Hi,

I believe the parameter max_n_modes is not used in class SphericalConv(BaseSpectralConv) layer which is the conv_module for the SFNO model, whereas its usage has been implemented in the SpectralConv class, the default conv_module for FNO.

Weight shape in SphericalConv.__init__() is defined as:

weight_shape = (in_channels, out_channels, *self.n_modes[:-1])

But in SpectralConv.__init__() we have:

weight_shape = (in_channels, out_channels, *max_n_modes)

There is also handling the different n_modes VS. max_n_modes in SpectralConv.forward():

# if current modes are less than max, start indexing modes closer to the center of the weight tensor
starts = [(max_modes - min(size, n_mode)) for (size, n_mode, max_modes) in zip(fft_size, self.n_modes, self.max_n_modes)]

which needs to be addressed accordingly in SphericalConv.forward() knowing that sht is happenning instead of fft. Maybe this is just a "todo" but I wanted to point it out :)

Best,
Reza