aipipeline is a library for running ai pipelines and monitoring the performance of the pipelines, e.g. accuracy, precision, recall, F1 score. This may include object detection, clustering, classification, and vector search algorithms. It is designed to be used for a number of projects at MBARI that require advanced workflows to process large amounts of images or video.

See the MBARI AI documentation for more information on the tools and services used in the pipelines.

---

Example plots from the t-SNE, confusion matrix and accuracy analysis of examplar data.

---

Requirements

Three tools are required to run the code in this repository:

Anaconda environment

This is a package manager for python. We recommend using the Miniconda version of Anaconda. Install on Mac OS X with the following command:

brew install miniconda

or on Ubuntu with the following command:

sudo apt install miniconda

Docker

This is a containerization tool that allows you to run code in a container.

just tool.

This is a handy tool for running scripts in the project. This is easier to use than make and more clean than bash scripts. Try it out!

Install on Mac OS X with the following command:

port install just

or on Ubuntu with the following command:

sudo apt install just

Installation

Clone the repository and run the setup command.

git clone http://github.com/mbari-org/aipipeline.git
cd aipipeline
just setup

Sensitive information is stored in a .env file in the root directory of the project, so you need to create a .env file with the following contents in the root directory of the project:

TATOR_TOKEN=your_api_token
REDIS_PASSWORD=your_redis_password
ENVIRONMENT=testing or production

Usage

Recipes are available to run the pipelines. To see the available recipes, run the following command:

just list

Recipe	Description
list	List recipes
install	Setup the environment
update_trackers	Update the environment. Run this command after checking out any code changes
plot-tsne-vss project='uav'	Generate a tsne plot of the VSS database
optimize-vss project='uav' *more_args=""	Optimize the VSS database
calc-acc-vss project='uav'	Calculate the accuracy of the VSS database; run after download, then optimize
reset-vss-all	Reset the VSS database, removing all data. Proceed with caution!!
reset-vss project='uav'	Reset the VSS database, removing all data. Run before init-vss or when creating the database. Run with e.g. uav
remove-vss project='uav' *more_args=""	Remove an entry from the VSS database, e.g., remove-vss i2map --doc 'doc:marine organism:*'
init-vss project='uav' *more_args=""	Initialize the VSS database for a project
load-vss project='uav'	Load already computed exemplars into the VSS database
cluster-uav *more_args=""	Cluster mission in aipipeline/projects/uav/data/missions2process.txt, add --vss to classify clusters with VSS
detect-uav *more_args=""	Detect mission in aipipeline/projects/uav/data/missions2process.txt
detect-uav-test	Detect mission data in aipipeline/projects/uav/data/missions2process.txt
load-uav-images	Load UAV mission images in aipipeline/projects/uav/data/missions2process.txt
load-uav type="cluster"	Load UAV detections/clusters in aipipeline/projects/uav/data/missions2process.txt
fix-uav-metadata	Fix UAV metadata lat/lon/alt
compute-saliency project='uav' *more_args=""	Compute saliency for downloaded VOC data and update the Tator database
crop project='uav' *more_args=""	Crop detections from VOC formatted downloads
download-crop-unknowns project='uav' labels='Unknown' download-dir='/tmp/download' *more_args=""	Download and crop Unknown detections
download project='uav'	Download only
predict-vss project='uav' image_dir='/tmp/download' *more_args=""	Predict images using the VSS database
run-ctenoA-test	Run the strided inference on a single video
run-ctenoA-prod	Run the strided inference on a collection of videos in a TSV file
run-mega-inference	Run the mega strided inference only on a single video
run-mega-track-bio	Run the mega strided tracking pipeline on a single video for the bio project
run-mega-track-i2map	Run the mega strided tracking pipeline on a single video for the i2map project

--

Related projects

• aidata -A tool to extract, transform, load and download operations on AI data.
• sdcat - Sliced Detection and Clustering Analysis Toolkit; a tool to detect and cluster objects in images.
• deepsea-ai - A tool to train and run object detection and tracking on video at scale in the cloud (AWS).
• fastapi-yolov5 - A RESTful API for running YOLOv5 object detection models on images either locally or in the cloud (AWS).
• fastapi-vss - A RESTful API for vector similarity search using foundational models.
• fastapi-tator - A RESTful API server for bulk operations on a Tator annotation database.

updated: 2024-10-22