使用pytorch+transformers复现了SimCSE论文中的有监督训练和无监督训练方法

Related tags

Text Data & NLPSimCSE
Overview

SimCSE复现

项目描述

SimCSE是一种简单但是很巧妙的NLP对比学习方法,创新性地引入Dropout的方式,对样本添加噪声,从而达到对正样本增强的目的。 该框架的训练目的为:对于batch中的每个样本,拉近其与正样本之间的距离,拉远其与负样本之间的距离,使得模型能够在大规模无监督语料(也可以使用有监督的语料)中学习到文本相似关系。 详见论文:Simple Contrastive Learning of Sentence EmbeddingsSimCSE官方代码仓库

本项目使用pytorch+transformers复现了SimCSE论文中的有监督训练和无监督训练方法,并且在STS-B数据集上进行消融实验,评价指标为Spearman相关系数,预训练模型为Bert-base-uncased, 验证了SimCSE的有效性。在STS-B数据集上,有监督训练和无监督训练的复现效果如下表。

在无监督训练中,dropout=0.1时,复现效果比原文略差,但也比较接近。当dropout=0.2时,复现效果比原文略高。 ** 但在有监督训练中,不知是否由于batch size过小(原论文使用512),复现效果与论文的效果相差较远,后续会进行排查。 **

训练方法 learning rate batch size dropout Spearman’s correlation
原论文 无监督 3e-5 64 0.1 0.763
复现 无监督 3e-5 64 0.2 0.771
复现 无监督 3e-5 64 0.1 0.748
原论文 有监督 5e-5 512 0.1 0.816
复现 有监督 5e-5 64 0.1 0.764

运行环境

python==3.6、transformers==3.1.0、torch==1.6.0

项目结构

  • data:存放训练数据
    • stsbenchmark:STS-B数据集
      • sts-dev.csv:STS-B验证集
      • sts-test.csv:STS-B验测试集
    • nli_for_simcse.csv:数量275601为的NLI数据集
    • wiki1m_for_simcse.txt:维基百科上获取的100w的文本
  • output:输出目录
  • pretrain_model:预训练模型存放位置
  • script:脚本存放位置。
  • dataset.py
  • model.py:模型代码,包含有监督和无监督损失函数的计算方式
  • train.py:训练代码

使用方法

Quick Start

下载训练数据:

bash script/download_nli.sh
bash script/download_wiki.sh

无监督训练,运行脚本

bash script/run_unsup_train.sh

有监督训练,运行脚本

bash script/run_sup_train.sh

实验

无监督训练

从前四条实验数据中可以看到,较大的batch size在一定程度上可以增加模型的泛化性。

dropout为0.2的时候,训练效果比0.1与0.3更好,有可能dropout=0.1加入的噪声过小,而dropout=0.3加入的噪声过大,增强得到的样本与原始样本差异较大。

learning rate batch size dropout 在哪一步得到best checkpoint 验证集上的得分 测试集上的得分
3e-5 256 0.1 6000 0.800 0.761
3e-5 128 0.1 4200 0.799 0.747
3e-5 64 0.1 10900 0.803 0.748
3e-5 32 0.1 21300 0.787 0.714
3e-5 64 0.2 11200 0.811 0.771
3e-5 64 0.3 6300 0.781 0.745
1e-5 64 0.1 16400 0.798 0.751

有监督训练

有监督实验的复现结果未达到预期,超参数相同时,在验证集上的得分略高于无监督,但是在测试集上,得分基本没有差异。增大有监督训练的学习率,有监督的训练的得分略高于无监督训练, 但还是与论文声称的0.816相差较远,原论文使用512的batch size, 不知是否由于batch size的设置有关,后续会对有监督的训练代码进一步排查。

不过从训练曲线可以看到,有监督训练的收敛速度明显快于无监督训练,这也符合我们的认知。

训练方法 learning rate batch size dropout 在哪一步得到best checkpoint 验证集上的得分 测试集上的得分
无监督 3e-5 64 0.1 10900 0.803 0.748
有监督 3e-5 64 0.1 200 0.810 0.748
有监督 5e-5 64 0.1 2300 0.809 0.764
有监督 3e-5 32 0.1 200 0.808 0.743
有监督 5e-5 32 0.1 200 0.806 0.746

无监督训练过程中,验证集得分的变化曲线: avatar

有监督训练过程中,验证集得分的变化曲线: avatar

REFERENCE

TODO

  • 排查有监督学习的效果不符合预期的原因
🤗 The largest hub of ready-to-use NLP datasets for ML models with fast, easy-to-use and efficient data manipulation tools

🤗 The largest hub of ready-to-use NLP datasets for ML models with fast, easy-to-use and efficient data manipulation tools

Hugging Face 15k Jan 02, 2023
Semantic search through a vectorized Wikipedia (SentenceBERT) with the Weaviate vector search engine

Semantic search through Wikipedia with the Weaviate vector search engine Weaviate is an open source vector search engine with build-in vectorization a

SeMI Technologies 191 Dec 26, 2022
this repository has datasets containing information of Uber pickups in NYC from April 2014 to September 2014 and January to June 2015. data Analysis , virtualization and some insights are gathered here

uber-pickups-analysis Data Source: https://www.kaggle.com/fivethirtyeight/uber-pickups-in-new-york-city Information about data set The dataset contain

