This rebo is a pytorch implementation of Show, Attend and Tell: Neural Image Caption Generation with Visual Attention on Flickr 8k Dataset which consists of 8,000 images that are each paired with five different captions which provide clear descriptions of the salient entities and events.
Code is developed under following library dependencies
python 3.7
torch 1.10.1
torchtext 0.11.1
torchvision 0.11.2
Start with creating a virtual environment then open your terminal and follow the following steps:
git clone "https://github.com/zaghlol94/Image-captioning"
cd Image-captioning
pip install -r requirements.txt
bash download_assets.sh
cd src
python inference.py -i "image path"
Flickr8k consists of 8,000 images that are each paired with five different captions which provide clear descriptions of the salient entities and events. you could download the dataset from here.
├── captions.txt
└── Images
├── 10815824_2997e03d76.jpg
├── 12830823_87d2654e31.jpg
├── 17273391_55cfc7d3d4.jpg
├── 19212715_20476497a3.jpg
├── 3637013_c675de7705.jpg
└── 667626_18933d713e.jpg
There is Images folder contains all the images and caption file consist of two columns, the first columns is the name of the image and the second columns is the caption.
To train the model on custom data-set of different image-caption pair, you could add your data files in the same structure described in the previous section and change the root folder/caption files' path in config.py then follow the following steps in the command line:
cd src
python train.py
After these steps you will find my_checkpoint.pth.tar in src folder, which is the weights of the seq2sq model.
you could figure out how the loss declined over epochs using tensorboard by writing in the command line:
tensorboard --logdir=runs/
then go to http://localhost:6006/
@misc{https://doi.org/10.48550/arxiv.1502.03044,
doi = {10.48550/ARXIV.1502.03044},
url = {https://arxiv.org/abs/1502.03044},
author = {Xu, Kelvin and Ba, Jimmy and Kiros, Ryan and Cho, Kyunghyun and Courville, Aaron and Salakhutdinov, Ruslan and Zemel, Richard and Bengio, Yoshua},
keywords = {Machine Learning (cs.LG), Computer Vision and Pattern Recognition (cs.CV), FOS: Computer and information sciences, FOS: Computer and information sciences},
title = {Show, Attend and Tell: Neural Image Caption Generation with Visual Attention},
publisher = {arXiv},
year = {2015},
copyright = {arXiv.org perpetual, non-exclusive license}
}