README.md

VS-CNN

This is the code for the paper:

Visual Sentiment Analysis for Review Images with Item-Oriented and User-Oriented CNN
Quoc-Tuan Truong and Hady W. Lauw
Presented at MM 2017

We provide:

• Code to train and evaluate the models
• Data used for the experiments
• Pre-trained weights of the base models

If you find the code and data useful in your research, please cite:

@inproceedings{vs-cnn,
  title={Visual sentiment analysis for review images with item-oriented and user-oriented CNN},
  author={Truong, Quoc-Tuan and Lauw, Hady W},
  booktitle={Proceedings of the ACM on Multimedia Conference},
  year={2017},
}

Requirements

• Python 3
• Tensorflow > 1.0
• Tqdm
• Pre-trained weights of AlexNet for initialization

Training and Evaluation

• Base model:

python train_base.py --dataset [user,business]

• Factor model:

To train the factor models, we need pre-trained weights from the base models for initialization. If you want to save time, the weights can be downloaded from here.

python train_factor.py --dataset [user,business] --factor_layer [conv1,conv3,conv5,fc7] --num_factors 16