Using my own dataset ?

[ahariri13]

Hi ! I am using my own dataset named PHY on DGCNN. I have shaped my data as mentioned in the readme file.
However, my data is big enough (300 000) and I prefer to proceed without 10-fold cross-validation (training on 230 000-testing on 70 000). For this, I ran ./run_DGCNN.sh Phy 1 70000. Despite the fact that it's training and the loss is decreasing, I am getting a 0 test accuracy and the following message:
_UndefinedMetricWarning: No positive samples in y_true, true positive value should be meaningless._I googled the warning and apparently it thinks all my labels are 0 (while I should be having 4 class labels). What do you think the problem is ?

Also, could you point out where I can find how much Dropout is used after a Dense layer, and where I can check the number of Fully-Connected layers used? (not the #of nodes just the # of hidden layers).

[muhanzhang]

Hi, along with accuracy, DGCNN also tries to output AUC results (which only makes sense for binary classification problems). The warning might be related to AUC. Also, can you post a screenshot of your output results and error messages? Without them it is hard to diagnose the problem. Thanks.

For the second question, please refer to line.

[ahariri13]

Thank you for your reply. Apparently the problem was from an error in Graph labeling. However i still have a slight problem: I have 4 nodes each having 2 features and 3 edges connecting them having 1 feature each. But how do I account for edge features in the text file? I even ran the matlab code with the edge attributes and labels in it but I only saw the 2 node features accounted for in the text file.

Any help is appreciated :)

[muhanzhang]

Hi, unfortunately this implementation does not support edge features. Using edge features requires extra modules during the message passing.