Create a Linear(in_features, out_features), then re-initialise it: weights with Xavier (Glorot) uniform initialisation and bias set to zeros.

Call torch.manual_seed(seed) immediately before the Xavier call, so the draw is reproducible. Return the tuple (weight, bias).

Xavier scales the bound by both fan-in and fan-out, which is what keeps the forward and backward variance balanced for symmetric activations.

Input

in_features = 4
out_features = 3
seed = 0

Output

(tensor([[-0.0069,  0.4967, -0.7620, -0.6813],
        [-0.3566,  0.2483, -0.0183,  0.7341],
        [-0.0822,  0.2450, -0.2798, -0.1820]]), tensor([0., 0., 0.]))