deinked

Port of https://github.com/vijishmadhavan/SkinDeep from FastAI 1 to FastAI 2.

Running Deinked

Locally

To run the Jupyter Notebooks I use a an extended version of the FastAI docker containers. https://github.com/fastai/docker-containers. The changes I've made are:

• Changed the root from fastai to local
• Changed the base ubuntu image to an NVIDIA/CUDA one (nvidia/cuda:11.3.1-base-ubuntu20.04) for GPU training
• Added an image (-ext) with my dependencies and endpoint
  • Locally install ffmpeg libsm6 libxext6
  • Pip install nvidia-ml-py3 opencv-python Pillow

This is the command that I use to run the customized container.

docker run --rm \
    --gpus all --privileged \
    --name fastai --ipc=host \
    -p 8888:8888 \
    -v `pwd`:/home \
    local/fastai-ext \
    jupyter notebook

Collab

Haven't set this up yet

Setting everything up

Prepare the training data

Place your training set in data/rawdata. The tattoo image and clean image should be named _tattoo.jpeg and _clean.jpeg respectively.

Run the Deink - Process Raw Data notebook. This will prep the images for training and put them in data/tattoo and data/clean respectively.

Running training

The Deink - Train Model notebook, unsurprisingly, is for training the model. There is a cell for setting training variables.

Predicting images

Use the Deink - Predict Image notebook to load the model and process an image of your choosing.