《论文阅读》Structured Pruning Learns Compact and Accurate Models

89 篇文章 51 订阅 ¥29.90 ¥99.00
本文介绍了如何结合剪枝和知识蒸馏技术来压缩模型,提出了一种名为Coarse- and Fine-grained Pruning (CoFi) 的方法。CoFi 在多头自注意力层和前馈神经网络层进行剪枝,并动态学习教师模型和学生模型之间的层映射关系,以优化蒸馏过程。实验表明,这种方法可以学到紧凑且精确的模型。
摘要由CSDN通过智能技术生成

Structured Pruning Learns Compact and Accurate Models

前言

今天读了一篇论文《Structured Pruning Learns Compact and Accurate Models》,主要讲的是如何结合剪枝和蒸馏对模型进行压缩,以实现好的效果。

主流策略

目前压缩模型的主流的方法无非两种:剪枝(Pruning)蒸馏(Distillation),下面分别说一下两者的优势和劣势。

剪枝

Advantages Disadvantages
显著地减少模型的大小 很难像蒸馏那样实现大幅度地加速

蒸馏

Advantages Disadvantages
显著地减少模型的大小 需要大量无标注数据
速度快
  • 0
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 0
    评论
Structured relation discovery is the process of identifying and extracting relationships between entities in a given text corpus. This is a challenging task due to the complexity of natural language and the varied ways in which relationships can be expressed. Generative models offer a promising approach to structured relation discovery, as they can learn to generate realistic text that exhibits the desired relationships. Generative models are a type of machine learning model that can learn to create new examples of text by modeling the underlying probability distribution of the text. This can be achieved through various techniques, such as recurrent neural networks (RNNs), Markov chains, and probabilistic graphical models. These models are trained on a large corpus of text, and they learn to generate new text that is similar to the training data. To use generative models for structured relation discovery, we need to first define the relationships we are interested in. For example, we might want to identify relationships between people and organizations, or between products and their features. We can then train a generative model on a corpus of text that contains these relationships. The model will learn to generate text that exhibits the desired relationships, and we can use this to identify and extract relationships in new text. One popular approach to generative modeling for structured relation discovery is the use of conditional random fields (CRFs). CRFs are a type of probabilistic graphical model that can learn to predict the probability of a particular sequence of labels (i.e., the relationships between entities) given a sequence of observations (i.e., the text). By training a CRF on a corpus of text that contains the desired relationships, we can learn to predict the relationships in new text. Another approach to generative modeling for structured relation discovery is the use of RNNs. RNNs are a type of neural network that can learn to model the sequential nature of text. By training an RNN on a corpus of text that contains the desired relationships, we can learn to generate new text that exhibits the relationships. We can then use this generated text to identify and extract relationships in new text. In conclusion, generative models offer a promising approach to structured relation discovery, as they can learn to generate realistic text that exhibits the desired relationships. By training these models on a corpus of text that contains the relationships we are interested in, we can use them to identify and extract relationships in new text.
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

365JHWZGo

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值