Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 

Repository files navigation

Multimodal Graph Representation Learning with Dynamic Information Pathways (DiP)

This is the official implementation of DiP

DiP framework

Preparation

Requirements

  • Python 3.7+
  • PyTorch 1.10.0+
  • PyTorch Geometric 2.1.0+

Our library versions are as follows:

  • torch 2.0.0+cu117
  • torch-geometric 2.3.0
  • torch-scatter 2.1.1+cu117
  • torch-sparse 0.6.17+cu117
  • torchmetrics 1.2.0

Data Preparation

Download our datasets from this link. You may save them to any directory you like, such as ./Multimodal-Graph-Completed-Graph. The structure should look like the following tree diagram. You can easily add new datasets following this format.

.
├── books-lp
│   ├── lp-edge-split-random.pt
│   ├── clip_feat.pt
│   ├── imagebind_feat.pt
│   ├── t5vit_feat.pt
│   └── t5dino_feat.pt
├── sports-copurchase
│   ├── lp-edge-split-hard.pt
│   ├── clip_feat.pt
│   ├── imagebind_feat.pt
│   ├── t5vit_feat.pt
│   └── t5dino_feat.pt
├── cloth-copurchase
│   ├── lp-edge-split-hard.pt
│   ├── clip_feat.pt
│   ├── imagebind_feat.pt
│   ├── t5vit_feat.pt
│   └── t5dino_feat.pt
├── ele-fashion
│   ├── nc_edges-nodeid.pt
│   ├── split.pt
│   ├── labels-w-missing.pt
│   ├── clip_feat.pt
│   ├── imagebind_feat.pt
│   ├── t5vit_feat.pt
│   └── t5dino_feat.pt
└── books-nc
    ├── nc_edges-nodeid.pt
    ├── split.pt
    ├── labels-w-missing.pt
    ├── clip_feat.pt
    ├── imagebind_feat.pt
    ├── t5vit_feat.pt
    └── t5dino_feat.pt

Raw Images

Raw images can be downloaded by using node_mapping.pt (which provides 1-1 mapping for node id and raw file id) for each dataset. A reference code for downloading can be found in download_img.py. The products metadata can be obtained from: https://cseweb.ucsd.edu/~jmcauley/datasets.html#amazon_reviews and https://mengtingwan.github.io/data/goodreads.html. The zipped folder can be found here.

Raw text

Raw images can be downloaded by using node_mapping.pt (which provides 1-1 mapping for node id and raw file id) for each dataset. The products metadata can be obtained from: https://cseweb.ucsd.edu/~jmcauley/datasets.html#amazon_reviews and https://mengtingwan.github.io/data/goodreads.html. The zipped folder can be found here.

Getting Started

  • For Link Prediction Task:
cd lp
CUDA_VISIBLE_DEVICES=0 python train.py --config_path configs/sports-copurchase.yaml --npnode_v 128 --npnode_t 32
  • For Node Classification Task:
cd nc
CUDA_VISIBLE_DEVICES=5 python train.py --config_path configs/ele-fashion.yaml --npnode_v 512 --npnode_t 64

Acknowledgements

About

Multimodal Graph Representation Learning with Dynamic Information Pathways

Resources

Stars

4 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors