Ship Segmentation

• Setup
• App
• Pipelines
• Library
• EDA
• Next Steps

Setup

First clone the repo and download the data.

GPU Requirements if you want to use GPU you must make sure you have your CUDA environment set up properly and you update the requirements-cuda.txt to point to the correct index based on your cuda driver. I have cuda 11.8 so I point to that index. See https://pytorch.org/get-started/locally/ for details on available cuda installs.

Docker

A Dockerfile is provided to run the shiny app.

# data dir must be present before building image
docker build -t ship-seg .
docker run \
    -d \
    --name ship-seg \
    --gpus all \
    -p 8000:8000 \
    ship-seg

Local

Local was run on python==3.10.`

# make env
mkvirtualenv shipseg
# install reqs
pip install -r requirements.txt
pip install -r requirements-cuda.txt
# install model library (shipseg)
pip install -e .

App

To view predictions and evaluation on the test set images call:

shiny run app.py

and navigate to the reported url. The first time running it will likely need to download fasterrcnn weights.

To look at model training runs on MLFlow call:

mlflow server -h 0.0.0.0

and navigate to the reported url.

Pipelines

Training

To train and save a model. Run info will be logged to the local MLFlow directory.

python script/train.py config/baseline.yaml