即这里的不变性特指：推理时不随输入变化而变化，但在训练过程中，矩阵是可以根据需要去做梯度下降而变化的Eliminates the bias of subword tokenisation: where by popular subwords are overrepresented and rare or new text are underrepresented or split into significantly less meaningful units.torchaudio: Docum