Model training method and device and computer equipment

Abstract

The invention discloses a model training method, a model comprises a first twin network composed of a first encoder and a second encoder, and the method comprises the steps: obtaining a first high-dimensional vector outputted by the first encoder after processing first data, and a second high-dimensional vector outputted by the second encoder after processing second data; training the first twin network through a first loss function until the first loss function reaches a minimum value, and converging the training of the first twin network; obtaining a first parameter set corresponding to thefirst encoder and a second parameter set corresponding to the second encoder when the first twin network training converges; forming a second twin network by the first encoder and the first recovery network under the first parameter set, and forming a third twin network by the second encoder and the second recovery network under the second parameter set; and respectively training the second twin network and the third twin network to converge through the second loss function to obtain a network model system for executing the dual task. Model construction and training cost is saved, and model training is improved.

Description

technical field [0001] This application relates to the field of computers, in particular to a model training method, device and computer equipment. Background technique [0002] In deep learning, there are many Sequence-2-Sequence prediction tasks with a dual relationship, such as speech recognition and speech synthesis, machine translation (Chinese-to-English and English-to-Chinese), text-to-image and picture-to-speech, text summarization and text generation, etc. . Due to the diversity of sequences, generally larger models and large-scale training data are required to achieve satisfactory results. In practice, the construction of the above task models is one-way. For example, speech recognition can only process the reasoning process from voice data to text, and speech synthesis can only process the reasoning process from text data to audio; similarly, the Chinese-to-English model can only process Chinese is translated into English, and the English-to-Chinese model can on...

AI Technical Summary

Problems solved by technology

[0003] The main purpose of this application is to provide a model training method, aiming to solve the techni

Images