Disclaimer: I am not working on this anymore. I will be happy to answer questions and review & merge PRs though.

This is a Keras & Tensorflow implementation of a captioning model. In particular, it uses the attention models described in this paper, which is depicted below:

where V are the K local features from the last convolutional layer of a ConvNet (e.g. ResNet-50), x_t is the input (composed of the embedding of the previous word and the average image feature). h_t is the hidden state of the LSTM at time t, which is used to compute the attention weights to apply to V in order to obtain the context vector c_t. c_t and h_t are combined to predict the current word y_t. In (b), an additional gate is incorporated into the LSTM to produce the additional s_t output, which is combined with V to compute the attention weights. s_t is used as an alternative feature to look at rather than the image features in V.

• Clone this repository

# Make sure to clone with --recursive
git clone --recursive https://github.com/amaiasalvador/sat_keras.git

• Install python 2.7.
• Install tensorflow 0.12.
• pip install -r requirements.txt
• (Optional )Install this Keras PR with support for layer-wise learning rate multipliers:

git clone https://github.com/amaiasalvador/keras.git
cd keras
git checkout lr_mult
python setup.py install

This option is disabled by default, so you can use "regular" keras 1.2.2 if you don't want to set a different learning rate to the base model.

• Set tensorflow as the keras backend in ~/.keras/keras.json:

{
    "image_dim_ordering": "tf", 
    "epsilon": 1e-07, 
    "floatx": "float32", 
    "backend": "tensorflow"
}

• Download MS COCO Caption Challenge 2015 dataset. Note that test images are not required for this code to work.
• After extraction, the dataset folder must have the following structure:

$coco/                                    # dataset dir
$coco/annotations/                        # annotations directory
$coco/annotations/captions_train2014.json # caption anns for training set
$coco/annotations/captions_val2014.json   # ...
$coco/images/                             # image dir
$coco/images/train2014                    # train image dir
$coco/images/val2014                      # ...

• Navigate to imcap/utils and run:

python prepro_coco.py --output_json path_to_json --output_h5 path_to_h5 --images_root path_to_coco_images

this will create the vocabulary and HDF5 file with data.

• [Coming soon] Download pretrained model here.

Unless stated otherwise, run all commands from ./imcap:

Run sample_captions.ipynb to test the trained network on some images and visualize attention maps.

Run python train.py. Run python args.py --help for a list of the available arguments to pass.

• Run python test.py to forward all validation images through a trained network and create json file with results. Use --cnntrain flag if evaluating a model with fine tuned convnet.
• Navigate to ./imcap/coco_caption/.
• From there run:

  python eval_caps.py -results_file results.json -ann_file gt_file.json
  

  to get METEOR, Bleu, ROUGE_L & CIDEr scores for the previous json file with generated captions.

For the sake of comparison, the data processing script follows the one in NeuralTalk2 and AdaptiveAttention.

• Xu et al. Show, Attend and Tell: Neural Image Caption Generation with Visual Attention. ICML 2015.
• Lu et al. Knowing When to Look: Adaptive Attention via A Visual Sentinel for Image Captioning. CVPR 2017 (original code here).
• Caption evaluation code from this repository.

For questions and suggestions either use the issues section or send an e-mail to [email protected].