Usage of nn.ModuleList

1

Hi, I wrote a simple script here, utilizing the new nn.ModuleList.

 module = [nn.Conv2d(1,64, 3, stride = (1,1), padding = (1,1)), nn.BatchNorm2d(
 self.module_list = nn.ModuleList(module)
 self.module_list += module

Each “module”, is composed of a Conv2d, and a BatchNorm. What I would like self.module_list to be, is: Conv2d -> BatchNorm -> Conv2d -> BatchNorm

Is this the right way to go about it? I worry that I am not really composing a function, but re-using the same set of weights twice… which I don’t want of course. I want each Conv2d, BN to be unique of course.

Is this the right way to go about it?

Thanks!!

1 Like
2

No, you’re adding the same modules, with the same weights twice. You need to create a new list, and append that after construction.

def make_sequence():
    return [nn.Conv2d(1,64, 3, stride = (1,1), padding = (1,1)), ...]

self.module_list = nn.ModuleList()
for i in range(5):
    self.module_list += make_sequence()
5 Likes