do the global pooling operators no longer work with a null batch?

[davidireland3]

I used to use the code

summary_1_mean = global_mean_pool(x, batch)
summary_1_max = global_max_pool(x, batch)

which would work if the batch was None, but in the latest version of PyG this gives me an error as a tensor is expected for the dimensions. I can hard code my way around this, but it was nice to not have to do this in the older version.

The use case is that my network is designed for obtaining graph level embedding and so the batch was needed when training, but at inference I was often interested in obtaining an embedding of a single graph at a time so rather than set up a DataLoader object of just one graph I could make a Data object and set _DataObj_.batch=None and it would work nicely.

Is there a reason this has been removed from the latest version?

[rusty1s]

Thanks for reporting. I'm not sure why and when it was removed TBH. Nonetheless, I agree this is an important feature to have, so I added it back in via #4161.