Turkish Image Captioning

This repository contains the project I worked on during my internship at Somera

Installing dependencies

First clone the repository and change directory

$ git clone https://github.com/badayvedat/turkish-image-captioning.git
$ cd turkish-image-captioning

Install dependencies. Installing on an isolated environment such as venv or conda is preferred.

$ pip install -r requirements.txt

Training

Preparing dataset

All of TasvirEt dataset used in validation. Because unlike COCO and Flickr30K captions, TasvirEt is not automatically translated, making it a better choice for validation.

Download and create dataset

$ sh data/download_data.sh

Make COCO dataset

$ python3 data/coco.py

Make Flickr30k dataset

$ python3 data/flickr30k.py

Make Flickr8k dataset

$ python3 data/flickr8k.py

Running training code

Create a config file with necessary parameters. For example config files check configs/ folder.

$ python3 train.py configs/12heads20kvoc.yaml

Run python3 train.py -h to see all parameters.

Also, you can use train.ipynb notebook.

Usage - Testing

Download config file, pre-trained weights and vocabulary file.

$ sh download_pretrained.sh

Currently only jpg files are supported.

Using Dockerfile

$ docker build -t turkish-caption .
$ docker run -p 6006:6006 turkish-caption

Using Flask

$ cd app
$ flask run

Using CLI

$ python3 evaluate.py path_to_image

Run python3 evaluate.py -h to see all parameters.