Confuse about the num_hidden parameter in GATConv

Questions
#1

In the API documentation ：
https://docs.dgl.ai/api/python/nn.pytorch.html?highlight=gatconv#dgl.nn.pytorch.conv.GATConv
the API GATConv doesn’t contain a parameter called num-hidden

however, in the GAT example of pytorch, there is a num-hidden
%E5%BE%AE%E4%BF%A1%E5%9B%BE%E7%89%87_20191014105642

Besides, the argparse part in train.py use ‘num-hidden’ rather than ‘num_hidden’, but the code still works, why was that?
parse

I’m confused about the meaning of these parameter , whether is it needed and are ‘num-hidden’ and ‘num_hidden’ the same?

#2
  1. For the num_hidden in GATConv, since we did not explicitly specify the argument name, this will be just the second parameter for GATConv, i.e. out_feats.
  2. With argparse, any argument of name --A-B-...-C will become an attribute with name A_B_..._C.
1 Like
#3

thanks
so does num-hidden mean the dimention of the representation of the nodes in the hidden layers?

#4

Here they are in fact the output size of a GATConv layer. Since the model consists of multiple layers, they are also the hidden size of the whole model.

#5

sorry to reply late, But I’m still confused about the “hidden units” in GNN models , since In GNN models, it seems that there is no hidden units like that in CNN models. So what’s the hidden units on earth?

#6

Let’s say your input node features X is a matrix of shape (N, M_1), where N is the number of nodes, and M_1 is the size of node feature. With each GNN layer, we also project the node features with

H=XW+b

such that H is a matrix of shape (N, M_2). The num_hidden denotes stuff like M_2.

#7

So as I mentioned before,it’s the dimention of the hidden representation ,right?
But In the example of cluster_gcn，run_ppi.sh，the num_hidden parameter was 2048, while the dimention of the raw feature of ppi_dataset was 50, I do not quite understand why to transform a 50-dimensional vector into a 2048-dimensional vector?

#8

We want our model to have sufficient capacity for representation learning. The larger the hidden dimension is, the more parameters we can use and the more powerful our model is. May I ask whether you had background in deep learning before?

#9

thanks for replying ，I took some online courses and read some gnn papers before.Is my question too stupid?