Introduction

This is Stacked Cross Attention Network, source code of Stacked Cross Attention for Image-Text Matching (project page) from Microsoft AI and Research. The paper will appear in ECCV 2018. It is built on top of the VSE++ in PyTorch.

Requirements and Installation

We recommended the following dependencies.

Python 2.7
PyTorch 0.3
NumPy (>1.12.1)
TensorBoard
Punkt Sentence Tokenizer:

import nltk
nltk.download()
> d punkt

Download data

Download the dataset files and pre-trained models. We use splits produced by Andrej Karpathy. The raw images can be downloaded from from their original sources here, here and here.

The precomputed image features of MS-COCO are from here. The precomputed image features of Flickr30K are extracted from the raw Flickr30K images using the bottom-up attention model from here. All the data needed for reproducing the experiments in the paper, including image features and vocabularies, can be downloaded from:

wget https://scanproject.blob.core.windows.net/scan-data/data.zip
wget https://scanproject.blob.core.windows.net/scan-data/vocab.zip

We refer to the path of extracted files for data.zip as $DATA_PATH and files for vocab.zip to ./vocab directory. Alternatively, you can also run vocab.py to produce vocabulary files. For example,

python vocab.py --data_path data --data_name f30k_precomp
python vocab.py --data_path data --data_name coco_precomp

Data pre-processing (Optional)

The image features of Flickr30K and MS-COCO are available in numpy array format, which can be used for training directly. However, if you wish to test on another dataset, you will need to start from scratch:

Use the bottom-up-attention/tools/generate_tsv.py and the bottom-up attention model to extract features of image regions. The output file format will be a tsv, where the columns are ['image_id', 'image_w', 'image_h', 'num_boxes', 'boxes', 'features'].
Use util/convert_data.py to convert the above output to a numpy array.

If downloading the whole data package containing bottom-up image features for Flickr30K and MS-COCO is too slow for you, you can download the following package with everything but image features and compute image features locally from raw images.

wget https://scanproject.blob.core.windows.net/scan-data/data_no_feature.zip

Training new models

Run train.py:

python train.py --data_path "$DATA_PATH" --data_name coco_precomp --vocab_path "$VOCAB_PATH" --logger_name runs/coco_scan/log --model_name runs/coco_scan/log --max_violation --bi_gru

Arguments used to train Flickr30K models:

Method	Arguments
SCAN t-i LSE	`--max_violation --bi_gru --agg_func=LogSumExp --cross_attn=t2i --lambda_lse=6 --lambda_softmax=9`
SCAN t-i AVG	`--max_violation --bi_gru --agg_func=Mean --cross_attn=t2i --lambda_softmax=9`
SCAN i-t LSE	`--max_violation --bi_gru --agg_func=LogSumExp --cross_attn=i2t --lambda_lse=5 --lambda_softmax=4`
SCAN i-t AVG	`--max_violation --bi_gru --agg_func=Mean --cross_attn=i2t --lambda_softmax=4`

Arguments used to train MS-COCO models:

Method	Arguments
SCAN t-i LSE	`--max_violation --bi_gru --agg_func=LogSumExp --cross_attn=t2i --lambda_lse=6 --lambda_softmax=9 --num_epochs=20 --lr_update=10 --learning_rate=.0005`
SCAN t-i AVG	`--max_violation --bi_gru --agg_func=Mean --cross_attn=t2i --lambda_softmax=9 --num_epochs=20 --lr_update=10 --learning_rate=.0005`
SCAN i-t LSE	`--max_violation --bi_gru --agg_func=LogSumExp --cross_attn=i2t --lambda_lse=20 --lambda_softmax=4 --num_epochs=20 --lr_update=10 --learning_rate=.0005`
SCAN i-t AVG	`--max_violation --bi_gru --agg_func=Mean --cross_attn=i2t --lambda_softmax=4 --num_epochs=20 --lr_update=10 --learning_rate=.0005`

Evaluate trained models

from vocab import Vocabulary
import evaluation
evaluation.evalrank("$RUN_PATH/coco_scan/model_best.pth.tar", data_path="$DATA_PATH", split="test")

To do cross-validation on MSCOCO, pass fold5=True with a model trained using --data_name coco_precomp.

Reference

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

@article{lee2018stacked,
  title={Stacked Cross Attention for Image-Text Matching},
  author={Lee, Kuang-Huei and Chen, Xi and Hua, Gang and Hu, Houdong and He, Xiaodong},
  journal={arXiv preprint arXiv:1803.08024},
  year={2018}
}

License

Apache License 2.0

Acknowledgments

The authors would like to thank Po-Sen Huang and Yokesh Kumar for helping the manuscript. We also thank Li Huang, Arun Sacheti, and Bing Multimedia team for supporting this work.

songzuolong/SCAN