Changed the interface of coalesce to only return single edge_index information in case the edge_attr argument is not specified

[rusty1s]

No description provided.

[codecov]

Codecov Report

Merging #6879 (5d19333) into master (581dd01) will increase coverage by 0.00%.
The diff coverage is 95.23%.

❗ Current head 5d19333 differs from pull request most recent head b857e10. Consider uploading reports for the commit b857e10 to get more accurate results

@@           Coverage Diff           @@
##           master    #6879   +/-   ##
=======================================
  Coverage   91.53%   91.53%           
=======================================
  Files         427      427           
  Lines       23173    23166    -7     
=======================================
- Hits        21211    21205    -6     
+ Misses       1962     1961    -1     

Impacted Files	Coverage Δ
torch_geometric/utils/undirected.py	90.24% <66.66%> (-0.67%)	⬇️
torch_geometric/io/tu.py	90.58% <100.00%> (-0.22%)	⬇️
torch_geometric/transforms/line_graph.py	100.00% <100.00%> (ø)
torch_geometric/transforms/two_hop.py	100.00% <100.00%> (ø)
torch_geometric/utils/coalesce.py	100.00% <100.00%> (+2.32%)	⬆️
torch_geometric/utils/grid.py	100.00% <100.00%> (ø)
torch_geometric/utils/sort_edge_index.py	100.00% <100.00%> (ø)

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

[EdisonLeeeee]

@rusty1s thank you for the update. I've been struggling with this discrepancy between different functions, such as coalesce and add_self_loops.

However, this is indeed a breaking change in PyG. I'm concerned that there may be compatibility issues for users who upgrade from older versions of PyG. Most importantly, the returned edge_index is unpackable, which means that the same code may cause potential errors when edge_attr is None. For example:

# for PyG<=2.2
edge_index, _ = coalesce(...) # correct

# for PyG >=2.3
edge_index, _ = coalesce(...) # wrong but work properly, equivalent to row, col = edge_index

The worst-case scenario is that code written with an older version will still work in the latest version, but return incorrect results. Should we take steps to avoid this situation?

[rusty1s]

Yes, this is a breaking change and will be highlighted in the release notes. I am not totally convinced it will impact a lot of users, as sort_edge_index and coalesce are more used internally, and it would only cause problems if you would call these functions via edge_index = coalesce(edge_index, None) (which seems rare). If you have any idea on how to make this backward compatible, I would be happy to adapt it.

[EdisonLeeeee]

Thank you for your clarification. You are right. But I think we can add a warning in the doc-string, WDYT?

[rusty1s]

Yes, sg. Will take care of it.

[rusty1s]

Done in #6893.

[EdisonLeeeee]

Thank you!