Help! What does 'ptr' mean in 'from torch_geometric.loader import DataLoader'

[icesky2310]

Hi, everyone.
I am just learning the Graph Neural Network. So, I import the class using 'from torch_geometric.loader import DataLoader' and I can see there is an attribute in the dict which named 'ptr'.
I have looked around the net and went after some useful blogs or references to know what it is. Does anyone please tell me what does that mean?
train_loader = DataLoader(train_dataset, batch_size=64)

One batch in the train_loader look like this:
DataBatch(x=[154, 1], edge_index=[2, 124], y=[64], batch=[154], ptr=[65])

[rusty1s]

Thanks for asking. I am very sorry that this is not documented at all (mostly since we only make use of it internally). The ptr tensor describes the compressed representation of the batch vector (in a similar fashion to the CSR representation of sparse matrices). Each representation can be converted into the other format. For example, if you batch two graphs together (with 4 and 5 nodes, respectively), then ptr will denote the offsets between these graphs:

[0, 4, 9]

In contrast, the batch vector will look like:

[0, 0, 0, 0, 1, 1, 1, 1, 1]

The ptr vector allows us to make use of segment_reductions, which are a little bit faster than scatter operations since we can leverage the fact that values that point to the same index are grouped and sorted in memory.

[icesky2310]

Hi, Matthias,
Thanks a lot for your kind help and patient answer which is clear and detailed.
I have understood it.
Best wishes,
Jianhai Zhang