You can safely skip l=square_loss(y_pred, y_tr).

The derivative of the loss function is already reflected in gradient_W calculation.