MaxUnpool2d with indices from MaxPool2d, all in nn.Sequential

1

Hey all, I saw it recommended to reserve Sequential for only the most trivial networks, but I’m a big fan of the readability/simplicity and I’m trying to keep as much of my network in nn.Sequential as possible.

Is there a simple way to use MaxUnpool2d layers in one Sequential block with respect to the indices from MaxPool2d layers in a previous block?

1 Like
2

there isn’t at the moment

3

Thanks. Are there plans to add more of that kind of support for working in Sequential?

4

no, we dont plan to make Sequential work on complex networks, it was provided as a one-off convenience container for really simple networks

5

Fair enough, thanks.

6 
def forward(self, x):
        for layers in self.front_process:
            print(layers)
            if isinstance(layers, nn.MaxPool2d):
                print('\ngot target1\n')
            
        print('\n\nmiddle \n\n')
                      
        for layers in self.back_process:
            print(layers)
            if isinstance(layers, nn.MaxUnpool2d):
                print('\ngot target2\n')

You can try this to extract maxpool and maxunpool, but I don’t know whether it will influence the speed or not, since I’m quite new to pytorch.