Data-Analysis Metrics Library (DAML)

About DAML

The Data-Analysis Metrics Library, or DAML, focuses on characterizing image data and its impact on model performance across classification and object-detection tasks.

Model-agnostic metrics that bound real-world performance

• relevance/completeness/coverage
• metafeatures (data complexity)

Model-specific metrics that guide model selection and training

• dataset sufficiency
• data/model complexity mismatch

Metrics for post-deployment monitoring of data with bounds on model performance to guide retraining

• dataset-shift metrics
• model performance bounds under covariate shift
• guidance on sampling to assess model error and model retraining

Getting Started

Requirements

• Python 3.9-3.11

Installing DAML

You can install DAML directly from pypi.org using the following command. The optional dependencies of DAML are torch, tensorflow and all. Using torch enables Sufficiency metrics, and tensorflow enables OOD Detection.

pip install daml[all]

Installing DAML from GitHub

To install DAML from source locally on Ubuntu, you will need git-lfs to download larger, binary source files and poetry for project dependency management.

sudo apt-get install git-lfs
pip install poetry

Pull the source down and change to the DAML project directory.

git clone https://github.com/aria-ml/daml.git
cd daml

Install DAML with optional dependencies for development.

poetry install --all-extras --with dev

Now that DAML is installed, you can run commands in the poetry virtual environment by prefixing shell commands with poetry run, or activate the virtual environment directly in the shell.

poetry shell

Documentation and Tutorials

For more ideas on getting started using DAML in your workflow, additional information and tutorials are in our Sphinx documentation hosted on Read the Docs.

Attribution

This project uses code from the Alibi-Detect python library developed by SeldonIO. Additional documentation from the developers are also available here.

POCs

• POC: Scott Swan @scott.swan
• DPOC: Andrew Weng @aweng