🐛 Bug: ResNetMSA._init_params sets weight to 0 instead of bias + method never called

[vietanhrs]

Description

src/resnet.py lines 139-145

def _init_params(self):
    for m in self.modules():
        if isinstance(m, nn.Conv2d):
            nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
        elif isinstance(m, nn.BatchNorm2d):
            nn.init.constant_(m.weight, 1)
            nn.init.constant_(m.weight, 0)  # ← Bug: should be m.bias

Two issues:

1. Line 145 sets m.weight to 0, overriding the line above which just set it to 1. Should be nn.init.constant_(m.bias, 0)
2. _init_params is never called anywhere — the initialization is effectively dead code

Why It Matters

If anyone calls _init_params(), all BatchNorm weights become 0, which effectively disables the normalization layers and would cause severe training issues.

Suggested Fix

elif isinstance(m, nn.BatchNorm2d):
    nn.init.constant_(m.weight, 1)
    nn.init.constant_(m.bias, 0)  # Fix: m.bias instead of m.weight

And either call self._init_params() in __init__ or remove the method.