EEGSym spatial convolutions are dense (groups=1, bias=True) instead of depthwise/grouped no-bias #1071

The spatial convolutions in EEGSym (braindecode/models/eegsym.py, _InceptionBlock.spatial_convs and _ResidualBlock.spatial_convs) are dense nn.Conv3d with default bias=True:

nn.Conv3d(
    in_channels=filters_per_branch * len(scales_samples),
    out_channels=filters_per_branch * len(scales_samples),
    kernel_size=(1, 1, ncha),   # spatial conv
    padding=(0, 0, 0),
)   # groups=1 (dense), bias=True

The authors' EEGSym (EEGSym_architecture.py, unit_dconv) uses a depthwise / grouped spatial convolution with no bias — the paper's explicit "grouped convolutions to emulate depthwise" design (§II.D):

Conv3D(kernel_size=(1,1,ncha), filters=filters*len(scales),
       groups=filters*len(scales), use_bias=False)

So braindecode mixes across all filters (dense) instead of per-filter spatial filtering, changes the parameter count, and adds a bias the source omits.

Fix

Build the spatial convs with groups=out_channels (depthwise) and bias=False, matching unit_dconv. (Minor related note: the braindecode default spatial_resnet_repetitions=5 differs from the source default 1 — "set to 1 and not tested".) Happy to PR.