Grammatically Recognizing Images with Tree Convolution

This is a re-implementation of our KDD 2020 paper "Grammatically Recognizing Images with Tree Convolution" (https://dl.acm.org/doi/abs/10.1145/3394486.3403133).

Guangrun Wang, Guangcong Wang, Keze Wang, Xiaodan Liang, and Liang Lin*

Sun Yat-sen University (SYSU)

Table of Contents

0. Introduction
1. Requirement
2. Pretrained model on ImageNet
3. Training on ImageNet
4. Citation

Introduction

This repository contains the training & testing code of "Grammatically Recognizing Images with Tree Convolution" (TreeConv) on ImageNet.

Requirement

This code was tested on：

• Python 3.6.7
• TensorFlow 1.15.0
• Tensorpack The code depends on Yuxin Wu's Tensorpack. The code was tested on tensorpack-0.9.8 where "0.9.8" denotes the tensorpack version.

  # install tensorpack local:
  cd tensorpack/tensorpack/
  python setup.py install --user
  

Pretrained model on ImageNet

• ImageNet accuracy and pretrained model (baidu pan code: ow9z):

Model	Top 5 Error	Top 1 Error	Download
ResNet50	6.9%	23.6%	⬇️
ResNet50-TreeConv	6.16%	22.08%	Baidu pan: ⬇️ ; passport: ijcb

• Testing script:

cd TreeConv

python imagenet-resnet.py  --gpu 0,1,2,3,4,5,6,7   --data [ROOT-OF-IMAGENET-DATASET]  --log_dir  [ROOT-OF-TEST-LOG] --load   [ROOT-TO-LOAD-MODEL]  --eval --data-format NHWC

Training on ImageNet

• Training script:

cd TreeConv

python imagenet-resnet.py  --gpu 0,1,2,3,4,5,6,7   --data [ROOT-OF-IMAGENET-DATASET]  --log_dir  [ROOT-OF-TRAINING-LOG-AND-MODEL]  --data-format NHWC

Citation

If you use these models in your research, please cite:

@inproceedings{Wang2020Grammatically_KDD,
  author    = {Guangrun Wang and
               Guangcong Wang and
               Keze Wang and
               Xiaodan Liang and
               Liang Lin},
  title     = {Grammatically Recognizing Images with Tree Convolution},
  booktitle = {{KDD} '20: The 26th {ACM} {SIGKDD} Conference on Knowledge Discovery
               and Data Mining, Virtual Event, CA, USA, August 23-27, 2020},
  pages     = {903--912},
  year      = {2020},
}