Create a Linear(in_features, out_features) and re-initialise its weights with Kaiming (He) normal initialisation configured for ReLU -- fan-in mode, ReLU nonlinearity -- then zero the bias.

Seed with torch.manual_seed(seed) immediately before the Kaiming call. Return (weight, bias).

Kaiming differs from Xavier by assuming a ReLU throws away half the signal, so it scales variance up to compensate.

Input

in_features = 4
out_features = 3
seed = 0

Output

(tensor([[ 1.0896, -0.2075, -1.5406,  0.4019],
        [-0.7669, -0.9890,  0.2852,  0.5926],
        [-0.5086, -0.2852, -0.4219,  0.1287]]), tensor([0., 0., 0.]))