FilterBankLayer IIR path NaNs in float32 (coefficients downcast before filtfilt) #1067

Bug

FilterBankLayer._apply_iir (braindecode/modules/filter.py) downcasts the float64 filter coefficients to the input dtype before torchaudio.functional.filtfilt:

filtered = filtfilt(
    x,
    a_coeffs=a_coeffs.type_as(x).to(x.device),   # float64 coeffs -> float32
    b_coeffs=b_coeffs.type_as(x).to(x.device),
    clamp=False,
)

For an IIR band-pass with poles near the unit circle (e.g. a 4th-order Chebyshev-II bank, as used by FBMSNet/FBCNet), the recursion diverges in float32 and the whole forward becomes NaN from the first batch (train loss NaN, accuracy at chance). Verified: NaN in float32, stable in float64.

Reproduce

import torch
from braindecode.modules.filter import FilterBankLayer
fb = FilterBankLayer(n_chans=22, sfreq=250, method="iir",
                     iir_params=dict(order=4, ftype="cheby2", rs=30, output="ba"))
out = fb(torch.randn(4, 22, 1000))   # float32
print(torch.isnan(out).any())        # tensor(True)

Fix

Run the recursion in float64 and cast back (the coefficient buffers are already float64):

orig_dtype = x.dtype
filtered = filtfilt(x.double(),
                    a_coeffs=a_coeffs.double().to(x.device),
                    b_coeffs=b_coeffs.double().to(x.device), clamp=False)
return filtered.to(orig_dtype).unsqueeze(1)

This makes any IIR filter bank (cheby2 etc.) usable. (Separately, the IIR path is restricted to 'ba' coefficients — SOS would be more robust for higher orders, but the float64 cast already fixes the NaN.) Happy to PR.