EEGITNet: missing source max-norm constraints + doubled dimensionality-reduction width #1068

braindecode/models/eegitnet.py diverges from the authors' EEG-ITNet (Salami et al. 2022, Main.ipynb) in two ways that materially hurt within-subject accuracy (~11pp on BCI-IV-2a in a replicability study; the model overfits without them):

1. Missing max_norm weight constraints

The source constrains the regularizing weights; braindecode applies none (grep -c max_norm eegitnet.py == 0):

• the 3 inception spatial depthwise convs (Spatial_filter_1/2/3) use depthwise_constraint = MaxNorm(max_value=1);
• the final dense uses kernel_constraint = max_norm(0.25).

braindecode already ships the primitives (Conv2dWithConstraint, MaxNormLinear, MaxNormParametrize) and uses them in EEGNetv4 — they were just omitted here. The dilated TC blocks correctly have no constraint in the source, so leave those unconstrained.

2. Dimensionality-reduction width doubled (28 vs 14)

The source DR conv has 14 filters (paper III-C: "the number of filters in the 1×1 convolutional layer … was selected to be 14"; source Conv2D(14,(1,1))). braindecode doubles it (eegitnet.py:188):

nn.Conv2d(tcn_in_channel, tcn_in_channel * 2, kernel_size=(1, 1)),   # 14 -> 28
nn.BatchNorm2d(tcn_in_channel * 2),

This inflates the classifier input (644 vs 322 features) and adds capacity.

Fix

Apply MaxNormParametrize(1.0) to the 3 inception spatial depthwise convs and MaxNormLinear(max_norm=0.25) (or parametrize) to the final layer; set the DR conv width to tcn_in_channel (14). Both are verified against the authors' code. Happy to PR.