fix: .normal_() on self.eps_weight.data

[ana-tamais]

Now it works with siamese architectures and when we need to double feedforward before gathering loss and backpropagating.