Skip to content

Latest commit

 

History

History
51 lines (42 loc) · 2.83 KB

README.md

File metadata and controls

51 lines (42 loc) · 2.83 KB

Dialog-based interactive image retrieval

About this repository

This repository contains an implementation of the models introduced in the paper Dialog-based Interactive Image Retrieval. The network is implemented using PyTorch and the rest of the framework is in Python. The user model is built directly on top of ImageCaptioning.pytorch.

Citing this work

If you find this work useful in your research, please consider citing:

@incollection{NIPS2018_7348,
title = {Dialog-based Interactive Image Retrieval},
author = {Guo, Xiaoxiao and Wu, Hui and Cheng, Yu and Rennie, Steven and Tesauro, Gerald and Feris, Rogerio},
booktitle = {Advances in Neural Information Processing Systems 31},
editor = {S. Bengio and H. Wallach and H. Larochelle and K. Grauman and N. Cesa-Bianchi and R. Garnett},
pages = {676--686},
year = {2018},
publisher = {Curran Associates, Inc.},
url = {http://papers.nips.cc/paper/7348-dialog-based-interactive-image-retrieval.pdf}
}

Project page

The project page is available at https://www.spacewu.com/posts/fashion-retrieval/.

Dependencies

To get started with the framework, install the following dependencies:

Dataset

The dataset used in the paper is built on the Attribute Discovery Dataset. Please refer to the dataset README for our dataset details. The pre-computed image features and user captioning model can be downloaded from here.

Train and evaluate a model

Follow the steps to train a model:

  1. Prepare following dataset instructions or download the pre-computed image features and user captioning model from here.
  2. Move the pre-computed image features into features folder and the user captioning model into caption_models folder.
  3. Edit the training script experiments/scripts/train.sh such that all paths agree with the files on your file system.
  4. To train the model with the pre-training loss, run:
python train_sl.py --log-interval=50 --lr=0.001  --batch-size=128 --model-folder="models/"

The program saves trained models into the folder models/ every epoch.

  1. To fine-tune the final model with the policy improvement loss, run:
python train_rl.py --log-interval=10 --lr=0.0001 --top-k=4 --batch-size=128 --tau=0.2 --pretrained-model="models/sl-10.pt" --save-folder="models/"

The program saves trained models into the folder models/ every epoch.

License

MIT License