Inference acceleration method and device based on collaborative meta-learning
The invention discloses a reasoning acceleration method and device based on collaborative meta-learning. The method comprises the following steps: constructing an inference model; wherein the inference model is a transform-based model in which an early-leaving classifier is additionally added behind...
Gespeichert in:
Hauptverfasser: | , , , , , , |
---|---|
Format: | Patent |
Sprache: | chi ; eng |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | The invention discloses a reasoning acceleration method and device based on collaborative meta-learning. The method comprises the following steps: constructing an inference model; wherein the inference model is a transform-based model in which an early-leaving classifier is additionally added behind each transform layer, and the inference model is a transform-based model in which an early-leaving classifier is additionally added behind each transform layer; training an inference model; wherein the step of training the inference model comprises the steps of optimizing initial parameters of the inference model, and transmitting the optimized parameters into an Adam optimizer for gradient updating so as to train an early leaving classifier; and performing task prediction by using the trained inference model, and outputting a prediction result of the task based on an entropy exit mechanism. According to the method, the reasoning speed of the model can be improved while the small performance loss is kept.
本发明公开一种基 |
---|