rip_current_detection

Download model weights

# From rip_current_detection repo root
RCD_DIR="models/rip_current_detector/1" && \
    mkdir -p "${RCD_DIR}" && \
    wget https://www.dropbox.com/s/dcsdi36jbc570u9/fasterrcnn_resnet50_fpn.pt -O "${RCD_DIR}/fasterrcnn_resnet50_fpn.pt";

Set up environment

Python environment setup

micromamba create -f environment.yml
micromamba activate rip_current_detection

GPU setup on host

This project can run against an NVIDIA CUDA-capable GPU, which will greatly accelerate detection requests. You must have the following installed and configured:

• An NVIDIA GPU, capable of CUDA 11.x or 12.x, with up-to-date drivers installed for your target operating system.

• If running within a Docker container, you will need the nvidia-container-toolkit to be installed, and the nvidia container runtime configured as runtime within your Docker daemon.

If CUDA capabilities are detected at runtime, the API should detect and use the device (using torch.cuda.is_available('cuda')). If not, the cpu device is used as a fallback.

GPU hardware developed and tested on:

• OS: Pop!_OS/Ubuntu 22.04
  CPU: Intel Xeon CPU E3-1245 v5 @ 3.50GHz
  GPU: Quadro P1000 (4GB VRAM)
  Driver: 550.67
  CUDA: 12.4
  nvidia-container-toolkit: 1.12.1

Run app

Run on the host in 'development' mode (restarting on file changes):

uvicorn api:app --port 8888 --reload

To run in a Docker container (using the default docker-compose.yml):

docker compose up

To run a Docker container with the NVIDIA GPU capabilities assigned to the container, you should use the docker-compose.gpu.yml Docker compose spec, like so:

docker compose -f docker-compose.gpu.yml up