README.md

DeepSeaAI is a Python package to simplify processing deep sea video in AWS from a command line.

It includes reasonable defaults that have been optimized for deep sea video. The goal is to simplify running these algorithms in AWS.

DeepSea-AI currently supports:

• Training YOLOv5 object detection models with up to 8 GPUs using the best available instances in AWS
• Processing video with YOLOv5 detection and tracking pipeline using
  • StrongSort tracking
• Scaling processing with AWS Elastic Container Service [click image below to see larger example]

The cost to process a video is typically less than $1.25 per 1-hour video using a model designed for a 640 pixel size.

The cost to train a YOLOv5 model depends on your data size and the number of GPUs you use.A large collection with 30K images and 300K localizations may cost $300-$600 to process, depending on the instance you choose to train on. This is reasonably small for a research project, and small in comparison to purchasing your own GPU hardware.

See the full documentation at MBARI deepsea-ai.

Processing

The processing technology uses the AWS Elastic Container Service with an architecture that includes a SQS messaging queue to start the processing. Simply upload a video to an S3 bucket then submit a job with the location of that video to the queue to start processing. The result is returned to a S3 bucket and the video is optionally removed to reduce storage cost.

Getting Started

Install

There are two main requirements to use this:

1. An account with AWS Amazon Web Services.
2. An account with Docker.
3. Install and update using pip in a Python>=3.8.0 environment:

After you have setup your AWS account, configure it using the awscli tool

pip install awscli
aws configure
aws --version

Then install directly from pypi

pip install deepsea-ai

Setting up the AWS environment is done with the setup mirror command. This only needs to be done once, or when you upgrade the module. This command will setup the appropriate AWS permissions and mirror the images used in the commands from Docker Hub to your ECR Elastic Container Registry.

Be patient - this takes a while, but only needs to be run once.

deepsea-ai setup --mirror

Tutorials

• FathomNet ✨ Recommended first step to learn more about how to train a YOLOv5 object detection model using freely available FathomNet data

Create the Anaconda environment

The fastest way to get started is to use the Anaconda environment. This will create a conda environment called deepsea-ai and make that available in your local jupyter notebook as the kernel named deepsea-ai

conda env create 
conda activate deepsea-ai
pip install ipykernel
python -m ipykernel install --user --name=deepsea-ai

Launch jupyter

cd docs/notebooks
jupyter notebook

---

Commands

• deepsea-ai setup --help - Setup the AWS environment. Must run this once before any other commands.
• deepsea-ai train --help - Train a YOLOv5 model and save the model to a bucket
• deepsea-ai process --help - Process one or more videos and save the results to a bucket
• deepsea-ai ecsprocess --help - Process one or more videos using the Elastic Container Service and save the results to a bucket
• deepsea-ai split --help - Split your training data. This is required before the train command.
• deepsea-ai monitor --help - Monitor processing. Use this after the ecsprocess train command.
• deepsea-ai -h - Print help message and exit.

Setting up an Elastic Container Service (ECS) cluster

To process videos in bulk, you can setup an ECS cluster to process videos in parallel. See the ECS setup documentation for more details.

---

Source code is available at github.com/mbari-org/deepsea-ai.

For more details, see the official documentation.