GCAT (Graph Collaborate Attention Network)

Graph Collaborate Attention Network

Experiment Result

GCAT	H@1	H@10	MR	MRR
FB15K	70.08	91.64	38	0.784
FB15k-237	36.06	58.32	211	0.435
WN18RR	35.12	57.01	1974	0.430

Structure

Root
├── data
│   └── {dataset*} // Dataset
│   │   ├── train.txt
│   │   ├── test.txt
│   │   └── valid.txt
├── output
│   ├── {dataset*} // Result training of each dataset
│   │   ├── WN18RR_cuda_gat_3599.pt ~ "{dataset}_{device}_{model-name}_{last-epoch}"
│   │   └── WN18RR_cuda_result.txt
├── config.json # Config for traning
└── *.py # Source code
└── README.md

Installation

Public Colab : https://drive.google.com/file/d/1uVd_w6vE5C70rmgKLI7BvnhCWegXTMhk/view?usp=sharing

Requirements

Using Google Colab with :

• Python >= 3.6x
• Pytorch >= 1.x

Clone

git clone https://github.com/hmthanh/GCAT.git

Config

All config store in config.json file

"dataset": "WN18RR", # Dataset
"data_folder": "./data",
"output_folder": "./output",
"save_gdrive": false, # Use Google Drive to save object
"drive_folder": "/content/drive/My Drive",
"cuda": false, # Use GPU to training
"epochs_gat": 1,
"epochs_conv": 1,
"weight_decay_gat": 5e-06,
"weight_decay_conv": 1e-05,
"pretrained_emb": false,
"embedding_size": 50,
"lr": 0.001,
"get_2hop": true,
"use_2hop": true,
"partial_2hop": false,
"batch_size_gat": 86835,
"valid_invalid_ratio_gat": 2,
"drop_GAT": 0.3,
"alpha": 0.2,
"entity_out_dim": [100, 200],
"nheads_GAT": [2, 2],
"margin": 5,
"batch_size_conv": 128,
"alpha_conv": 0.2,
"valid_invalid_ratio_conv": 40,
"out_channels": 500,
"drop_conv": 0.0

Run

Because we use Google Colab for training, if you training with larger device, just run python main.py

• Step 1 : Create corpus

python 1_create_corpus.py

• Step 2 : Training embedding

python 2_training_encoder.py

• Step 3 : Training prediction

python 3_training_decoder.py

• Step 4 : Evaluation

python 4_evalution.py

Dataset

• FB15k (Free Base)
• FB15k-237
• WN18 (Word Net)
• WN18RR

Contact

Email : [email protected] | [email protected]

License

• MIT license
• Copyright 2020 © Minh-Thanh Hoang.

GCAT was modify from KBGAT repos (https://github.com/deepakn97/relationPrediction )