Accessing layers and their weights/parameters inside Sequential

Hello,

First, thanks a lot for your work on this framework, it’s a very powerful tool!

I had a concern about your customized nn.Sequential. It does not allowed to access to the layers inside it:

m = Sequential(
  (0): BatchNorm(200)
  (1): GCNConv(200, 10)
  (2): ELU(alpha=1.0, inplace=True)
  (3): <function global_mean_pool at 0x0000026F7CA36AF0>
  (4): BatchNorm(10)
  (5): Dropout(p=0.2, inplace=False)
  (6): Linear(in_features=10, out_features=2, bias=True)
)
m[1]     
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: 'Sequential_dbe74c' object is not subscriptable

It would be useful to be able to access them, for example to check the weights and parameters of the layers.

For example, the initial nn.Sequential allows it, and propose to name our layer like that:

# Using Sequential with OrderedDict. This is functionally the same as the above code
model = nn.Sequential(OrderedDict([
          ('conv1', nn.Conv2d(1,20,5)),
          ('relu1', nn.ReLU()),
          ('conv2', nn.Conv2d(20,64,5)),
          ('relu2', nn.ReLU())
        ]))

(from https://pytorch.org/docs/stable/generated/torch.nn.Sequential.html)

Would it be possible to add a similar feature, to get a direct access to our layers?

Thanks for your attention,

[rusty1s]

Every layer is different, some layers may utilize multiple linear layers, some may use MLPs, etc. It is best to look up the code on a GNN layer in order to see how to access its parameters.

[rusty1s]

Added support for OrderedDict as well, see https://github.com/pyg-team/pytorch_geometric/pull/4075.