Skip to content

Repository files navigation

albert_zh_pytorch

This repository contains a PyTorch implementation of the albert model from the paper

A Lite Bert For Self-Supervised Learning Language Representations

by Zhenzhong Lan. Mingda Chen....

arxiv: https://arxiv.org/pdf/1909.11942.pdf

Pre-LN and Post-LN

  • Post-LN: . 在原始的Transformer中,Layer Norm在跟在Residual之后的,我们把这个称为Post-LN Transformer

  • Pre-LN: 把Layer Norm换个位置,比如放在Residual的过程之中(称为Pre-LN Transformer

paper: On Layer Normalization in the Transformer Architecture

使用方式

按照]brightmart大佬提供的模型权重文件,需要在配置文件中添加ln_type参数,如下:

{
"attention_probs_dropout_prob": 0.0,
"directionality": "bidi", "hidden_act": "gelu", "hidden_dropout_prob": 0.0,
"hidden_size": 768,
"embedding_size": 128,
"initializer_range": 0.02, "intermediate_size": 3072 ,
"max_position_embeddings": 512, "num_attention_heads": 12,
"num_hidden_layers": 12,
"pooler_fc_size": 768,
"pooler_num_attention_heads": 12,
"pooler_num_fc_layers": 3, "pooler_size_per_head": 128, "pooler_type": "first_token_transform", "type_vocab_size": 2, "vocab_size": 21128,
"ln_type":"postln"# postln or preln
}

show type

Cross-Layer Parameter Sharing: ALBERT use cross-layer parameter sharing in Attention and FFN(FeedForward Network) to reduce number of parameter.

modify the share_type parameter:

  • all: attention和FFN层参数都共享
  • ffn: 只共享FFN层参数
  • attention: 只共享attention层参数
  • None: 无参数共享

使用方式

在加载config时,指定share_type参数,如下:

config=AlbertConfig.from_pretrained(bert_config_file,share_type=share_type)

Download Pre-trained Models of Chinese

感谢brightmart大佬提供中文模型权重:github

  1. albert_large_zh 参数量,层数24,大小为64M

  2. albert_base_zh(小模型体验版), 参数量12M, 层数12,大小为40M

  3. albert_xlarge_zh 参数量,层数24,文件大小为230M

预训练

n-gram: 原始论文中按照以下分布随机生成n-gram,默认max_n为3

1.将文本数据转化为一行一句格式,并且不同document之间使用`\n`分割

2.运行python prepare_lm_data_ngram.py --do_data分别生成ngram mask格式数据集

3.运行python run_pretraining.py --share_type=all进行模型预训练

** 模型大小**

以下是对bert-base进行实验的结果

embedding_sizeshare_typemodel_size
768None476.5M
768attention372.4M
768ffn268.6M
768all164.6M
128None369.1M
128attention265.1M
128ffn161.2M
128all57.2M

下游任务Fine-tuning

1.下载预训练的albert模型,例如下载albert_large_zh.zip,解压到 ~/tmp文件夹下:

$ tree ~/tmp/
/home/dell/tmp/
└── albert_large_zh
├── albert_config_large.json
├── albert_model.ckpt.data-00000-of-00001
├── albert_model.ckpt.index
├── albert_model.ckpt.meta
├── checkpoint
└── vocab.txt

2.运行python convert_albert_tf_checkpoint_to_pytorch.py将TF模型权重转化为pytorch模型权重(默认情况下shar_type=all)

$python convert_albert_tf_checkpoint_to_pytorch.py \
--tf_checkpoint_path ~/tmp/albert_large_zh/ \
--bert_config_file configs/albert_config_large.json \
--pytorch_dump_path pretrain/pytorch/pytorch_model.bin

请参考 convert.sh.

3.下载对应的数据集,比如LCQMC数据集,包含训练、验证和测试集,训练集包含24万口语化描述的中文句子对,标签为1或0,1为句子语义相似,0为语义不相似,将下载文件解压到dataset/lcqmc/。

$ tree dataset/lcqmc/
dataset/lcqmc/
├── dev.txt
├── __init__.py
├── test.txt
└── train.txt

4.运行python run_classifier.py --do_train进行Fine-tuning训练

python run_classifier.py \
--arch albert_large \
--albert_config_path configs/albert_config_large.json \
--bert_dir pretrain/pytorch/albert_large_zh \
--train_batch_size 24 \
--num_train_epochs 10 \
--do_train 

请参考 train.sh.

5. 运行python run_classifier.py --do_test进行test评估

python run_classifier.py \
--arch albert_large \
--albert_config_path configs/albert_config_large.json \
--bert_dir pretrain/pytorch/albert_large_zh \
--do_test

请参考 test.sh.

结果

问题匹配语任务:LCQMC(Sentence Pair Matching)

模型开发集(Dev)测试集(Test)
ALBERT-zh-base(tf)86.486.3
ALBERT-zh-base(pytorch)87.486.4

About

albert_zh对应的pytorch版本

Resources

Stars

6 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages