguments for the Domain constructor are more unsual, but they can also be used to set your special service endpoints:

• protocol (string) Protocol for the service (when different from HTTPS)
• verify (boolean) Sets on/off the SSL verification
• prediction_verify (boolean) Sets on/off the SSL verification for the prediction server (when different from the general SSL verification)

Note that the previously existing dev_mode flag:

api = BigML(dev_mode=True)

that caused the connection to work with the Sandbox Development Environment has been deprecated because this environment does not longer exist. The existing resources that were previously created in this environment have been moved to a special project in the now unique Production Environment, so this flag is no longer needed to work with them.

Quick Start

Imagine that you want to use this csv file containing the Iris flower dataset to predict the species of a flower whose petal length is 2.45 and whose petal width is 1.75. A preview of the dataset is shown below. It has 4 numeric fields: sepal length, sepal width, petal length, petal width and a categorical field: species. By default, BigML considers the last field in the dataset as the objective field (i.e., the field that you want to generate predictions for).

sepal length,sepal width,petal length,petal width,species
5.1,3.5,1.4,0.2,Iris-setosa
4.9,3.0,1.4,0.2,Iris-setosa
4.7,3.2,1.3,0.2,Iris-setosa
...
5.8,2.7,3.9,1.2,Iris-versicolor
6.0,2.7,5.1,1.6,Iris-versicolor
5.4,3.0,4.5,1.5,Iris-versicolor
...
6.8,3.0,5.5,2.1,Iris-virginica
5.7,2.5,5.0,2.0,Iris-virginica
5.8,2.8,5.1,2.4,Iris-virginica

You can easily generate a prediction following these steps:

from bigml.api import BigML

api = BigML()

source = api.create_source('./data/iris.csv')
dataset = api.create_dataset(source)
model = api.create_model(dataset)
prediction = api.create_prediction(model, \
    {"petal width": 1.75, "petal length": 2.45})

You can then print the prediction using the pprint method:

>>> api.pprint(prediction)
species for {"petal width": 1.75, "petal length": 2.45} is Iris-setosa

Certainly, any of the resources created in BigML can be configured using several arguments described in the API documentation. Any of these configuration arguments can be added to the create method as a dictionary in the last optional argument of the calls:

from bigml.api import BigML

api = BigML()

source_args = {"name": "my source",
     "source_parser": {"missing_tokens": ["NULL"]}}
source = api.create_source('./data/iris.csv', source_args)
dataset_args = {"name": "my dataset"}
dataset = api.create_dataset(source, dataset_args)
model_args = {"objective_field": "species"}
model = api.create_model(dataset, model_args)
prediction_args = {"name": "my prediction"}
prediction = api.create_prediction(model, \
    {"petal width": 1.75, "petal length": 2.45},
    prediction_args)

The iris dataset has a small number of instances, and usually will be instantly created, so the api.create_ calls will probably return the finished resources outright. As BigML's API is asynchronous, in general you will need to ensure that objects are finished before using them by using api.ok.

from bigml.api import BigML

api = BigML()

source = api.create_source('./data/iris.csv')
api.ok(source)
dataset = api.create_dataset(source)
api.ok(dataset)
model = api.create_model(dataset)
api.ok(model)
prediction = api.create_prediction(model, \
    {"petal width": 1.75, "petal length": 2.45})

Note that the prediction call is not followed by the api.ok method. Predictions are so quick to be generated that, unlike the rest of resouces, will be generated synchronously as a finished object.

The example assumes that your objective field (the one you want to predict) is the last field in the dataset. If that's not he case, you can explicitly set the name of this field in the creation call using the objective_field argument:

from bigml.api import BigML

api = BigML()

source = api.create_source('./data/iris.csv')
api.ok(source)
dataset = api.create_dataset(source)
api.ok(dataset)
model = api.create_model(dataset, {"objective_field": "species"})
api.ok(model)
prediction = api.create_prediction(model, \
    {'sepal length': 5, 'sepal width': 2.5})

You can also generate an evaluation for the model by using:

test_source = api.create_source('./data/test_iris.csv')
api.ok(test_source)
test_dataset = api.create_dataset(test_source)
api.ok(test_dataset)
evaluation = api.create_evaluation(model, test_dataset)
api.ok(evaluation)

If you set the storage argument in the api instantiation:

api = BigML(storage='./storage')

all the generated, updated or retrieved resources will be automatically saved to the chosen directory.

Alternatively, you can use the export method to explicitly download the JSON information that describes any of your resources in BigML to a particular file:

api.export('model/5acea49a08b07e14b9001068',
           filename="my_dir/my_model.json")

This example downloads the JSON for the model and stores it in the my_dir/my_model.json file.

In the case of models that can be represented in a PMML syntax, the export method can be used to produce the corresponding PMML file.

api.export('model/5acea49a08b07e14b9001068',
           filename="my_dir/my_model.pmml",
           pmml=True)

You can also retrieve the last resource with some previously given tag:

api.export_last("foo",
                resource_type="ensemble",
                filename="my_dir/my_ensemble.json")

which selects the last ensemble that has a foo tag. This mechanism can be specially useful when retrieving retrained models that have been created with a shared unique keyword as tag.

For a descriptive overview of the steps that you will usually need to follow to model your data and obtain predictions, please see the basic Workflow sketch document. You can also check other simple examples in the following documents:

• model 101
• logistic regression 101
• linear regression 101
• ensemble 101
• cluster 101
• anomaly detector 101
• association 101
• topic model 101
• deepnet 101
• time series 101
• fusion 101
• scripting 101

Additional Information

We've just barely scratched the surface. For additional information, see the full documentation for the Python bindings on Read the Docs. Alternatively, the same documentation can be built from a local checkout of the source by installing Sphinx ($ pip install sphinx) and then running

$ cd docs
$ make html

Then launch docs/_build/html/index.html in your browser.

How to Contribute

Please follow the next steps:

1. Fork the project on github.com.
2. Create a new branch.
3. Commit changes to the new branch.
4. Send a pull request.

For details on the underlying API, see the BigML API documentation.