【摘要】 针对预训练模型推理代价过大的问题，提出一种基于傅里叶变换的加速推理方法。首先，使用傅里叶变换子层替代部分Transformer模块的自注意力子层，从而降低推理时间；然后，使用贪婪微调方法，即在预训练模型上对每一个编码层进行总体微调，使得模型中低层也具有高级语义知识，提高模型中低层的准确率与推理速度。为验证所提方法的有效性，在6个英文数据集上进行了实验。实验结果表明，在熵阈值为0.1时，与Dee-BERT(Dynamic early exiting for BERT)相比，所提方法的准确率平均下降了0.19个百分点，推理速度平均提升了73%；在熵阈值为0.5时，与Dee-BERT相比，所提方法的准确率平均下降了0.41个百分点，推理速度平均提升了62%，验证了所提方法可以有效提高推理速度。更多还原

【Abstract】 To solve the problem of excessive reasoning cost in pretraining models,an accelerated reasoning method based on Fourier transform was proposed.Firstly,the Fourier transform sublayer was used to replace the self-attention sublayer of some Transformer modules,thereby reducing inference time.Then,a greedy fine-tuning method was used,that was,the overall fine-tuning of each coding layer was performed on the pretraining model,so that the middle and lower layers of the model also had high-level semantic knowledge,improving the accuracy and reasoning speed of the middle and lower layers of the model.To verify the effectiveness of the proposed method,experiments were conducted on six English datasets.Experimental results show that when the entropy threshold is 0.1,the accuracy of the proposed method decreases by 0.19percentage points on average compared to Dynamic early exiting for BERT(Dee-BERT),and the inference speed increases by 73% on average;when the entropy threshold is 0.5,compared to Dee-BERT,the accuracy of the proposed method decreases by 0.41 percentage points on average,and the inference speed increases by 62% on average,verifying that the proposed method can effectively improve the inference speed.更多还原