AI 科技评论按：刚刚在Github上发布了开源 Pytorch-Transformers 1.0，该项目支持BERT, GPT, GPT-2, Transfo-XL, XLNet, XLM等，并包含27个预训练模型。
我们来看。
哪些支持PyTorch-Transformers（此前叫做pytorch-pretrained-bert）是面向自然语言处理，当前性能最高的预训练模型开源库。 该开源库现在包含了 PyTorch 实现、预训练模型权重、运行脚本和以下模型的转换工具：
1、谷歌的 BERT
论文：“BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”
2、OpenAI 的GPT
论文：“ Improving Language Understanding by Generative Pre-Training”
3、OpenAI 的 GPT-2
论文：“ Language Models are Unsupervised Multitask Learners”
4、谷歌和 CMU 的 Transformer-XL
论文：“ Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context”
5、谷歌和 CMU 的XLNet
论文：“XLNet: Generalized Autoregressive Pretraining for Language Understanding”
6、Facebook的 XLM
论文：“ Cross-lingual Language Model Pretraining”
这些实现都在几个数据集（参见示例脚本）上进行了测试，性能与原始实现相当，例如 BERT中文全词覆盖在 SQuAD数据集上的F1分数为93 ， OpenAI GPT 在 RocStories上的F1分数为88， Transformer-XL在 WikiText 103 上的困惑度为18.3， XLNet在STS-B的皮尔逊相关系数为0.916。
27个预训练模型项目中提供了27个预训练模型，下面是这些模型的完整列表，以及每个模型的简短介绍。
ArchitectureShortcut nameDetails of the modelBERTbert-base-uncased12-layer, 768-hidden, 12-heads, 110M parameters Trained on lower-cased English textbert-large-uncased24-layer, 1024-hidden, 16-heads, 340M parameters Trained on lower-cased English textbert-base-cased12-layer, 768-hidden, 12-heads, 110M parameters Trained on cased English textbert-large-cased24-layer, 1024-hidden, 16-heads, 340M parameters Trained on cased English textbert-base-multilingual-uncased(Original, not recommended) 12-layer, 768-hidden, 12-heads, 110M parameters Trained on lower-cased text in the top 102 languages with the largest Wikipedias (see details)bert-base-multilingual-cased(New, recommended) 12-layer, 768-hidden, 12-heads, 110M parameters Trained on cased text in the top 104 languages with the largest Wikipedias (see details)bert-base-chinese12-layer, 768-hidden, 12-heads, 110M parameters Trained on cased Chinese Simplified and Traditional textbert-base-german-cased12-layer, 768-hidden, 12-heads, 110M parameters Trained on cased German text by Deepset.ai (see details on deepset.ai website)bert-large-uncased-whole-word-masking24-layer, 1024-hidden, 16-heads, 340M parameters Trained on lower-cased English text using Whole-Word-Masking (see details)bert-large-cased-whole-word-masking24-layer, 1024-hidden, 16-heads, 340M parameters Trained on cased English text using Whole-Word-Masking (see details)bert-large-uncased-whole-word-masking-finetuned-squad24-layer, 1024-hidden, 16-heads, 340M parameters The bert-large-uncased-whole-word-maskingmodel fine-tuned on SQuAD (see details of fine-tuning in the example section)bert-large-cased-whole-word-masking-finetuned-squad24-layer, 1024-hidden, 16-heads, 340M parameters The bert-large-cased-whole-word-maskingmodel fine-tuned on SQuAD (see details of fine-tuning in the example section)bert-base-cased-finetuned-mrpc12-layer, 768-hidden, 12-heads, 110M parameters The bert-base-casedmodel fine-tuned on MRPC (see details of fine-tuning in the example section)GPTopenai-gpt12-layer, 768-hidden, 12-heads, 110M parameters OpenAI GPT English modelGPT-2gpt212-layer, 768-hidden, 12-heads, 117M parameters OpenAI GPT-2 English modelgpt2-medium24-layer, 1024-hidden, 16-heads, 345M parameters OpenAI’s Medium-sized GPT-2 English modelTransformer-XLtransfo-xl-wt10318-layer, 1024-hidden, 16-heads, 257M parameters English model trained on wikitext-103XLNetxlnet-base-cased12-layer, 768-hidden, 12-heads, 110M parameters XLNet English modelxlnet-large-cased24-layer, 1024-hidden, 16-heads, 340M parameters XLNet Large English modelXLMxlm-mlm-en-204812-layer, 1024-hidden, 8-heads XLM English modelxlm-mlm-ende-102412-layer, 1024-hidden, 8-heads XLM English-German Multi-language modelxlm-mlm-enfr-102412-layer, 1024-hidden, 8-heads XLM English-French Multi-language modelxlm-mlm-enro-102412-layer, 1024-hidden, 8-heads XLM English-Romanian Multi-language modelxlm-mlm-xnli15-102412-layer, 1024-hidden, 8-heads XLM Model pre-trained with MLM on the 15 XNLI languages.xlm-mlm-tlm-xnli15-102412-layer, 1024-hidden, 8-heads XLM Model pre-trained with MLM + TLM on the 15 XNLI languages.xlm-clm-enfr-102412-layer, 1024-hidden, 8-heads XLM English model trained with CLM (Causal Language Modeling)xlm-clm-ende-102412-layer, 1024-hidden, 8-heads XLM English-German Multi-language model trained with CLM (Causal Language Modeling)
例子BERT-base和BERT-large分别是110M和340M参数模型，并且很难在单个GPU上使用推荐的批量大小对其进行微调，来获得良好的性能（在大多数情况下批量大小为32）。
为了帮助微调这些模型，我们提供了几种可以在微调脚本中激活的技术 run_bert_classifier.py 和 run_bert_squad.py：梯度累积（gradient-accumulation），多GPU训练（multi-gpu training），分布式训练（distributed training ）和16- bits 训练（ 16-bits training）。注意，这里要使用分布式训练和16- bits 训练，你需要安装NVIDIA的apex扩展。
作者在doc中展示了几个基于BERT原始实现（/）和扩展的微调示例，分别为：
九个不同GLUE任务的序列级分类器；
问答集数据集SQUAD上的令牌级分类器；
SWAG分类语料库中的序列级多选分类器；
另一个目标语料库上的BERT语言模型。
我们这里仅展示GLUE的结果：
这里是使用uncased BERT基础模型在GLUE基准测试开发集上得到的结果。所有实验均在批量大小为32的P100 GPU上运行。尽管比较原始，但结果看起来还不错。
安装该项目是在Python 2.7和3.5+上测试（例子只在python 3.5+上测试）和PyTorch 0.4.1到1.1.0测试pip 安装：
pip install pytorch-transformers
测试：python -m pytest -sv ./pytorch_transformers/tests/python -m pytest -sv ./examples/传送门：
源码：
文档：
