Out of Memory in HeteroGraph Explanation

[Mellandd]

Hello,
I am trying to use the interpretability package with a model created for a heterogeneous graph. The problem comes in that when trying to generate the interpretation there is not enough memory.

It is a model with 6 layers and hidden channel of 256, based on GraphSAGE for nodes of 2 categories. The dataset is:

`HeteroData(
a={ x=[16810, 768] },

b={ x=[4619, 53] },

(a, related_to, b)={ edge_index=[2, 721540] },

(a, is_a, a)={ edge_index=[2, 42142] },

(b, rev_related_to, a)={ edge_index=[2, 721540] }
)`

The GPU has 8 GB of memory. Making the model smaller right now is not an option. Is there anything I can do to be able to use the interpretability package?

• I am using CaptumExplainer because as far as I have seen it is the only one that works with heterogeneous graphs. Are there any others or are there any plans to have compatibility soon?
• According to what I have seen, CaptumExplainer has not implemented the use of batches yet. Is there any news about this?
• I have 2 GPUs in the computer. Is it possible to use both in order not to have memory problems in this task?

Thank you all!

[wsad1]

You could save memory by only explaining a subgraph.
For example lets say you want to explain nodes [1,10]. You could pass this to NeighborLoader and explain only the subgraphs returned by NeighborLoader.

[rusty1s]

Yes, it seems that CaptumExplainer has a higher memory footprint than other explainers, so there is not much we can do about this right now besides reducing the input data to a subgraph.