It’s probably such an error reporting method. I’ve been using torch for so many years. I first encountered this error NotImplementedError
I’m not using a nightly version

Traceback (most recent call last):

  File "xxxxx\x.py", line 268, in <module>
    print(x(y).shape)

  File "xxxxx\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)

  File "xxxxx\x.py", line 259, in forward
    x = self.features(x)

  File "xxxxx\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)

  File "xxxxx\lib\site-packages\torch\nn\modules\container.py", line 119, in forward
    input = module(input)

  File "xxxxx\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)

  File "xxxxx\lib\site-packages\torch\nn\modules\module.py", line 201, in _forward_unimplemented
    raise NotImplementedError

NotImplementedError

Call self.forward in _call_impl

result = self.forward(*input, **kwargs)

If you inherit nn.Module, and if you don’t implement self.forward, it will

raise NotImplementedError

It turns out that when I use this function, I really don’t have the forward method:

class Hswish(nn.Module):

    def __init__(self, inplace=True):
        super(Hswish, self).__init__()
        self.inplace = inplace

    def __swish(self, x, beta, inplace=True):
        # But this swish is not used by H-swish
        # The reason it's called H-swish is to make the sigmoid hard
        # approximated by Relu6(x+3)/6
        # Reduced computational effort for embedded deployment
        return x * F.sigmoid(beta * x, inplace)

    @staticmethod
    def Hsigmoid(x, inplace=True):
        return F.relu6(x + 3, inplace=inplace)/6

    def foward(self, x):
        return x * self.Hsigmoid(x, self.inplace)

forward Write as foward …