Harvard transformer NLP 模型 openNMT 简介入门_opennmt结构图

项目网址：

OpenNMT - Open-Source Neural Machine Translation

logo：

一，从应用的层面先跑通 Harvard transformer

GitHub - harvardnlp/annotated-transformer: An annotated implementation of the Transformer paper.

​
 
git clone https://github.com/harvardnlp/annotated-transformer.git
cd annotated-transformer/
 
​

1. 环境搭建

 conda create --name ilustrate_transformer_env python=3.9
 conda activate ilustrate_transformer_env
 pip install -r requirements.txt  -i https://pypi.tuna.tsinghua.edu.cn/simple

问题：TypeError: issubclass() arg 1 must be a class

原因： 这是由python中的后端包之一的兼容性问题引起的问题，包“pydantic”

执行下面命令可以解决

 
 
python -m pip install -U pydantic spacy

 

会遇到下载不到数据的问题，因为有个网址废弃了：www.quest......

改成最新版本的torchtext的内容即可：

/home/hipper/anaconda3/envs/ilustrate_transformer_env/lib/python3.9/site-packages/torchtext/datasets/multi30k.py

 13 '''LL::
 14 URL = {
 15     "train": r"http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/training.tar.gz",
 16     "valid": r"http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/validation.tar.gz",
 17     "test": r"http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/mmt16_task1_test.tar.gz",
 18 }
 19
 20 MD5 = {
 21     "train": "20140d013d05dd9a72dfde46478663ba05737ce983f478f960c1123c6671be5e",
 22     "valid": "a7aa20e9ebd5ba5adce7909498b94410996040857154dab029851af3a866da8c",
 23     "test": "0681be16a532912288a91ddd573594fbdd57c0fbb81486eff7c55247e35326c2",
 24 }
 25 '''
 26 # TODO: Update URL to original once the server is back up (see https://github.com/pytorch/text/issues/1756)
 27 URL = {
 28     "train": r"https://raw.githubusercontent.com/neychev/small_DL_repo/master/datasets/Multi30k/training.tar.gz",
 29     "valid": r"https://raw.githubusercontent.com/neychev/small_DL_repo/master/datasets/Multi30k/validation.tar.gz",
 30     "test": r"https://raw.githubusercontent.com/neychev/small_DL_repo/master/datasets/Multi30k/mmt16_task1_test.tar.gz",
 31 }
 32
 33 MD5 = {
 34     "train": "20140d013d05dd9a72dfde46478663ba05737ce983f478f960c1123c6671be5e",
 35     "valid": "a7aa20e9ebd5ba5adce7909498b94410996040857154dab029851af3a866da8c",
 36     "test": "6d1ca1dba99e2c5dd54cae1226ff11c2551e6ce63527ebb072a1f70f72a5cd36",
 37 }

运行：

未完待续 ...

__________________________________________________

参考：

《The Annotated Transformer》翻译——注释和代码实现《Attention Is All You Need》_神洛华的博客-CSDN博客

图解transformer | The Illustrated Transformer_Ann's Blog的博客-CSDN博客

GitHub - harvardnlp/annotated-transformer: An annotated implementation of the Transformer paper.

OpenNMT - Open-Source Neural Machine Translation

flash attention 1,2:

Stanford CRFM

GitHub - Dao-AILab/flash-attention: Fast and memory-efficient exact attention