def content_loss(Y_hat, Y):
    # We detach the target content from the tree used to dynamically compute
    # the gradient: this is a stated value, not a variable. Otherwise the loss
    # will throw an error.
    return torch.square(Y_hat - Y.detach()).mean()