3 个月前

FlowSeq:基于生成流的非自回归条件序列生成

FlowSeq:基于生成流的非自回归条件序列生成

摘要

大多数序列到序列(seq2seq)模型均为自回归模型,其通过基于先前生成的标记来逐个生成每个标记。相比之下,非自回归seq2seq模型能够在单次前向传播中生成全部标记,从而通过GPU等硬件实现并行处理,显著提升效率。然而,直接联合建模所有标记的联合分布极具挑战性,即便采用日益复杂的模型结构,其生成准确率仍显著落后于自回归模型。本文提出一种基于隐变量模型的简单、高效且有效的非自回归序列生成方法。具体而言,我们引入生成流(generative flow)这一优雅的神经网络技术,用于建模复杂分布,并设计了多层流结构,专门用于建模序列隐变量的条件密度。我们在三个神经机器翻译(NMT)基准数据集上对该模型进行了评估,结果表明其性能可与当前最先进的非自回归NMT模型相媲美,且解码时间几乎不随序列长度变化,保持恒定。

代码仓库

XuezheMax/flowseq
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
machine-translation-on-iwslt2015-germanFlowSeq-base
BLEU score: 24.75
machine-translation-on-wmt2014-english-germanFlowSeq-large (IWD n = 15)
BLEU score: 22.94
Hardware Burden:
Operations per network pass:
machine-translation-on-wmt2014-english-germanFlowSeq-base
BLEU score: 18.55
Hardware Burden:
Operations per network pass:
machine-translation-on-wmt2014-english-germanFlowSeq-large (NPD n = 15)
BLEU score: 23.14
Hardware Burden:
Operations per network pass:
machine-translation-on-wmt2014-english-germanFlowSeq-large (NPD n = 30)
BLEU score: 23.64
Hardware Burden:
Operations per network pass:
machine-translation-on-wmt2014-english-germanFlowSeq-large
BLEU score: 20.85
Hardware Burden:
Operations per network pass:
machine-translation-on-wmt2014-german-englishFlowSeq-large (NPD n = 15)
BLEU score: 27.71
machine-translation-on-wmt2014-german-englishFlowSeq-large
BLEU score: 25.4
machine-translation-on-wmt2014-german-englishFlowSeq-base
BLEU score: 23.36
machine-translation-on-wmt2014-german-englishFlowSeq-large (IWD n=15)
BLEU score: 27.16
machine-translation-on-wmt2014-german-englishFlowSeq-large (NPD n = 30)
BLEU score: 28.29
machine-translation-on-wmt2016-english-1FlowSeq-large (NPD n=15)
BLEU score: 31.97
machine-translation-on-wmt2016-english-1FlowSeq-base
BLEU score: 29.26
machine-translation-on-wmt2016-english-1FlowSeq-large
BLEU score: 29.86
machine-translation-on-wmt2016-english-1FlowSeq-large (NPD n = 30)
BLEU score: 32.35
machine-translation-on-wmt2016-english-1FlowSeq-large (IWD n = 15)
BLEU score: 31.08
machine-translation-on-wmt2016-romanianFlowSeq-large (IWD n = 15)
BLEU score: 32.03
machine-translation-on-wmt2016-romanianFlowSeq-large (NPD n = 30)
BLEU score: 32.91
machine-translation-on-wmt2016-romanianFlowSeq-large (NPD n = 15)
BLEU score: 32.46
machine-translation-on-wmt2016-romanianFlowSeq-large
BLEU score: 30.69
machine-translation-on-wmt2016-romanianFlowSeq-base
BLEU score: 30.16

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
FlowSeq:基于生成流的非自回归条件序列生成 | 论文 | HyperAI超神经