ETA: Extensible Toolkit for Analytics

An open and extensible computer vision, machine learning and video analytics infrastructure.

Requirements

ETA is very portable:

• Installable on Mac or Linux
• Supports Python 2.7 and Python 3.6 or later
• Supports TensorFlow 1.X and 2.X
• Supports OpenCV 2.4+ and OpenCV 3.0+
• Supports CPU-only and GPU-enabled installations
• Supports CUDA 8, 9 and 10 for GPU installations

Installation

You can install the latest release of ETA via pip:

pip install voxel51-eta

This will perform a lite installation of ETA. If you use an ETA feature that requires additional dependencies (e.g., ffmpeg or tensorflow), you will be prompted to install the relevant packages.

Docker Installation

If you prefer to operate via Docker, see the Docker Build Guide for simple instructions for building a Docker image with an ETA environment installed.

Installation from source

Step 0: Setup your Python environment

It is assumed that you already have Python installed on your machine.

IMPORTANT: ETA assumes that the version of Python that you intend to use is accessible via python and pip on your path. In particular, for Python 3 users, this means that you may need to alias python3 and pip3 to python and pip, respectively.

We strongly recommend that you install ETA in a virtual environment to maintain a clean workspace.

Step 1: Clone the repository

git clone https://github.com/voxel51/eta
cd eta

Step 2: Run the install script

bash install.bash

Note that the install script supports flags that control things like (on macOS) whether port or brew is used to install packages. Run bash install.bash -h for more information.

For Linux installs, the script inspects your system to see if CUDA is installed via the lspci command. If CUDA is available, TensorFlow is installed with GPU support.

The table below lists the version of TensorFlow that will be installed by the installer, as recommended by the tested build configurations:

Note that ETA also supports TensorFlow 2.X. The only problems you may face when using ETA with TensorFlow 2 are when trying to run inference with ETA models that only support TensorFlow 1. A notable case here are TF-slim models. In such cases, you should see an informative error message alerting you of the requirement mismatch.

Lite installation

Some ETA users are only interested in using the core ETA library defined in the eta.core package. In such cases, you can perform a lite installation using the -l flag of the install script:

bash install.bash -l

Lite installation omits submodules and other large dependencies that are not required in order for the core library to function. If you use an ETA feature that requires additional dependencies (e.g., ffmpeg or tensorflow), you will be prompted to install the relevant packages.

Developer installation

If you are interested in contributing to ETA or generating its documentation from source, you should perform a developer installation using the -d flag of the install script:

bash install.bash -d

Setting up your execution environment

When the root eta package is imported, it tries to read the eta/config.json file to configure various package-level constants. Many advanced ETA features such as pipeline building, model management, etc. require a properly configured environment to function.

To setup your environment, create a copy the example configuration file:

cp config-example.json eta/config.json

If desired, you can edit your config file to customize the various paths, change default constants, add environment variables, customize your default PYTHONPATH, and so on. You can also add additional paths to the module_dirs, pipeline_dirs, and models_dirs sections to expose custom modules, pipelines, and models to your system.

Note that, when the config file is loaded, any {{eta}} patterns in directory paths are replaced with the absolute path to the eta/ directory on your machine.

The default config includes the modules/, pipelines/, and models/ directories on your module, pipeline, and models search paths, respectively. These directories contain the necessary information to run the standard analytics exposed by the ETA library. In addition, the relative paths ./modules/, ./pipelines/, and ./models/ are added to their respective paths to support the typical directory structure that we adopt for our custom projects.

CLI

Installing ETA automatically installs eta, a command-line interface (CLI) for interacting with the ETA Library. This utility provides access to many useful features of ETA, including building and running pipelines, downloading models, and interacting with remote storage.

To explore the CLI, type eta --help, and see the CLI Guide for complete information.

Quickstart

Get your feet wet with ETA by running some of examples in the examples folder.

Also, see the docs folder for more documentation about the various components of the ETA library.

Organization

The ETA package is organized as described below. For more information about the design and function of the various ETA components, read the documentation in the docs folder.

Generating Documentation

This project uses Sphinx-Napoleon to generate its documentation from source.

To generate the documentation, you must install the developer dependencies by running the install.bash script with the -d flag.

Then you can generate the docs by running:

bash sphinx/generate_docs.bash

To view the documentation, open the sphinx/build/html/index.html file in your browser.

Uninstallation

pip uninstall voxel51-eta

Acknowledgements

This project was gratefully supported by the NIST Public Safety Innovation Accelerator Program.

Citation

If you use ETA in your research, feel free to cite the project (but only if you love it 😊):

@article{moore2017eta,
  title={ETA: Extensible Toolkit for Analytics},
  author={Moore, B. E. and Corso, J. J.},
  journal={GitHub. Note: https://github.com/voxel51/eta},
  year={2017}
}