Adding trim_to_layer utility function

[mszarma]

Adding hierarchical graph adjacency matrix feature to significantly improve performance

Contributors: @rBenke @andreazanetti

[rusty1s]

Thanks for this PR. I have a few questions:

• Can we split this PR into multiple such that we first integrate sampled_info into the NeighborLoader before we think about integration in MessagePassing/benchmark scripts?
• I am not entirely sure why we need the concept of a HierarchicalSparseTensor. Can't we just let NeighborLoader return a list of sliced edge indices/sparse tensors?
• I would like to avoid adding any logic of this to MessagePassing and its instances. IMO, any customization of x and edge_index should happen outside of it. Would that be possible?

[mszarma]

Thanks for this PR. I have a few questions:

• Can we split this PR into multiple such that we first integrate sampled_info into the NeighborLoader before we think about integration in MessagePassing/benchmark scripts?
• I am not entirely sure why we need the concept of a HierarchicalSparseTensor. Can't we just let NeighborLoader return a list of sliced edge indices/sparse tensors?
• I would like to avoid adding any logic of this to MessagePassing and its instances. IMO, any customization of x and edge_index should happen outside of it. Would that be possible?

Thanks @rusty1s for the review. Let's proceed first with pyg-lib part.

[codecov]

Codecov Report

Merging #6661 (7b3b6fc) into master (bf96b63) will increase coverage by 0.00%.
The diff coverage is 92.68%.

❗ Current head 7b3b6fc differs from pull request most recent head 3c1dde7. Consider uploading reports for the commit 3c1dde7 to get more accurate results

@@           Coverage Diff           @@
##           master    #6661   +/-   ##
=======================================
  Coverage   91.46%   91.46%           
=======================================
  Files         431      432    +1     
  Lines       23458    23498   +40     
=======================================
+ Hits        21455    21492   +37     
- Misses       2003     2006    +3     

Impacted Files	Coverage Δ
torch_geometric/sampler/base.py	98.02% <ø> (ø)
torch_geometric/nn/models/basic_gnn.py	89.18% <87.50%> (-0.08%)	⬇️
torch_geometric/utils/trim_to_layer.py	92.59% <92.59%> (ø)
torch_geometric/nn/summary.py	100.00% <100.00%> (ø)
torch_geometric/typing.py	100.00% <100.00%> (ø)
torch_geometric/utils/__init__.py	100.00% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more