ConvBERT-Prod

Overview

ConvBERT

目录

0. 仓库结构

root:[./]
|--convbert_base_outputs
|      |--args.json
|      |--best.pdparams
|      |      |--model_config.json
|      |      |--model_state.pdparams
|      |      |--tokenizer_config.json
|      |      |--vocab.txt
|--convbert_infer
|      |--inference.pdiparams
|      |--inference.pdiparams.info
|      |--inference.pdmodel
|      |--tokenizer_config.json
|      |--vocab.txt
|--deploy
|      |--inference_python
|      |      |--infer.py
|      |      |--README.md
|      |--serving_python
|      |      |--config.yml
|      |      |--convbert_client
|      |      |      |--serving_client_conf.prototxt
|      |      |      |--serving_client_conf.stream.prototxt
|      |      |--convbert_server
|      |      |      |--inference.pdiparams
|      |      |      |--inference.pdmodel
|      |      |      |--serving_server_conf.prototxt
|      |      |      |--serving_server_conf.stream.prototxt
|      |      |--PipelineServingLogs
|      |      |      |--pipeline.log
|      |      |      |--pipeline.log.wf
|      |      |      |--pipeline.tracer
|      |      |--pipeline_http_client.py
|      |      |--ProcessInfo.json
|      |      |--README.md
|      |      |--web_service.py
|--images
|      |--convbert_framework.jpg
|      |--py_serving_client_results.jpg
|      |--py_serving_startup_visualization.jpg
|--LICENSE
|--output_inference_engine.npy
|--output_predict_engine.npy
|--paddlenlp
|--print_project_tree.py
|--README.md
|--requirements.txt
|--shell
|      |--export.sh
|      |--inference_python.sh
|      |--predict.sh
|      |--train.sh
|      |--train_dist.sh
|--test_tipc
|      |--common_func.sh
|      |--configs
|      |      |--ConvBERT
|      |      |      |--model_linux_gpu_normal_normal_serving_python_linux_gpu_cpu.txt
|      |      |      |--train_infer_python.txt
|      |--docs
|      |      |--test_serving.md
|      |      |--test_train_inference_python.md
|      |      |--tipc_guide.png
|      |      |--tipc_serving.png
|      |      |--tipc_train_inference.png
|      |--output
|      |      |--python_infer_cpu_usemkldnn_False_threads_null_precision_null_batchsize_null.log
|      |      |--python_infer_gpu_usetrt_null_precision_null_batchsize_null.log
|      |      |--results_python.log
|      |      |--results_serving.log
|      |      |--server_infer_gpu_pipeline_http_usetrt_null_precision_null_batchsize_1.log
|      |--README.md
|      |--test_serving.sh
|      |--test_train_inference_python.sh
|--tools
|      |--export_model.py
|      |--predict.py
|--train.log
|--train.py

1. 简介

论文: ConvBERT: Improving BERT with Span-based Dynamic Convolution

摘要: 像BERT及其变体这样的预训练语言模型最近在各种自然语言理解任务中取得了令人印象深刻的表现。然而,BERT严重依赖全局自注意力块,因此需要大量内存占用和计算成本。 虽然它的所有注意力头从全局角度查询整个输入序列以生成注意力图,但我们观察到一些头只需要学习局部依赖,这意味着存在计算冗余。 因此,我们提出了一种新颖的基于跨度的动态卷积来代替这些自注意力头,以直接对局部依赖性进行建模。新的卷积头与其余的自注意力头一起形成了一个新的混合注意力块,在全局和局部上下文学习中都更有效。 我们为 BERT 配备了这种混合注意力设计并构建了一个ConvBERT模型。实验表明,ConvBERT 在各种下游任务中明显优于BERT及其变体,具有更低的训练成本和更少的模型参数。 值得注意的是,ConvBERT-base 模型达到86.4GLUE分数,比ELECTRA-base高0.7,同时使用不到1/4的训练成本。

