An Unsupervised Learning approach to active learning.
Latens uses Python 3.6 or higher; see here for recent downloads, or install from brew. Additionally, it relies on Tensorflow 1.12.0 or higher, which can be found here.
After cloning this repository, add its root to the PYTHON_PATH by running
export PYTHONPATH=PATH_TO_LATENS:$PYTHONPATH
where PATH_TO_LATENS is replaced with the path to the root directory (where
this file is located). You can add the same line to your ~/.bash_profile or
equivalent config file to make the change permanent.
latens can be easily imported, once it has been added to the PYTHONPATH, but
to run directly, the latens.py script contains the main functionality. Run
python latens.py -h
to see the available options.
Getting started can be tricky because of the data format that latens expects
data to be in. It was originally developed with the MNIST dataset. Similar
datasets should also be compatible.
Before it can begin training, latens requires data to be stored in a
.tfrecord format. TFRecords are not very well standardized, so we provide the
convert command, which should format data as expected. Store images and
labels in a single .npz file with keywords "data" and "labels"
respectively. data/mnist.npz is provided for reference. Once this is done, run
python latens.py convert -i data/mnist.npz
to create a .tfrecord file in the same directory.