snap-stanford/neural-subgraph-learning-GNN

Augment dim error

WxxShirley opened this issue · 4 comments

WxxShirley commented

Hi,
thanks for your work on neural subgraph match. It runs quite well on my dataset.

However, I wanted to fix the augmentation step in ./common/feature_preprocess.py . In detail, add features such as degree, clustering_coefficient, etc. How to correctly implement the FEATURE_AUGMENT and FEATURE_AUGMENT_DIMS?

Auctally, I set FEATURE_AUGMENT as ['degree','betweenness_centrality'] and set FEATURE_AUGMENT_DIMS as [1,1], it encounters error ......

qema commented

So sorry for the late response. Great to hear the method works on your dataset!

The FEATURE_AUGMENT and FEATURE_AUGMENT_DIMS look good to me (although to be sure, degree should instead be node_degree? Also, it may be necessary to use higher values for the degree feature, as the degree is encoded via a one-hot encoding).

Additionally, to add new features, it is necessary to register the features: see the line at

neural-subgraph-learning-GNN/common/feature_preprocess.py

Line 136 in d6ad4da

self.node_feature_funs = {"node_degree": degree_fun,
. The functions above that line give examples for other features, where each function actually creates the feature by creating a new property graph.[feature_name] which is a tensor of features, one row per node in the graph.

WxxShirley commented

oh, that's just a spelling error😅

If I set:

FEATURE_AUGMENT = ["node_degree", "node_clustering_coefficient"]
FEATURE_AUGMENT_DIMS = [1, 1]

there will be error like this:

Training:   0%|          | 0/100 [00:00<?, ?it/s]Process Process-3:
Process Process-1:
Process Process-2:
Traceback (most recent call last):
Traceback (most recent call last):
·······
  File "/Users/xxwu/Downloads/subgraph/common/models.py", line 170, in forward
    x = self.pre_mp(x)
  File "/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/container.py", line 117, in forward
    input = module(input)
  File "/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/linear.py", line 93, in forward
    return F.linear(input, self.weight, self.bias)
  File "/opt/anaconda3/lib/python3.8/site-packages/torch/nn/functional.py", line 1690, in linear
    ret = torch.addmm(bias, input, weight.t())
RuntimeError: mat1 and mat2 shapes cannot be multiplied (527x2 and 3x64)
qema commented
WxxShirley commented
  • does it work if you set the corresponding dimension to a higher value: does not work😅
  • Does it work with just one feature but not the other : only use "node_clustering_coefficient" it works
  • Also how are you creating the data (which data source class are you using)?: build the data source with my own dataset..