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.

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!

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

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!

Error Message

Error message when running main.py:

Traceback (most recent call last):
  File "main.py", line 72, in <module>
    main(opt)
  File "main.py", line 68, in main
    test(test_dataloader, net, criterion, optimizer, opt)
  File ".../ggnn.pytorch/utils/test.py", line 24, in test
    test_loss += criterion(output, target).data[0]
IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number

Steps to reproduce:

1. Create a new virtualenv with python2.7.
2. pip install torch numpy
3. python main.py

Suggested Reason

I believe the bug occured due to a depreciating of indexing 0-dim tensors in an update to PyTorch.

I have implemented the fix.