2. 数据集和复现精度

数据集为SST-2

模型 sst-2 dev acc (复现精度)
ConvBERT 0.9461

3. 准备环境与数据

3.1 准备环境

  • 下载代码
git clone https://github.com/junnyu/ConvBERT-Prod.git
  • 安装paddlepaddle
# 需要安装2.2及以上版本的Paddle,如果
# 安装GPU版本的Paddle
pip install paddlepaddle-gpu==2.2.0
# 安装CPU版本的Paddle
pip install paddlepaddle==2.2.0

更多安装方法可以参考:Paddle安装指南

  • 安装requirements
pip install -r requirements.txt

3.2 准备数据

SST-2数据已经集成在paddlenlp仓库中。

3.3 准备模型

如果您希望直接体验评估或者预测推理过程,可以直接根据第2章的内容下载提供的预训练模型,直接体验模型评估、预测、推理部署等内容。

4. 开始使用

4.1 模型训练

  • 单机单卡训练
export CUDA_VISIBLE_DEVICES=0
python -m paddle.distributed.launch --gpus "0" train.py \
    --model_type convbert \
    --model_name_or_path convbert-base \
    --task_name sst-2 \
    --max_seq_length 128 \
    --learning_rate 1e-4 \
    --num_train_epochs 3 \
    --output_dir ./convbert_base_outputs/ \
    --logging_steps 100 \
    --save_steps 400 \
    --batch_size 32   \
    --warmup_proportion 0.1

部分训练日志如下所示。

====================================================================================================
global step 2500/6315, epoch: 1, batch: 394, rank_id: 0, loss: 0.140546, lr: 0.0000671182, speed: 3.7691 step/s
global step 2600/6315, epoch: 1, batch: 494, rank_id: 0, loss: 0.062813, lr: 0.0000653589, speed: 4.1413 step/s
global step 2700/6315, epoch: 1, batch: 594, rank_id: 0, loss: 0.051268, lr: 0.0000635996, speed: 4.1867 step/s
global step 2800/6315, epoch: 1, batch: 694, rank_id: 0, loss: 0.133289, lr: 0.0000618403, speed: 4.1769 step/s
eval loss: 0.342346, acc: 0.9461009174311926,
eval done total : 1.9056718349456787 s
====================================================================================================
  • 单机多卡训练
export CUDA_VISIBLE_DEVICES=0,1,2,3
python -m paddle.distributed.launch --gpus "0,1,2,3" train.py \
    --model_type convbert \
    --model_name_or_path convbert-base \
    --task_name sst-2 \
    --max_seq_length 128 \
    --learning_rate 1e-4 \
    --num_train_epochs 3 \
    --output_dir ./convbert_base_outputs/ \
    --logging_steps 100 \
    --save_steps 400 \
    --batch_size 32   \
    --warmup_proportion 0.1

更多配置参数可以参考train.pyget_args_parser函数。

4.2 模型评估

该项目中,训练与评估脚本同时进行,请查看训练过程中的评价指标。

4.3 模型预测

  • 使用GPU预测
python tools/predict.py --model_path=./convbert_base_outputs/best.pdparams

对于下面的文本进行预测

the problem , it is with most of these things , is the script .

最终输出结果为label_id: 0, prob: 0.9959235191345215,表示预测的标签ID是0,置信度为0.9959

  • 使用CPU预测
python tools/predict.py --model_path=./convbert_base_outputs/best.pdparams --device=cpu

对于下面的文本进行预测

the problem , it is with most of these things , is the script .

