/sca-cnn

image captions generation with spatial and channel-wise attention

Primary LanguagePython

SCA-CNN

Source code for the paper: SCA-CNN: Spatial and Channel-wise Attention in Convolution Networks for Imgae Captioning

This code is modified based on two previous works arctic-captions and arctic-capgen-vid.

This code is only for two layers attention model in ResNet-152 Network for MS COCO dataset. Other networks (VGG-19) or other datasets (Flickr30k/Flickr8k) can also be used by some little modifications.

Dependencies

  1. This code is written in python with powerful Theano library.

  2. Some other python package dependencies like numpy/scipy, skimage, opencv, sklearn, hdf5 module can be installed by pip, or directly run command $ pip install -r requirements.txt

  3. For image CNN feature extraction, we use Caffe. You should install caffe and building the pycaffe interface to extract the image CNN feature.

  4. For results evaluation, we use the official coco evaluation scrpits coco-caption. Install it by simply adding it into $PYTHONPATH.

Getting Started

  • Get the code $ git clone the repo and install the dependencies

  • Save the pretrained CNN weights Save the ResNet-152 weights pretrained on ImageNet. Before running the code, set the variable deploy and model in save_resnet_weight.py to your own path. Then run:

$ cd cnn
$ python save_resnet_weight.py
  • Preprocessing the dataset For the preprocessing of captioning, we directly use the processed JSON blob from neuraltalk. Similar as step 2, set the PATH in cnn_until.py and make_coco.py to your own install path. Then :
$ cd data
$ python make_coco.py
  • Training The results are saved in the directory exp.
$ THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32 python sca_resnet_branch2b.py

Citation

If you found this code useful, please cite the following paper:

@inproceedings{chen2016sca,
  title={SCA-CNN: Spatial and Channel-wise Attention in Convolutional Networks for Image Captioning},
  author={Chen, Long and Zhang, Hanwang and Xiao, Jun and Nie, Liqiang and Shao, Jian and Liu, Wei and Chua, Tat-Seng},
  booktitle={CVPR},
  year={2017}
}