1 Nov 02, 2021
Code to reprudece NeurIPS paper: Accelerated Sparse Neural Training: A Provable and Efficient Method to Find N:M Transposable Masks

Accelerated Sparse Neural Training: A Provable and Efficient Method to FindN:M Transposable Masks Recently, researchers proposed pruning deep neural n

itay hubara 4 Feb 23, 2022
自然言語で書かれた時間情報表現を抽出/規格化するルールベースの解析器

ja-timex 自然言語で書かれた時間情報表現を抽出/規格化するルールベースの解析器 概要 ja-timex は、現代日本語で書かれた自然文に含まれる時間情報表現を抽出しTIMEX3と呼ばれるアノテーション仕様に変換することで、プログラムが利用できるような形に規格化するルールベースの解析器です。

Yuki Okuda 116 Nov 09, 2022
Creating a chess engine using GPT-3

GPT3Chess Creating a chess engine using GPT-3 Code for my article : https://towardsdatascience.com/gpt-3-play-chess-d123a96096a9 My game (white) vs GP

19 Dec 17, 2022
A website which allows you to play with the GPT-2 transformer

transformers A website which allows you to play with the GPT-2 model Built with ❤️ by raphtlw Table of contents Model Setup About Contributors Model T

raphtlw 2 Jan 27, 2022
Train 🤗-transformers model with Poutyne.

poutyne-transformers Train 🤗 -transformers models with Poutyne. Installation pip install poutyne-transformers Example import torch from transformers

Lennart Keller 2 Dec 18, 2022
An open source framework for seq2seq models in PyTorch.

pytorch-seq2seq Documentation This is a framework for sequence-to-sequence (seq2seq) models implemented in PyTorch. The framework has modularized and

International Business Machines 1.4k Jan 02, 2023
SimBERT升级版(SimBERTv2)!

RoFormer-Sim RoFormer-Sim,又称SimBERTv2,是我们之前发布的SimBERT模型的升级版。 介绍 https://kexue.fm/archives/8454 训练 tensorflow 1.14 + keras 2.3.1 + bert4keras 0.10.6 下载

317 Dec 23, 2022
Dual languaged (rus+eng) tool for packing and unpacking archives of Silky Engine.

SilkyArcTool English Dual languaged (rus+eng) GUI tool for packing and unpacking archives of Silky Engine. It is not the same arc as used in Ai6WIN. I

Tester 5 Sep 15, 2022
A simple Speech Emotion Recognition (SER) API created using Flask and running in a Docker container.

keyword_searching Steps to use this Python scripts: (1)Paste this script into the file folder containing the PDF files you need to search from; (2)Thi

2 Nov 11, 2022
Extract city and country mentions from Text like GeoText without regex, but FlashText, a Aho-Corasick implementation.

flashgeotext ⚡ 🌍 Extract and count countries and cities (+their synonyms) from text, like GeoText on steroids using FlashText, a Aho-Corasick impleme

Ben 57 Dec 16, 2022
The ability of computer software to identify words and phrases in spoken language and convert them to human-readable text

speech-recognition-py Speech recognition is the ability of computer software to identify words and phrases in spoken language and convert them to huma

Deepangshi 1 Apr 03, 2022
Code for using and evaluating SpanBERT.

SpanBERT This repository contains code and models for the paper: SpanBERT: Improving Pre-training by Representing and Predicting Spans. If you prefer

Meta Research 798 Dec 30, 2022
Nmt - TensorFlow Neural Machine Translation Tutorial

Neural Machine Translation (seq2seq) Tutorial Authors: Thang Luong, Eugene Brevdo, Rui Zhao (Google Research Blogpost, Github) This version of the tut

6.1k Dec 29, 2022
Under the hood working of transformers, fine-tuning GPT-3 models, DeBERTa, vision models, and the start of Metaverse, using a variety of NLP platforms: Hugging Face, OpenAI API, Trax, and AllenNLP

Transformers-for-NLP-2nd-Edition @copyright 2022, Packt Publishing, Denis Rothman Contact me for any question you have on LinkedIn Get the book on Ama

Denis Rothman 150 Dec 23, 2022
EMNLP'2021: Can Language Models be Biomedical Knowledge Bases?

BioLAMA BioLAMA is biomedical factual knowledge triples for probing biomedical LMs. The triples are collected and pre-processed from three sources: CT

DMIS Laboratory - Korea University 41 Nov 18, 2022
基于Transformer的单模型、多尺度的VAE模型

UniVAE 基于Transformer的单模型、多尺度的VAE模型 介绍 https://kexue.fm/archives/8475 依赖 需要大于0.10.6版本的bert4keras(当前还没有推到pypi上,可以直接从GitHub上clone最新版)。 引用 @misc{univae,

苏剑林(Jianlin Su) 49 Aug 24, 2022
MASS: Masked Sequence to Sequence Pre-training for Language Generation

MASS: Masked Sequence to Sequence Pre-training for Language Generation

Microsoft 1.1k Dec 17, 2022