ranzcr_clip

This repo contains the code for RANZCR CLiP Competition on Kaggle.

Interactive Demo

You can play with already trained models here.

Usage

1. Prepare

Clone the repo and install requirements

git clone https://github.com/asnorkin/ranzcr_clip.git
python -m pip install -r requirements.txt

Install DVC, load the dataset and pretrained models

python -m pip install dvc dvc[gdrive]
dvc pull

Or just models

dvc pull ranzcr_clip/models

2. Training

Run training

cd ranzcr_clip
./train_segmentation.sh

Check training output log

tail -F -n 10000 segmentation/train_outputs/<experiment>.log

Visualize training logs on tensorboard

tensorboard --logdir=segmentation/logs

3. Run Demo with custom model

Copy trained model into the service

cp segmentation/checkpoints/<experiment>/<model_name>.ckpt models/<model>/<model_name>.ckpt

Run the demo with new model locally

PYTHONPATH=ranzcr_clip:${PYTHONPATH} streamlit run app/app.py

4. Deploy demo

docker-compose up --build -d