ucsdtnel/autolfads repository overview

⁠AutoLFADS

This registry provides a set of images for running LFADS/AutoLFADS⁠ in different compute environments and workflows.

⁠Local Installation & Usage

Running LFADS in a container provides isolation from your host operating system and instead relies on a system installed container runtime. This workflow is suitable for evaluating algorithm operation on small datasets or exploring specific model parameter changes. It is suitable for use on shared compute environments and other platforms where there is limited system package isolation.

Prerequisites: Container runtime (e.g. Docker - Linux / Mac / Windows⁠, Podman - Linux / Mac / Windows⁠, containerD - Linux / Windows⁠) and the Nvidia Container Toolkit⁠ (GPU only).

1. Specify latest for CPU operation and latest-gpu for GPU compatible operation

    TAG=latest

1. (OPTIONAL) Pull the docker image to your local machine. This step ensures you have the latest version of the image.

    docker pull ucsdtnel/autolfads:$TAG

1. Browse to a directory that has access to your data and LFADS configuration file

    #    The general structure should be as follows (names can be changed, just update the paths in the run parameters)
    #    \<my-data-directory>
    #       \data
    #          <data files>
    #          config.yaml (LFADS model parameter file)
    #       \output
    #          <location for generated outputs>
    cd <my-data-directory>

1. Run LFADS

    docker run --rm -it -v $(pwd):/share ucsdtnel/autolfads:$TAG \
        --data /share/data \
        --checkpoint /share/output \
        --config-file /share/data/config.yaml

⁠KubeFlow Usage

Running AutoLFADS using KubeFlow enables scaling your experiments across an entire cluster. This workflow allows for isolated multi-user utilization and is ideal for running on managed infrastructure (e.g. University, public or private cloud) or on service-oriented clusters (i.e. no direct access to compute instances). It leverages industry standard tooling and enables scalable compute workflows beyond AutoLFADS for groups looking to adopt a framework for scalable machine learning.

Prerequisites: KubeFlow >= 0.14

Directly run experiments using the ucsdtnel/autolfads images (example⁠)