chingyaoc / ggnn.pytorch Goto Github PK

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!

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

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

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!!!

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!

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.

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