最终输出结果为label_id: 0, prob: 0.995919406414032,表示预测的标签ID是0,置信度为0.9959

5. 模型推理部署

5.1 基于Inference的推理

Inference推理教程可参考:链接

5.2 基于Serving的服务化部署

Serving部署教程可参考:链接

6. TIPC自动化测试脚本

以Linux基础训练推理测试为例,测试流程如下。

  • 运行测试命令
bash test_tipc/test_train_inference_python.sh test_tipc/configs/ConvBERT/train_infer_python.txt whole_train_whole_infer

如果运行成功,在终端中会显示下面的内容,具体的日志也会输出到test_tipc/output/文件夹中的文件中。

�[33m Run successfully with command - python train.py --save_steps 400      --max_steps=6315           !  �[0m
�[33m Run successfully with command - python tools/export_model.py --model_path=./convbert_base_outputs/best.pdparams --save_inference_dir ./convbert_infer      !  �[0m
�[33m Run successfully with command - python deploy/inference_python/infer.py --model_dir ./convbert_infer --use_gpu=True               > ./test_tipc/output/python_infer_gpu_usetrt_null_precision_null_batchsize_null.log 2>&1 !  �[0m
�[33m Run successfully with command - python deploy/inference_python/infer.py --model_dir ./convbert_infer --use_gpu=False --benchmark=False               > ./test_tipc/output/python_infer_cpu_usemkldnn_False_threads_null_precision_null_batchsize_null.log 2>&1 !  �[0m

7. 注意

为了可以使用静态图导出功能,本项目修改了paddlenlp仓库中的convbert模型,主要修改部分如下。

    1. 使用paddle.shape而不是tensor.shape获取tensor的形状。
    1. F.unfold对于静态图不怎么友好,只好采用for循环。
if self.conv_type == "sdconv":
    bs = paddle.shape(q)[0]
    seqlen = paddle.shape(q)[1]
    mixed_key_conv_attn_layer = self.key_conv_attn_layer(query)
    conv_attn_layer = mixed_key_conv_attn_layer * q

    # conv_kernel_layer
    conv_kernel_layer = self.conv_kernel_layer(conv_attn_layer)
    conv_kernel_layer = tensor.reshape(
        conv_kernel_layer, shape=[-1, self.conv_kernel_size, 1])
    conv_kernel_layer = F.softmax(conv_kernel_layer, axis=1)
    conv_out_layer = self.conv_out_layer(query)
    conv_out_layer = paddle.stack(
        [
            paddle.slice(F.pad(conv_out_layer, pad=[
                            self.padding, self.padding], data_format="NLC"), [1], starts=[i], ends=[i+seqlen])
            for i in range(self.conv_kernel_size)
        ],
        axis=-1,
    )
    conv_out_layer = tensor.reshape(
        conv_out_layer,
        shape=[-1, self.head_dim, self.conv_kernel_size])
    conv_out_layer = tensor.matmul(conv_out_layer, conv_kernel_layer)
    conv_out = tensor.reshape(
        conv_out_layer,
        shape=[bs, seqlen, self.num_heads, self.head_dim])

8. LICENSE

本项目的发布受Apache 2.0 license许可认证。

9. 参考链接与文献

TODO

Owner
yujun
Please show me your code.
yujun
Ecco is a python library for exploring and explaining Natural Language Processing models using interactive visualizations.

Visualize, analyze, and explore NLP language models. Ecco creates interactive visualizations directly in Jupyter notebooks explaining the behavior of Transformer-based language models (like GPT2, BER

Jay Alammar 1.6k Dec 25, 2022
NLP Text Classification

多标签文本分类任务 近年来随着深度学习的发展,模型参数的数量飞速增长。为了训练这些参数,需要更大的数据集来避免过拟合。然而,对于大部分NLP任务来说,构建大规模的标注数据集非常困难(成本过高),特别是对于句法和语义相关的任务。相比之下,大规模的未标注语料库的构建则相对容易。为了利用这些数据,我们可以

Jason 1 Nov 11, 2021
Persian Bert For Long-Range Sequences

ParsBigBird: Persian Bert For Long-Range Sequences The Bert and ParsBert algorithms can handle texts with token lengths of up to 512, however, many ta

Sajjad Ayoubi 63 Dec 14, 2022
A simple word search made in python

Word Search Puzzle A simple word search made in python Usage $ python3 main.py -h usage: main.py [-h] [-c] [-f FILE] Generates a word s

