ATTEND-GAN Model

TensorFlow implementation of Towards Generating Stylized Image Captions via Adversarial Training.

Reference

if you use our codes or models, please cite our paper:

@article{nezami2019towards,
  title={Towards Generating Stylized Image Captions via Adversarial Training},
  author={Nezami, Omid Mohamad and Dras, Mark and Wan, Stephen and Paris, Cecile and Hamey, Len},
  journal={arXiv preprint arXiv:1908.02943},
  year={2019}
}

Data

We pretrain our models using Microsoft COCO Dataset. Then, we train the models using SentiCap Dataset.

Requiremens

1. Python 2.7.12
2. Numpy 1.15.2
3. Hickle
4. Python-skimage
5. Tensorflow 1.8.0

Content

1. Model Train Code
2. Model Test Code
3. ATTEND-GAN Generator
4. ATTEND-GAN Discriminator

Train

1. Download Microsoft COCO Dataset including neutral image caption data and SentiCap Dataset including sentiment-bearing image caption data.
2. Reseize the downloded images into [224, 224] and put them in "./images".
3. Preprosses the COCO image caption data and place them in "./data/neutral". You can do this by prepro.py and the ResNet-152 network trained on ImageNet, which is generating [7,7,2048] feature map (we use the Res5c layer of the network).
4. Preprosses the SentiCap image caption data and place its positve part in "./data/positive" and its negative part in "./data/negative". You can do this by prepro.py and the ResNet-152 network trained on ImageNet, which is generating [7,7,2048] feature map (we use the Res5c layer of the network).
5. Pretrain the generator and discriminator using "./data/neutral". (python model_train.py)
6. Train the generator and the discriminator using "./data/positive" for the positive part and "./data/negative" for the negative part. (python model_train.py)

Test

1. Dowload pretrained models and unzip the models in "./models".
2. python model_test.py

Results

	BLEU-1	BLEU-4	METEOR	ROUGE-L	CIDEr	SPICE
ATTEND-GAN	56.55%	13.05%	18.35%	44.45%	62.85%	16.05%

ATTEND-GAN is inspired from Self-critical Sequence Training and SeqGAN in TensorFlow.