Authors: Harsh Bhatia (hbhatia@llnl.gov) and Joseph Y Moon
DynIm is a pure-python package to perform dynamic-importance (DynIm)
sampling on a high-dimensional data set.
DynIm is designed to minimize redundancy and maximize the coverage of the sampled points. DynIm uses the notion of "dissimilarity" from previously selected samples to define the importance of potential selections, and selects the ones that are most dissimilar. Simply, DynIm provides a farthest-point sampling approach.
Currently, DynIm uses L2 distances in the given high-dimensional space to define similarity and can be configured to use exact as well as approximate distances. Approximate distances are useful for computational viability for large data sizes and large data dimensionality. DynIm also provides a random sampler for comparison of sampling quality.
Researchers who want to run multiscale simulations.
Requires installation of faiss (outside of DynIm). After that, the code is pure-python and straightforward to use/edit.
The code is robust, however, approximate calculation of distances may require experimentation from the user to identify suitable parameters.
Data does not need to be preprocessed.
This is the first tool to perform “dynamic” sampling, that is, the input distribution can change over time, and the sampling adapts itself to the new distribution. This is the key feature that makes it possible to use this in-situ and enable large multiscale simulations.
Software: contained in this GitHub repository.
DynIm uses faiss to implement
nearest neighbor searches for sampling, and has been tested with faiss v1.6.3.
Currently, we ask the user to install faiss explicitly from source. Please
see here
for installation instructions.
Other dependencies are numpy and pyyaml (if needed, will be installed with
DynIm).
Once the dependencies are installed, DynIm can be installed as follows:
git clone https://github.com/CBIIT/NCI-DOE-Collab-Pilot2-DynIm
cd dynim
pip install .
Please test your installation as follows.
python3 -m unittest examples/test_dynim.py
See the examples directory.
DynIm is distributed under the terms of the MIT license.
See LICENSE and NOTICE for details.
SPDX-License-Identifier: (MIT)
LLNL-CODE-813147