chingyaoc / ggnn.pytorch Goto Github PK

A PyTorch Implementation of Gated Graph Sequence Neural Networks (GGNN)

License: MIT License

Shell 7.47% Python 92.53%

ggnn.pytorch's Introduction

A PyTorch Implementation of GGNN

This is a PyTorch implementation of the Gated Graph Sequence Neural Networks (GGNN) as described in the paper Gated Graph Sequence Neural Networks by Y. Li, D. Tarlow, M. Brockschmidt, and R. Zemel. This implementation gets 100% accuracy on node-selection bAbI task 4, 15, and 16. Their official implementation are available in the yujiali/ggnn repo on GitHub.

What is GGNN?

Solve graph-structured data and problems
A gated propagation model to compute node representations
Unroll recurrence for a fixed number of steps and use backpropogation through time
An output model to make predictions on nodes

Requirements

python==2.7
PyTorch>=0.2

Run

Train and test the GGNN:

python main.py --cuda (use GPUs or not)

Suggesting configurations for each task:

# task 4
python main.py --task_id 4 --state_dim 4 --niter 10
# task 15
python main.py --task_id 15 --state_dim 5 --niter 10
# task 16
python main.py --task_id 16 --state_dim 10 --niter 150

Results

I followed the paper, randomly picking only 50 training examples for training. Performances are evaluated on 50 random validation examples.

bAbI Task	Performance
4	100%
15	100%
16	100%

Here's an example of bAbI deduction task (task 15)

Disclaimer

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

TODO

GraphLevel Output

References

Gated Graph Sequence Neural Networks, ICLR 2016
yujiali/ggnn

ggnn.pytorch's People

Contributors

Stargazers

Watchers

Forkers

shubhampachori12110095 hbcbh1999 johndpope kazuhiradz tangyoubao reeshark hyzcn samangel93 psavine42 paul841029 songfgh christinaliang hvdthong xxarbiter hsientzucheng skkuggi zy20091082 saeidall ricardozitseng bdqnghi nmuralid1 raihan2108 zhengxuyu cocoaaa carolzxyzxy kaihuatang lezhang0912 tsingzao science4fun ranganthaya bkwapong lichao88 axinletter tungk longjinxiang janeshenyy xuyou314 taewook-ko ygan queenie88 lpsunny juexinwang yueyuep chengli0327 jxh4945777 david-lee-1990 zhhhzhang rohanchandra30 fuzhiyi qq826346094 ammieqi balcilar rainmaker22 zyynnn joaorico gabeochieng wuyff zeroqiaoba admdev8 noobanti flash-barry greitzmann njuhxc liang813 cqm1998 caomaowsh trendingtechnology world4jason nnguyen19 dot23 nashid shi-d qsingularityai mint-yuan llxx1998 fjgao jzxycsjzy lucky-nolby li-lindong zpc2024 layingdownboy

ggnn.pytorch's Issues

solved

how to save or extract the GraphLevel Output model

i wanna to know how to save or extract the GraphLevel Output model and use the feature to apply on classification application,thanks

Question about Graph-Level output

In your todo list, you list out the task : GraphLevel Output, then I assume that you haven't implemented this task in the current code set.

But so far when I read the source code, I can see that you nearly finish it.
This line: https://github.com/JamesChuanggg/ggnn.pytorch/blob/0c7897fe9b05e9b4f9a963ff55bd3ad917ea734e/model.py#L123 is to compute the vector representation of the graph that will use to predict the target class and compute the CrossEntropy loss in the latter step. In the current Babi tasks, the prediction target is the label of the node (in most of the task). But for the graph-level output, e.g graph classification, I guess what we need to do is instead of predicting the label of the node, we predict the label of the whole graph, and with the current set of code, we have 99% of the code ready, no need to do more. Not sure If I understand this correctly, please help me to clarify.

I don't know how process the original bAbI data to the form of proceed_1,and symbolic_preprocess.py is not process the original bAbI data?

Creating adjacency matrix

For the part to create the adjacency matrix, it's not clear to me.

For example, we have 2 edges:
2 1 3
1 1 2

Then from the code implementation here:
https://github.com/JamesChuanggg/ggnn.pytorch/blob/master/utils/data/dataset.py#L79

Supposed there are 4 node types and 2 edges types, then Row = 4 , Column = 4 X 2 X 2 = 16
Then we get the 4 X 16 Matrix M :
[[0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0.]
[1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0.]
[0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]

Where:
M[2][1] = 1
M[1][10] = 1
M[1][0] = 1
M[0][9] = 1

What kind of adjacency matrix like this? I've checked the paper, it doesn't mention anything similar to this. I suspect that this is the matrix from Figure 1, section 3.2 of the paper that illustrates how nodes in the graph communicate with each other, in this case, it should be a 4 X 8 Matrix, not 4 X 16.

Please help me to enlighten my understanding.
Thanks!!!

What is the meaning of annotation? Have trouble to figure it out..

https://github.com/JamesChuanggg/ggnn.pytorch/blob/master/utils/data/dataset.py#L74-L75

annotation = np.zeros([n_nodes, n_annotation_dim])
annotation[target[1]-1][0] = 1

@JamesChuanggg Thank you!

Poor performance on Task19

According to the paper, this should achieve 97% acc on task19. which I can only get around 30%.
Did you test on the task19?

have trouble to figure out what is task_type?

https://github.com/JamesChuanggg/ggnn.pytorch/blob/master/utils/data/dataset.py#L72

    for item in data_list:
        edge_list = item[0]
        target_list = item[1]
        for target in target_list:
            task_type = target[0]

Thank you very much!!
@JamesChuanggg

Removing these code and then it become origin GNN?

https://github.com/JamesChuanggg/ggnn.pytorch/blob/master/model.py#L29-L40

        self.reset_gate = nn.Sequential(
            nn.Linear(state_dim*3, state_dim),
            nn.Sigmoid()
        )
        self.update_gate = nn.Sequential(
            nn.Linear(state_dim*3, state_dim),
            nn.Sigmoid()
        )
        self.tansform = nn.Sequential(
            nn.Linear(state_dim*3, state_dim),
            nn.Tanh()
        )

@JamesChuanggg Thank you very much!