imbalanced_sampler

[zuluokonkwo]

Hi,
I'm presently working on an imbalanced graph data. I tried using the torch geometric imbalanced sampler class to address this issue, but it always gives an error.

Scenario 1: When I define a sampler with the whole dataset as in "sampler=ImbalancedSampler(dataset)" I get an out of index error when training with training data.
Scenario 2: When I define a sampler with the training dataset as in "sampler=ImbalancedSampler(train_dataset)" I get an assertion error. I understand what's going on in scenario 2 "assert len(dataset) == y.numel()".

Is there a way around this?

[rusty1s]

Do you have a minimal example to reproduce? This would help to track down the issue.

[zuluokonkwo]

After installing from master the same error still pops up. Still can't perform sampling on the sliced dataset.

[rusty1s]

Are you sure the master is correctly picked up? You can try to re-install torch-geometric first (multiple times), and then install again. The following snippet works for me in master:

dataset = TUDataset('/tmp/ENZYMES', name='ENZYMES')
dataset = dataset[:50]

sampler = ImbalancedSampler(dataset)
loader = DataLoader(dataset, batch_size=10, sampler=sampler)

for batch in loader:
    print(batch)

[zuluokonkwo]

Can you please change the name of the sliced data to something else and see if it works?

dataset = TUDataset('/tmp/ENZYMES', name='ENZYMES')
sliced_dataset = dataset[:50]

sampler = ImbalancedSampler(sliced_dataset )
loader = DataLoader(sliced_dataset , batch_size=10, sampler=sampler)

for batch in loader:
    print(batch

[rusty1s]

Still works.

[zuluokonkwo]

Its all good now,
Many thanks for your help Matthias.

[nasrinsaalehi]

Converting y Tensor to torch long will probably solve the assertion error.

data = Data(x=self.drug_encodings, y=y.to(torch.long), edge_index=self.edge_list)
# Sample the graph according to the ImbalancedSampler based on y distribution
sampler = ImbalancedSampler(dataset= data.y)