liangjiubujiu / ggnn.pytorch

A PyTorch Implementation of Gated Graph Sequence Neural Networks (GGNN) and Residual Gated Graph ConvNets (RGGC) for FYP

Final Year Project Code

This repository contains the implementation of the graph neural networks implemented for the Final Year Project (FYP). It includes models for Gated Graph Neural Networks (GGNN) and Residual Gated Graph ConvNets (RGGC). This was originally forked from JamesChuanggg/ggnn.pytorch and modified to include the RGGC model here. Both models are tested against the bAbi tasks dataset. Here's an example of bAbI deduction task (task 15):

The blog for this project can be found here.

Requirements

• python==3.7
• PyTorch>=1.0
• XlsxWriter>=1.1.5

Installation

Using conda environments:

conda env create -f environment.yml
activate fyp

Run

View help:

python main.py --help

Suggesting configurations for each task (GGNN):

# task 1
python main.py --net "GGNN" --task_id 1 --lr 0.005 --state_dim 8 --n_steps 5 --niter 100 --cuda
# task 2
python main.py --net "GGNN" --task_id 2 --lr 0.005 --state_dim 4 --n_steps 5 --niter 100 --cuda
# task 4
python main.py --net "GGNN" --task_id 4 --lr 0.05 --state_dim 10 --n_steps 10 --niter 100 --cuda
# task 9
python main.py --net "GGNN" --task_id 9 --lr 0.005 --state_dim 8 --n_steps 5 --niter 100 --cuda
# task 11
python main.py --net "GGNN" --task_id 11 --lr 0.005 --state_dim 8 --n_steps 10 --niter 100 --cuda
# task 12
python main.py --net "GGNN" --task_id 12 --lr 0.01 --state_dim 8 --n_steps 5 --niter 100 --cuda
# task 13
python main.py --net "GGNN" --task_id 13 --lr 0.005 --state_dim 8 --n_steps 5 --niter 100 --cuda
# task 15
python main.py --net "GGNN" --task_id 15 --lr 0.05 --state_dim 4 --n_steps 10 --niter 100 --cuda
# task 16
python main.py --net "GGNN" --task_id 16 --lr 0.01 --state_dim 8 --n_steps 5 --niter 100 --cuda
# task 17
python main.py --net "GGNN" --task_id 17 --lr 0.005 --state_dim 4 --n_steps 10 --niter 100 --cuda
# task 18
python main.py --net "GGNN" --task_id 18 --lr 0.005 --state_dim 8 --n_steps 5 --niter 100 --cuda

Suggesting configurations for each task (RGGC):

# task 1
python main.py --net "RGGC" --task_id 1 --lr 0.05 --state_dim 10 --n_steps 5 --niter 100 --cuda
# task 2
python main.py --net "RGGC" --task_id 2 --lr 0.05 --state_dim 8 --n_steps 10 --niter 100 --cuda
# task 4
python main.py --net "RGGC" --task_id 4 --lr 0.01 --state_dim 8 --n_steps 10 --niter 100 --cuda
# task 9
python main.py --net "RGGC" --task_id 9 --lr 0.005 --state_dim 8 --n_steps 5 --niter 100 --cuda
# task 11
python main.py --net "RGGC" --task_id 11 --lr 0.01 --state_dim 10 --n_steps 10 --niter 100 --cuda
# task 12
python main.py --net "RGGC" --task_id 12 --lr 0.005 --state_dim 4 --n_steps 10 --niter 100 --cuda
# task 13
python main.py --net "RGGC" --task_id 13 --lr 0.005 --state_dim 4 --n_steps 5 --niter 100 --cuda
# task 15
python main.py --net "RGGC" --task_id 15 --lr 0.05 --state_dim 10 --n_steps 10 --niter 100 --cuda
# task 16
python main.py --net "RGGC" --task_id 16 --lr 0.01 --state_dim 4 --n_steps 5 --niter 100 --cuda
# task 17
python main.py --net "RGGC" --task_id 17 --lr 0.01 --state_dim 4 --n_steps 5 --niter 100 --cuda
# task 18
python main.py --net "RGGC" --task_id 18 --lr 0.005 --state_dim 10 --n_steps 10 --niter 100 --cuda

Results

Only 50 randomly selected training examples for training for each task. Performances are evaluated on 50 random examples.

bAbI Task	GGNN Accuracy	RGGC Accuracy
1	46.00%	47.60%
2	36.00%	37.60%
4	100.00%	58.95%
9	8.00%	38.40%
11	31.20%	35.20%
12	28.00%	30.40%
13	27.60%	28.40%
15	100.00%	73.50%
16	100.00%	100.00%
17	29.67%	43.96%
18	33.86%	28.35%

Disclaimer

The data processing codes are from official implementation yujiali/ggnn.