giannisnik / cnn-graph-classification Goto Github PK

Hi @giannisnik ,

Could you please shed some light on why to append a zero row on top of Q_t?
Q_t = np.vstack([np.zeros(embedding_dim), Q_t]) # Line 154 in utils.py

Thanks.

Hi giannisnik,

Thanks for your kind sharing code!

As I run "python main.py" without any code change, I get a RecursionError. The whole trace is posted below, could you pls help check?

$ python main.py
Computing graph maps
Number of communities: 2590
Average size: 7.53
Starting cross-validation
Create model
Traceback (most recent call last):
File "main.py", line 280, in
main()
File "main.py", line 101, in main
model = CNN(input_size=args.n_filters, hidden_size=args.hidden_size, n_classes=np.unique(y).size, d=args.d, n_kernels=n_kernels, max_n_communities=max_n_communities)
File "/Users/polarbear/Personal/NU/MSiA/Courses/Summer_Quarter_2018/OnPoint_Internship/cnn-graph-classification-master/kcnn/model.py", line 13, in init
self.init_weights()
File "/Users/polarbear/Personal/NU/MSiA/Courses/Summer_Quarter_2018/OnPoint_Internship/cnn-graph-classification-master/kcnn/model.py", line 16, in init_weights
xavier_uniform(self.conv.weight.data)
File "/Users/polarbear/Personal/NU/MSiA/Courses/Summer_Quarter_2018/OnPoint_Internship/cnn-graph-classification-master/kcnn/init.py", line 43, in xavier_uniform
xavier_uniform(tensor.data, gain=gain)
File "/Users/polarbear/Personal/NU/MSiA/Courses/Summer_Quarter_2018/OnPoint_Internship/cnn-graph-classification-master/kcnn/init.py", line 43, in xavier_uniform
xavier_uniform(tensor.data, gain=gain)
File "/Users/polarbear/Personal/NU/MSiA/Courses/Summer_Quarter_2018/OnPoint_Internship/cnn-graph-classification-master/kcnn/init.py", line 43, in xavier_uniform
xavier_uniform(tensor.data, gain=gain)
[Previous line repeated 985 more times]
File "/Users/polarbear/Personal/NU/MSiA/Courses/Summer_Quarter_2018/OnPoint_Internship/cnn-graph-classification-master/kcnn/init.py", line 42, in xavier_uniform
if isinstance(tensor, Variable):
RecursionError: maximum recursion depth exceeded

Thanks!

Hi @giannisnik,

Seems the function orthogonal is not used in your code.
What is it for? Is it to accelerate the model training?

Br,

Hi there. Thank you for the code. In which line in the main.py file can I specify the dataset and the hyperparameters? When I run python main.py it gives me the error below:

Computing graph maps
Number of communities: 2594
Average size: 7.52
Traceback (most recent call last):
File "main.py", line 280, in
main()
File "main.py", line 62, in main
Q, subgraphs, labels, shapes = compute_nystrom(args.dataset, use_node_labels, args.d, args.community_detection, kernels)
File "/home/n/Documents/GDL/cnn-graph-classification/kcnn/utils.py", line 152, in compute_nystrom
model.fit(communities)
File "/home/n/Documents/GDL/cnn-graph-classification/kcnn/nystrom.py", line 30, in fit
basis_kernel = self.kernel(basis, basis, **self._get_kernel_params())
File "/home/n/Documents/GDL/cnn-graph-classification/kcnn/graph_kernels.py", line 111, in wl_kernel
if len(neighbors) > 0:
TypeError: object of type 'generator' has no len()

Thank you so much @giannisnik for publishing this code. Is a very interesting use of a CNN. I'm trying to reproduce some of your results but keep getting Accuracy = 0 measurements. Any thoughts on where to focus my debugging? Maybe related to the missing "best model"?

Hi giannisnik,
Thanks for your kind sharing code!
Can you give detailed code comments to help me learn?

And could you pls help explain what is the meaning below?
Epoch: [99] Avg accuracy 0.762; Average Loss 0.451; Avg Time x Batch 0.007

• Average accuracy 0.700; Average Loss 0.727
  => loading best model './checkpoint/kcnn/model_best.pth'
  => loaded best model './checkpoint/kcnn/model_best.pth' (epoch 4)
• Average accuracy 0.660; Average Loss 0.593
  Accuracy at iteration 10: 0.66
  Average accuracy: 0.6759999999999999
  std: 0.041279534881100574

@giannisnik ,

Thanks for sharing the code with us!
I am wondering how to handle categorical edges. For example, there are two types of edges in graphs among nodes, how coud I represent/identify type of a edge in graphs to feed graph kernel? Thanks!

Whenever I try to input a graph to the KCNN which only represents one graph (i.e. N = 1), then I get the error "ValueError: all the input array dimensions except for the concatenation axis must match exactly". What hyperparameter in main.py would I have to adjust to fix this?