yonyeoseok / ggnn

Situation Recognition with Graph Neural Networks

This is the Torch implementation of Situation Recognition with Graph Neural Networks

Setup

• Train a CNN to predict 504 verbs with Cross-Entropy loss
• Train a CNN to predict top 2K most frequent nouns with Binary Cross-Entropy loss
• Extract the CNN features for each image and save them together with roles information into a HDF5 file

Training

Specify several options and then run th train.lua

• -input_h5: The input HDF5 file
• -embedding_size: Embedding size of verb and role
• -rnn_size: GGNN hidden state dimension
• -num_updates: Number of updates in GGNN
• -checkpoint_path: Where to save the models

Test

There are two steps to test the model

1. Save the outputs by running th eval.lua, the options are the same as training
2. Calculate the accuracy th clc_accuracy.lua

Cite

If you use this code, please consider citing

@inproceedings{li2017situation, 
title={Situation recognition with graph neural networks}, 
author={Li, Ruiyu and Tapaswi, Makarand and Liao, Renjie and Jia, Jiaya and Urtasun, Raquel and Fidler, Sanja}, 
booktitle={ICCV}, year={2017} 
}