Bug in PNAConv implementation

[adrianjav]

🐛 Describe the bug

When aggregating messages in PNAConv, scalers are applied one after the other and then stacked, but the variable used on the for loop is the same one always. That is, scalers are applied to the same variable one after the other, instead of applying each scaler independently to the same base variable.

This is the latest code in the repository:

pytorch_geometric/torch_geometric/nn/aggr/scaler.py

Lines 68 to 90 in 8bcc77c

	out = self.aggr(x, index, ptr, dim_size, dim)
	assert index is not None
	deg = degree(index, num_nodes=dim_size, dtype=out.dtype).clamp_(1)
	size = [1] * len(out.size())
	size[dim] = -1
	deg = deg.view(size)
	outs = []
	for scaler in self.scaler:
	if scaler == 'identity':
	pass
	elif scaler == 'amplification':
	out = out * (torch.log(deg + 1) / self.avg_deg['log'])
	elif scaler == 'attenuation':
	out = out * (self.avg_deg['log'] / torch.log(deg + 1))
	elif scaler == 'linear':
	out = out * (deg / self.avg_deg['lin'])
	elif scaler == 'inverse_linear':
	out = out * (self.avg_deg['lin'] / deg)
	else:
	raise ValueError(f"Unknown scaler '{scaler}'")
	outs.append(out)

and the variable out is reused over the entire loop.

Proposed solution

The solution is as simple as using a different variable name within the for loop and do, e.g., out_i = out at the start of the loop in line 77. Since it is such a simple change, I am happy to open a pull request if requested.

I might be wrong here, so please correct me if that's the case, but looking at the original paper and the official implementation, I think this is a bug:
https://github.com/lukecavabarrett/pna/blob/0c630c2e2d88bb1ef784c850dd8f3a069fcd9489/models/pytorch_geometric/pna.py#L158

Environment

• PyG version:
• PyTorch version:
• OS:
• Python version:
• CUDA/cuDNN version:
• How you installed PyTorch and PyG (conda, pip, source):
• Any other relevant information (e.g., version of torch-scatter):

[EdisonLeeeee]

Thanks for reporting. You are right. This is not expected behavior in PNAConv. Please feel free to send a PR for this :)