autonlab / autonbox

Primitives for the D3M program

The Auton Lab TA1 primitives

This repository contains additional Auton Lab TA1 primitives for the D3M program.

1. Iterative Labeling - Blackbox based iterative labeling for semi-supervised learning
2. Video featurizer - Video Feature Extraction for Action Classification With 3D ResNet

Installation

To install primitives, run:

pip install -U -e git+https://github.com/autonlab/autonbox.git#egg=autonbox

Video featurizer requires a static file, pre-trained model weights. To download it, run:

mkdir -p /tmp/cmu/pretrained_files
python3 -m d3m index download -o /tmp/cmu/pretrained_files # requires d3m core

Video featurizer

The primitive outputs a data frame of size N x M, where N is the number of videos and M is 2024 features of type float.

It supports running on GPUs.

Merge Partial MultiPredictions

This primitive allows to merge partial predictions. These partial predictions may happen when removing rows of a dataset. It is however necessary to provide a fallback predictions to offer a prediction to each initial row. The strategy adopted in this primitive is to take the first vote for each row; therefore the order of the inputs predictions is crucial (for instance, one can use a cross correlation score to sort this input).

Clean Augmentation

This primitive removes rows of a dataset if they contain less than x% of features.