Multi-Modal Image Retrieval System (MMIRS)

This repo holds the code for my Master Thesis at Language Technology (LT) Group of the University of Hamburg

NOTE THAT THIS IS WIP AND THE SYSTEM CHANGED A LOT. THIS README IS OUTDATED!

How to run locally

Assuming your in the root folder of this repository

Install pre-requisites

conda env create -v --file environment.yaml

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

Run the api locally

Make sure to change the path according to your setup! For more control edit: config.yaml (you can also set the data paths there and omit the environment variables)

Start the image server with the minimal config:

docker-compose -f docker-compose_img_svr.yml up -d

Start retrieval API

conda activate mt
export UNITER_DATA_DIR=/home/p0w3r/gitrepos/UNITER_fork/downloads
export UNITER_MODEL_CONFIG=/home/p0w3r/gitrepos/UNITER_fork/config/uniter-base.json
uvicorn main:app --host 0.0.0.0 --port 8081