Wrong number of parameters in ogbn-arxiv GCN submission?

[zjost]

The GCN submission for ogbn-arxiv reports 142,888 parameters on the leader-board. Looking at the code and default settings, it seems to use the following architecture:

• GCNConv(128, 256) (parms = 32,768 (W) + 256 (b) = 33,024)
• BatchNorm(256) (parms = 256*2 = 512)
• GCNConv(256, 256) (parms = 65,536 (W) + 256 (b) = 65,792)
• BatchNorm(256) (parms = 256*2 = 512)
• GCNConv(256, 40) (parms = 10,240 (W) + 40 (b) = 10,280)

This architecture gives a total number of parms of 110,120, which is 32,768 fewer than that reported on the leader-board. It's probably not a coincidence that this number is equal to the input*hidden dimensions. Is there some other projection layer coming into play, is the leader-board mistaken, or something else?

[weihua916]

Hi! You are right, that was a mistake. When I run sum(p.numel() for p in model.parameters()) I indeed got 110,120. I have updated the parameter count.