Restriction on reused modules might be too strict about pytorch-opcounter HOT 5 CLOSED

And if you truly want to show your attitude toward module re-use that much, warnings is more recommended as it won't interrupt the program. 😄

By raising a Warning, you're generating error, rather than real warning.

from pytorch-opcounter.

Question, does the example code you provide trigger the warning? Though ReLU is forwarded several times, the registration should only happen once.

from pytorch-opcounter.

@Lyken17 Sorry that I didn't dig deep into your code's logic, see the following NN design instead.

class SomeNet(nn.Module):
    def __init__(self, reduction=4):
        super(SomeNet, self).__init__()
        self.activation = nn.LeakyReLU(negative_slope=0.3, inplace=True)

        self.encoder_feature = nn.Sequential(OrderedDict([
            ("conv3x3_bn", conv3x3_bn(in_channel, 2)),
            ("activation", self.activation),
        ]))
        ...

The problem has nothing to do with module reuse in forward function. THOP will recursively enter each leaf nn.Module and detect if there exist any "re-defined" module.

However, the aforementioned re-defined is legal, since I want to define a globally same activation function and plant it in many sub-modules. And THOP stops me from doing that.

Like I said, even my definition of network is wrong, it is not THOP's job to point it out. And even if THOP would like to point it out, warnings rather than Raise Error should be used.

As a flops counting tools, it should not do extra judgement or security check. A correct flops and params output is the only thing to focus on.

from pytorch-opcounter.

It is a weird usage, I have to say. For your code example, each activation should have its own nn.Module. The proper case should be a recurrent structure where one layer is used in different depth. I have relaxed the check: now it only prints the warning instead of raises it.

from pytorch-opcounter.

Yet a legal one and a convinient one, too. Still unable to get your points of checking user's network... If I were you I would cancel the redundant check, but after all it is your project 😂

from pytorch-opcounter.