bitfount 0.5.25

Machine Learning and Federated Learning Library.

Bitfount

This repository enables quick and easy experimentation with machine learning and federated learning models.

Table of Contents

• Bitfount
  • Table of Contents
  • Using the Docker images
  • Running the Python code
    • Installation
    • Environment variables
    • Getting started (Tutorials)
    • Federated training scripts
    • Basic Local Usage
  • License

Using the Docker images

There are two docker images, one for running a Pod (ghcr.io/bitfount/pod:stable), and another for running a modelling task (ghcr.io/bitfount/modeller:stable).

Both of the images require a config.yaml file to be provided to them, by default they will try to load it from /mount/config/config.yaml inside the docker container. You can provide this file easily by mounting/binding a volume to the container, how you do this may vary depending on your platform/environment (Docker/docker-compose/ECS), if you have any problems doing this then feel free to reach out to us.

Alternative you could copy a config file into a stopped container using docker cp.

If you're using a CSV data source then you'll also need to mount your data to the container, this will need to be mounted at the path specified in your config, for simplicity it's easiest put your config and your CSV in the same directory and then mount it to the container.

Once your container is running you will need to check the logs and complete the login step, allowing your container to authenticate with Bitfount. The process is the same as when running locally (e.g. the tutorials), except that we can't open the login page automatically for you.

Running the Python code

Installation

Where to get it

Binary installers for the latest released version are available at the Python Package Index (PyPI).

pip install bitfount

If you are planning on using the bitfount package with Jupyter Notebooks, we recommend you install the splinter package bitfount[tutorials] which will make sure you are running compatible jupyter dependencies.

pip install "bitfount[tutorials]"

Installation from sources

To install bitfount from source you need to create a python virtual environment.

In the bitfount directory (same one where you found this file after cloning the git repo), execute:

pip install -r requirements/requirements.in

These requirements are set to permissive ranges but are not guaranteed to work for all releases, especially the latest versions. For a pinned version of these requirements which are guaranteed to work, run the following command instead:

pip install -r requirements/requirements.txt

For MacOS you also need to install libomp:

brew install libomp

Environment variables

The following environment variables can optionally be set:

• BITFOUNT_ENGINE: determines the backend used. Current accepted values are "basic" or "pytorch". If pytorch is installed, this will automatically be selected
• BITFOUNT_ENVIRONMENT: accepted values are "production" or "staging". Defaults to "production". Should only be used for development purposes.
• BITFOUNT_HOME: path to a directory where a .bitfount subdirectory will be created for local storage purposes and configuration. Defaults to ~.
• BITFOUNT_LOG_TO_FILE: determines whether bitfount logs to file as well as console. Accepted values are "true" or "false". Defaults to "true"
• BITFOUNT_LOGS_DIR: determines where logfiles are stored. If empty, logs will be stored in a subdirectory called bitfount_logs in the directory where the script is run from
• BITFOUNT_POD_VITALS_PORT: determines the TCP port number to serve the pod vitals health check over. You can check the state of a running pod's health by accessing http://localhost:{{ BITFOUNT_POD_VITALS_PORT }}/health. A random open port will be selected if BITFOUNT_POD_VITALS_PORT is not set.

Getting started (Tutorials)

In order to run the tutorials, you also need to install the tutorial requirements:

pip install -r requirements/requirements-tutorial.txt

To get started using the Bitfount package in a federated setting, we recommend that you start with our tutorials. Run jupyter notebook and open up the first tutorial at: tutorials/01_running_a_pod.ipynb

Federated training scripts

Some simple scripts have been provided to run a Pod or Modelling job from a config file.

⚠️ If you are running from a source install (such as from git clone) you will need to use python -m scripts.<script_name> rather than use bitfount <script_name> directly.

To run a pod:

bitfount run_pod --path_to_config_yaml=<CONFIG_FILE>

To run a modelling job:

bitfount run_modeller --path_to_config_yaml=<CONFIG_FILE>

Basic Local Usage

As well as providing the ability to use data in remote pods, this package also enables local ML training. Some example code for this purpose is given below.

1. Import bitfount

import bitfount as bf

2. Create DataSource and load data

census_income = bf.CSVSource(
    path="https://bitfount-hosted-downloads.s3.eu-west-2.amazonaws.com/adult.csv",
    ignore_cols=["fnlwgt"],
)
census_income.load_data()

3. Create Schema

schema = bf.BitfountSchema(
    census_income,
    table_name="census_income",
    force_stypes={
        "census_income": {
            "categorical":[
                "TARGET",
                "workclass",
                "marital-status",
                "occupation",
                "relationship",
                "race",
                "native-country",
                "gender",
                "education"
            ]
        }
    }
)

4. Transform Data

clean_data = bf.CleanDataTransformation()
processor = bf.TransformationProcessor([clean_data], schema.get_table_schema("census_income"))
census_income.data = processor.transform(census_income.data)
schema.add_datasource_tables(census_income, table_name="census_income")

5. Create DataStructure

adult_data_structure=bf.DataStructure(
  table="census_income",
  target="TARGET",
)

6. Create and Train Model

nn = bf.PyTorchTabularClassifier(
    datastructure=adult_data_structure,
    schema=schema,
    epochs=2,
    batch_size=256,
    optimizer=bf.Optimizer("RAdam", {"lr": 0.001}),
)
nn.fit(census_income)
nn.serialize("demo_task_model.pt")

7. Evaluate

preds, targs = nn.evaluate()
metrics = bf.MetricCollection.create_from_model(nn)
results = metrics.compute(targs, preds)
print(results)

8. Assert results

import numpy as np
assert nn._validation_results[-1]["validation_loss"] is not np.nan
assert results["AUC"] > 0.7

License

The license for this software is available in the LICENSE file. This can be found in the Github Repository, as well as inside the Docker image.