Single-Modal and Cross-Modal Retrieval Tasks

Dependencies

We recommended to use Anaconda for the following packages: Python 2.7, PyTorch (>0.1.12), NumPy (>1.12.1), TensorBoard

Punkt Sentence Tokenizer:

import nltk
nltk.download()
> d punkt

Download data

In this exeripments, we used MSCOCO image-caption as the dataset, for single model (image or text), you can only use image/text as the training data, corss-modal retrieval shares the same framework with single modal retrieval. You can download the data from MSCOCO image captioning website, or download the precomputed image features are from here and here. To use full image encoders, download the images from their original sources here, here and here.

wget http://www.cs.toronto.edu/~faghri/vsepp/vocab.tar
wget http://www.cs.toronto.edu/~faghri/vsepp/data.tar
wget http://www.cs.toronto.edu/~faghri/vsepp/runs.tar

We refer to the path of extracted files for data.tar as $DATA_PATH and files for models.tar as $RUN_PATH. Extract vocab.tar to ./vocab directory.

Training

Run train.sh:

python train.sh 0

Evaluation

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

Visualize Word Embeddings

python visualize_w2v.py

Reference

@article{gu2017look,
  title={Look, Imagine and Match: Improving Textual-Visual Cross-Modal Retrieval with Generative Models},
  author={Gu, Jiuxiang and Cai, Jianfei and Joty, Shafiq and Niu, Li and Wang, Gang},
  journal={CVPR},
  year={2018}
}