sashimi: LinearActivation initialization gives TypeError for weight_norm argument

[BavoKempen]

Hi,

I bounced upon an issue while attempting to run the sashimi.py script, namely the LinearActivation function in the ./standalone/s4.py script does not accept the weight_norm argument upon initialization passed on through the **kwargs in the DownPool and UpPool classes.

The error trace:

File ".\PycharmProjects\state-spaces\sashimi\sashimi.py", line 450, in <module>
    model = Sashimi(n_layers=2).cuda()

File ".\PycharmProjects\state-spaces\sashimi\sashimi.py", line 298, in __init__
    d_layers.append(DownPool(H, expand, p))

File ".\PycharmProjects\state-spaces\sashimi\sashimi.py", line 28, in __init__
    self.linear = LinearActivation(

File ".\PycharmProjects\state-spaces\src\models\sequence\ss\standalone\s4.py", line 137, in LinearActivation
    linear = linear_cls(d_input, d_output, bias=bias, **kwargs)

TypeError: __init__() got an unexpected keyword argument 'weight_norm'

Looking a bit further I noticed that neither the nn.Conv1d (chosen in DownPool due to transposed = True) nor the nn.Linear that could be called within LinearActivation(), have the explicit weight_norm argument in Pytorch.

Am I overlooking something?

Python: 3.9
Pytorch: 1.12 (latest stable release)

Thanks a lot for publishing your code with the papers!

Cheers,
Bavo

[albertfgu]

Thanks for pointing this out! It looks like I deleted some unused options in the V2.1 cleanup, but didn't track all of the other references. You should be able to just remove those arguments from the sashimi.py code. I am not planning to push small fixes like this yet as V3 is almost done.

[BavoKempen]

Hi @albertfgu,

Thanks a lot for your quick reply, I will adjust the code accordingly.
Looking forward to the V3 release!

Kind regards,
Bavo