Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Scatter_max overflows with int

See original GitHub issue

🐛 Bug

I have a torch_geometric.data.Data object with the x attribute in torch.int format. (the attribute represents vertex colours, hence the type).

If I attempt to use torch.nn.max_pool, it will yield an error in scatter_max. Note that using torch.nn.avg_pool gives the desired result, as does converting the x attribute to float before trying scatter_max (which is what I resorted to doing)

>>> avg_pool(clusters,data)
Batch(edge_index=[2, 64384], pos=[10758, 3], x=[10758, 3])
>>> max_pool(clusters,data)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/wp01/trainees/aykroydc/tools/miniconda3/lib/python3.7/site-packages/torch_geometric/nn/pool/max_pool.py", line 64, in max_pool
    x = _max_pool_x(cluster, data.x)
  File "/home/wp01/trainees/aykroydc/tools/miniconda3/lib/python3.7/site-packages/torch_geometric/nn/pool/max_pool.py", line 9, in _max_pool_x
    return scatter_('max', x, cluster, dim_size=size)
  File "/home/wp01/trainees/aykroydc/tools/miniconda3/lib/python3.7/site-packages/torch_geometric/utils/scatter.py", line 28, in scatter_
    out = op(src, index, 0, None, dim_size, fill_value)
  File "/home/wp01/trainees/aykroydc/tools/miniconda3/lib/python3.7/site-packages/torch_scatter/max.py", line 98, in scatter_max
    src, out, index, dim = gen(src, index, dim, out, dim_size, fill_value)
  File "/home/wp01/trainees/aykroydc/tools/miniconda3/lib/python3.7/site-packages/torch_scatter/utils/gen.py", line 18, in gen
    out = src.new_full(out_size, fill_value)
RuntimeError: value cannot be converted to type int without overflow: -1e+38

Nothing particular about the data I’m using: >>> data.x.dtype torch.int32 >>> data.x.shape torch.Size([53490, 3]) >>> data.x.max() tensor(239, dtype=torch.int32) >>> data.x.min() tensor(14, dtype=torch.int32) >>> clusters.shape torch.Size([53490])