NGNN

This is the code for the WebConf 2019 Paper: Dressing as a Whole: Outfit Compatibility Learning Based on Node-wise Graph Neural Networks.

Usage

Paper data and code

The original Polyvore dataset we used in our paper is first proposed here. After downloaded the datasets, you can put them in the folder NGNN/data/:

You can download the preprocessed data here, https://drive.google.com/open?id=1ibYEw0H9L9O9OLbxCiAlcZkt_IYuwKfd and also put them in the folder NGNN/data/.

There is a small dataset sample included in the folder NGNN/data/, which can be used to test the correctness of the code.

the data preprocess is written in the ./data/README.md

Quick Start

Then you can run the file NGNN/main_score.py to train the model.

You can change parameters according to the usage in NGNN/Config.py:

parameters arguments in `NGNN/Config.py`:

    epoch_num           the max epoch number
    train_batch_size    training batch size
    valid_batch_size    validation batch size
    hidden_size         hidden size of the NGNN
    lstm_forget_bias    forget bias in NGNN update
    max_grad_norm       the gradient clip during train
    init_scale          the scale of initialize parameter 0.05
    learning_rate       learning rate  0.01  # 0.001  # 0.2
    decay               the decay of 0.5
    decay_when = 0.002  # AUC
    decay_epoch = 200
    sgd_opt             train strategy can choose: 'RMSProp', 'Adam', 'Momentum', 'RMSProp', 'Adadelta'
    beta                the weight of regulartion
    GNN_step            the number of step of GNN
    dropout_prob        the dropout probability of our model
    adagrad_eps         eps
    gpu = 0             the gpu id            

Requirements

• Python 2.7
• Tensorflow 1.5.0

Citation

Please cite our paper if you use the code:

@inproceedings{cui2019dressing,
  title={Dressing as a Whole: Outfit Compatibility Learning Based on Node-wise Graph Neural Networks},
  author={Cui, Zeyu and Li, Zekun and Wu, Shu and Zhang, Xiao-Yu and Wang, Liang},
  booktitle={The World Wide Web Conference},
  pages={307--317},
  year={2019},
  organization={ACM}
}