使用UER-py库在下游数据集进一步预训练+微调

183****0229

使用UER-py库在下游数据集进一步预训练+微调

环境

RTX3090
pytorch 1.8.1
python 3.8
cuda 11.1

流程

# 切换路径
cd /hy-tmp
# 下载UER并安装其他依赖包
git clone https://hub.fastgit.org/dbiir/UER-py.git
cd UER-py
pip install -U six
pip install transformers

转化huggingface的预训练权重为UER适配的权重

convert.py如下

# convert.py
import torch
import argparse
import collections


def convert_bert_transformer_encoder_from_huggingface_to_uer(input_model, output_model, layers_num):
    for i in range(layers_num):
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.0.weight"] = input_model["bert.encoder.layer." + str(i) + ".attention.self.query.weight"]
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.0.bias"] = input_model["bert.encoder.layer." + str(i) + ".attention.self.query.bias"]
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.1.weight"] = input_model["bert.encoder.layer." + str(i) + ".attention.self.key.weight"]
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.1.bias"] = input_model["bert.encoder.layer." + str(i) + ".attention.self.key.bias"]
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.2.weight"] = input_model["bert.encoder.layer." + str(i) + ".attention.self.value.weight"]
        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers.2.bias"] = input_model["bert.encoder.layer." + str(i) + ".attention.self.value.bias"]
        output_model["encoder.transformer." + str(i) + ".self_attn.final_linear.weight"] = input_model["bert.encoder.layer." + str(i) + ".attention.output.dense.weight"]
        output_model["encoder.transformer." + str(i) + ".self_attn.final_linear.bias"] = input_model["bert.encoder.layer." + str(i) + ".attention.output.dense.bias"]
        output_model["encoder.transformer." + str(i) + ".layer_norm_1.gamma"] = input_model["bert.encoder.layer." + str(i) + ".attention.output.LayerNorm.weight"]
        output_model["encoder.transformer." + str(i) + ".layer_norm_1.beta"] = input_model["bert.encoder.layer." + str(i) + ".attention.output.LayerNorm.bias"]
        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"] = input_model["bert.encoder.layer." + str(i) + ".intermediate.dense.weight"]
        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.bias"] = input_model["bert.encoder.layer." + str(i) + ".intermediate.dense.bias"]
        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"] = input_model["bert.encoder.layer." + str(i) + ".output.dense.weight"]
        output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.bias"] = input_model["bert.encoder.layer." + str(i) + ".output.dense.bias"]
        output_model["encoder.transformer." + str(i) + ".layer_norm_2.gamma"] = input_model["bert.encoder.layer." + str(i) + ".output.LayerNorm.weight"]
        output_model["encoder.transformer." + str(i) + ".layer_norm_2.beta"] = input_model["bert.encoder.layer." + str(i) + ".output.LayerNorm.bias"]


def main():
    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument("--input_model_path", type=str, default="models/bert-base-chinese/pytorch_model.bin",
                        help=".")
    parser.add_argument("--output_model_path", type=str, default="models/google_zh_model.bin",
                        help=".")
    parser.add_argument("--layers_num", type=int, default=12, help=".")
    parser.add_argument("--target", choices=["bert", "mlm"], default="bert",
                        help="The training target of the pretraining model.")

    args = parser.parse_args()
    
    input_model = torch.load(args.input_model_path, map_location='cpu')
    
    output_model = collections.OrderedDict()
    
    output_model["embedding.word_embedding.weight"] = input_model["bert.embeddings.word_embeddings.weight"]
    output_model["embedding.position_embedding.weight"] = input_model["bert.embeddings.position_embeddings.weight"]
    output_model["embedding.segment_embedding.weight"] = torch.cat((torch.Tensor([[0]*input_model["bert.embeddings.token_type_embeddings.weight"].size()[1]]), input_model["bert.embeddings.token_type_embeddings.weight"]), dim=0)
    output_model["embedding.layer_norm.gamma"] = input_model["bert.embeddings.LayerNorm.weight"]
    output_model["embedding.layer_norm.beta"] = input_model["bert.embeddings.LayerNorm.bias"]
    
    convert_bert_transformer_encoder_from_huggingface_to_uer(input_model, output_model, args.layers_num)
    
    if args.target == "bert":
        output_model["target.nsp_linear_1.weight"] = input_model["bert.pooler.dense.weight"]
        output_model["target.nsp_linear_1.bias"] = input_model["bert.pooler.dense.bias"]
        output_model["target.nsp_linear_2.weight"] = input_model["cls.seq_relationship.weight"]
        output_model["target.nsp_linear_2.bias"] = input_model["cls.seq_relationship.bias"]
    output_model["target.mlm_linear_1.weight"] = input_model["cls.predictions.transform.dense.weight"]
    output_model["target.mlm_linear_1.bias"] = input_model["cls.predictions.transform.dense.bias"]
    output_model["target.layer_norm.gamma"] = input_model["cls.predictions.transform.LayerNorm.weight"]
    output_model["target.layer_norm.beta"] = input_model["cls.predictions.transform.LayerNorm.bias"]
    output_model["target.mlm_linear_2.weight"] = input_model["cls.predictions.decoder.weight"]
    output_model["target.mlm_linear_2.bias"] = input_model["cls.predictions.bias"]
    
    torch.save(output_model, args.output_model_path)

if __name__ == "__main__":
    from transformers import BertForPreTraining
    model = BertForPreTraining.from_pretrained("bert-base-chinese")
    model.save_pretrained("models/bert-base-chinese")
    main()

# 创建convert.py并将代码复制进去
touch convert.py
# 转换权重
python convert.py

# 数据预处理
python preprocess.py    --corpus_path corpora/book_review_bert.txt \
                        --vocab_path models/google_zh_vocab.txt \
                        --dataset_path dataset.pt --processes_num 8 \
                        --target bert

# 进一步预训练
python pretrain.py  --dataset_path dataset.pt --vocab_path models/google_zh_vocab.txt \
                    --pretrained_model_path models/google_zh_model.bin \
                    --output_model_path models/book_review_model.bin \
                    --gpu_ranks 0 \
                    --total_steps 5000 --save_checkpoint_steps 1000 --batch_size 16 \
                    --embedding word_pos_seg --encoder transformer --mask fully_visible --target bert
# 修改权重名
mv models/book_review_model.bin-5000 models/book_review_model.bin

# 微调
python finetune/run_classifier.py --pretrained_model_path models/book_review_model.bin \
                    --vocab_path models/google_zh_vocab.txt \
                    --train_path datasets/douban_book_review/train.tsv \
                    --dev_path datasets/douban_book_review/dev.tsv \
                    --test_path datasets/douban_book_review/test.tsv \
                    --epochs_num 3 --batch_size 64 \
                    --embedding word_pos_seg --encoder transformer --mask fully_visible

REFERENCE

https://github.com/dbiir/UER-py/wiki/Quickstart

Alice_恒源云

UER-py（Universal Encoder Representations）是一个用于对通用语料进行预训练并对下游任务进行微调的工具包