My expected action is a tensor with softmax activation. However, when I use TD3, the current implementation will add this action_bias to the softmax-activated tensor, which breaks the action distribution.