This implementation of a kind of Dense block is right or wrong?

vision
1

I recently referred to a kind of dense block implementation (not in a classic way), but what puzzles me is his use of nn.ModuleList. His implementation is shown below.

I think, just the first nn.Sequential in the layers (nn.ModuleList) is used, is right？Look forward replying to help me understand it. Thanks!

%E5%BE%AE%E4%BF%A1%E6%88%AA%E5%9B%BE_20191105142734
%E5%BE%AE%E4%BF%A1%E6%88%AA%E5%9B%BE_20191105142734672×522 14.1 KB

link is here: DeformingAutoencoders-pytorch/DAENet.py at master · zhixinshu/DeformingAutoencoders-pytorch · GitHub

2

I’m not familiar with the code base and am not sure, how the model is supposed to work, but you could use forward hooks to check, which layer is being called:

model = DenseBlockEncoder(1, 3)

def print_layer(name):
    def hook(model, input, output):
        print("{} called".format(name))
    return hook    
    
for idx in range(len(model.layers)):
    model.layers[idx].register_forward_hook(print_layer(str(idx)))

x = torch.randn(1, 1, 4, 4)
model(x)
> 0 called

In this example, only the first sequential block is being called.

3

Thank you so much! That means what I thought is right.