Home

Awesome

Multi-grained Uncertainty Regularization

Offical Implementation of Composed Image Retrieval with Text Feedback via Multi-grained Uncertainty Regularization.

pipeline

The baseline code is borrowed from CoSMo. We would like to thank this great work.

Install

Prerequisites

Required packages

pip install -r requirements.txt

Dataset

Download any of the three data sets, FashionIQ, Shoes and Fashion200k, to run the code.

We provided the three datasets in Google Drive temporarily because some image links are not permanent. Follow the requirements of the original author, datasets are only used for academic purposes.

Copy the dataset folders to the data folder. The data folder is structured as follows:

├── data/
│   ├── abc.py
│   ├── collate_fns.py
│   ├── fashion200k.py
│   ├── fashionIQ.py
│   ├── __init__.py
│   ├── shoes.py
│   └── utils.py
│   ├── fashionIQ/
│   │   ├── captions/
│   │   ├── captions_pairs/
│   │   ├── image_data/
│   │   ├── image_splits/
│   │   ├── image_tag_dataset/
│   │   ├── fashion_iq_vocab.pkl
│   │   ├── ...
│   ├── fashion200k/
│   │   ├── labels/
│   │   ├── women/
│   │   ├── fashion200k_vocab.pkl
│   │   ├── ...
│   ├── shoes/
│   │   ├── attributedata/
│   │   ├── shoes_vocab.pkl
│   │   ├── ...

The Core Code

Core code is relatively simple, and could be directly applied to other works.

Weights & Biases

We use Weights and Biases to log our experiments. You can register an account or provide an existing one, head it over to *config.json and fill out your wandb_account_name.You can also change the default at options/command_line.py.

Run

You can run the code by the following command:

CUDA_VISIBLE_DEVICES=0 python3 main.py --config_path=configs/fashionIQ_config.json --experiment_description=test_fashionIQ --device_idx=$CUDA_VISIBLE_DEVICES

Citation

@inproceedings{chen2024composed,
author = "Chen, Yiyang and Zheng, Zhedong and Ji, Wei and Qu, Leigang and Chua, Tat-Seng",
title = "Composed Image Retrieval with Text Feedback via Multi-grained Uncertainty Regularization",
booktitle = "International Conference on Learning Representations (ICLR)",
code = "https://github.com/Monoxide-Chen/uncertainty\_retrieval",
year = "2024"
}