So the issue is that the subclass __init__ methods have two fewer parameters (no transposed, output_padding) than _ConvNd. The current fix works, but technically the current way the class hierarchy is organized is not correct and the better way to do it would be:

    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
                 padding=0, dilation=1, groups=1, bias=True,
                 padding_mode='zeros'):
        # All subclasses have this signature
        raise NotImplementedError

    def _init(self, in_channels, out_channels, kernel_size, stride,
              padding, dilation,
              transposed, output_padding,
              groups, bias,
              padding_mode='zeros'):
        super(_ConvNd, self).__init__()

and then subclasses should call super(Conv1d, self)._init(

I think that would be a good change, but it is only possible if there are no subclasses of _ConvNd floating around (it's private, but you never know).

Let's see what other reviewers think.

Ok, I'll wait the other reviewers before changing this part of the code

Soooo it is a little tricky. torch.nn._ConvNd is sadly de facto public API; see https://github.com/pytorch/fairseq/blob/master/fairseq/modules/quantization/scalar/modules/qconv.py#L14 for an example. But this is quantized ConvNd, and this code is relatively new and there ought to be less uses. cc @jerryzh168 for more thoughts.

I'm trying to decide if I agree that the alternate version is better.

I guess the high level question I have is why the subclass init methods have fewer parameters. This seems like a blatant oversight to me.

I guess the high level question I have is why the subclass init methods have fewer parameters

Because those two parameters are implicitly defined by the dimensionality of subclass - e.g. Conv1d hardcodes transpose, output_padding to False, (0,). It would not make sense for the user to define those when initializing Conv1d.

transpose is always False, so that really doesn't make much sense - probably code just copied from the non-quantized version or something like that.

OK fair enough. The suggested alternative seems like a reasonable way to shut up the type checker then.

I've updated the code to use _init.

I see, are we going to have similar changes in https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/conv.py as well?