Magoninho 16 Mar 10, 2022
TPlinker for NER 中文/英文命名实体识别

本项目是参考 TPLinker 中HandshakingTagging思想,将TPLinker由原来的关系抽取(RE)模型修改为命名实体识别(NER)模型。

GodK 113 Dec 28, 2022
Dope Wars game engine on StarkNet L2 roll-up

RYO Dope Wars game engine on StarkNet L2 roll-up. What TI-83 drug wars built as smart contract system. Background mechanism design notion here. Initia

104 Dec 04, 2022
Awesome-NLP-Research (ANLP)

Awesome-NLP-Research (ANLP)

Language, Information, and Learning at Yale 72 Dec 19, 2022
Python bot created with Selenium that can guess the daily Wordle word correct 96.8% of the time.

Wordle_Bot Python bot created with Selenium that can guess the daily Wordle word correct 96.8% of the time. It will log onto the wordle website and en

Lucas Polidori 15 Dec 11, 2022
spaCy-wrap: For Wrapping fine-tuned transformers in spaCy pipelines

spaCy-wrap: For Wrapping fine-tuned transformers in spaCy pipelines spaCy-wrap is minimal library intended for wrapping fine-tuned transformers from t

Kenneth Enevoldsen 32 Dec 29, 2022
Smart discord chatbot integrated with Dialogflow

academic-NLP-chatbot Smart discord chatbot integrated with Dialogflow to interact with students naturally and manage different classes in a school. De

Tom Huynh 5 Oct 24, 2022
The following links explain a bit the idea of semantic search and how search mechanisms work by doing retrieve and rerank

Main Idea The following links explain a bit the idea of semantic search and how search mechanisms work by doing retrieve and rerank Semantic Search Re

Sergio Arnaud Gomez 2 Jan 28, 2022
Fine-tune GPT-3 with a Google Chat conversation history

Google Chat GPT-3 This repo will help you fine-tune GPT-3 with a Google Chat conversation history. The trained model will be able to converse as one o

Nate Baer 7 Dec 10, 2022
Pre-Training with Whole Word Masking for Chinese BERT

Pre-Training with Whole Word Masking for Chinese BERT

Yiming Cui 7.7k Dec 31, 2022
text to speech toolkit. 好用的中文语音合成工具箱,包含语音编码器、语音合成器、声码器和可视化模块。

ttskit Text To Speech Toolkit: 语音合成工具箱。 安装 pip install -U ttskit 注意 可能需另外安装的依赖包:torch,版本要求torch=1.6.0,=1.7.1,根据自己的实际环境安装合适cuda或cpu版本的torch。 ttskit的

KDD 483 Jan 04, 2023
[ICCV 2021] Instance-level Image Retrieval using Reranking Transformers

Instance-level Image Retrieval using Reranking Transformers Fuwen Tan, Jiangbo Yuan, Vicente Ordonez, ICCV 2021. Abstract Instance-level image retriev

UVA Computer Vision 86 Dec 28, 2022
Facilitating the design, comparison and sharing of deep text matching models.

MatchZoo Facilitating the design, comparison and sharing of deep text matching models. MatchZoo 是一个通用的文本匹配工具包,它旨在方便大家快速的实现、比较、以及分享最新的深度文本匹配模型。 🔥 News

Neural Text Matching Community 3.7k Jan 02, 2023
End-to-end MLOps pipeline of a BERT model for emotion classification.

image source EmoBERT-MLOps The goal of this repository is to build an end-to-end MLOps pipeline based on the MLOps course from Made with ML, but this

Dimitre Oliveira 4 Nov 06, 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
Active learning for text classification in Python

Active Learning allows you to efficiently label training data in a small-data scenario.

Webis 375 Dec 28, 2022
Final Project Bootcamp Zero

The Quest (Pygame) Descripción Este es el repositorio de código The-Quest para el proyecto final Bootcamp Zero de KeepCoding. El juego consiste en la

Seven-z01 1 Mar 02, 2022