Scatter operations such as scatter_max vs to_dense_batch

[jomalla123]

Hi,

I have to do multiple operations over a batch of graphs with different sizes, e.g. calculate the max.

After seeing #4488, I implemented it like this:

dense_batch, mask = pyg.utils.to_dense_batch(input, batch_index, fill_value=float("-inf"))
max_values, argmax_indices = torch.max(dense_batch, dim=dim)

Is this the right approach? There is also torch_scatter.scatter_max which would achieve the same thing but with converting to a dense batch. Would this be more efficient in general?

[rusty1s]

Yes, I recommend to use scatter_max for this use-case.