perception-benchmark

Overview

A benchmark that evaluates and analyzes multi-modal perception models.

Usage

Run locally

1. install the conda environment with docker/env.yml
2. install packages in (utils_package)[https://github.com/AutonomousSystemsResearch/utils_package].
3. run pip install -e . in the root directory of this repo
4. update configs files. For example, mounting directory for azure storage.
5. run python tasks/train.py base=configs/multiMAE.yaml to pretrain multi-MAE model over habitat dataset

Run aml

The jobs/amlt.yaml is already set up.

Note

Certain directories in this repo are cloned from exist repos as submodules. When you first git clone this repo, those folders of submodules will be empty.You must run two commands: git submodule init to initialize your local configuration file, and git submodule update to fetch all the data from that project and check out the appropriate commit listed in your superproject. Or you can pass --recurse-submodules to the git clone command, it will automatically initialize and update each submodule in the repository, including nested submodules if any of the submodules in the repository have submodules themselves.

Those submodules are usually official repo of open-sourced codes, you should avoid modifying them unless necessary.