shuxinyin/ner-pytorch-chinese

Implemention of NER model on chinese dataset.

ner-pytorch-chinese's Issues

请问一下，crf学习率是bert的100倍，在哪儿体现的？谢谢

processors模块问题

from processors.processor import LEBertProcessor, BertProcessor
ModuleNotFoundError: No module named 'processors.processor'
这是安装提示processors in /share/home/.local/lib/python3.7/site-packages (0.0.2)

processors模块应该是processors这个文件夹里的processor.py文件吧 怎么也找不到模块 难受，我直接把processor.py文件拖到了报错的model_class下面 但是我不知道是哪里的路径没改对 他才会找不到processor文件夹呢

这是我的配置，bert_chinese，tencent-ailab-embedding-zh-d200-v0.2.0-s.txt是下载的文件，不知道是不是还缺了要改的地方，运行不起来

• DEVICE=0
• LR=1e-5
• CRF_LR=1e-3
• ADAPTER_LR=1e-3
• REPO_PATH=/share/home/NER-Pytorch-Chinese-master
• DATA_DIR=/share/home/NER-Pytorch-Chinese-master/data
• OUTPUT_PATH=/share/home/NER-Pytorch-Chinese-master/output
• PRETRAIN_MODEL=/data/bert_chinese
• PRETRAIN_EMBED_PATH=/data/tencent-ailab-embedding-zh-d200-v0.2.0-s.txt
• for DATA_SET in ''''msra''''
• for MODEL_CLASS in ''''lebert-softmax''''
• echo ----------------------------------------msra:lebert-softmax----------------------------------------
• python /share/home//NER-Pytorch-Chinese-master/train/train.py --device gpu --output_path /share/home//NER-Pytorch-Chinese-master/output --add_layer 1 --loss_type lsr --lr 1e-5 --crf_lr 1e-3 --adapter_lr 1e-3 --weight_decay 0.01 --eps 1.0e-08 --epochs 1 --batch_size_train 24 --batch_size_eval 256 --num_workers 0 --eval_step 100 --max_seq_len 150 --max_word_num 3 --max_scan_num 3000000 --data_path /share/home/NER-Pytorch-Chinese-master/data/msra/ --dataset_name msra --model_class lebert-softmax --pretrain_model_path /data/bert_chinese --pretrain_embed_path /data/tencent-ailab-embedding-zh-d200-v0.2.0-s.txt --seed 42 --markup bios --grad_acc_step 1 --max_grad_norm 1.0 --num_workers 0 --warmup_proportion 0.1 --load_word_embed --do_train
  Traceback (most recent call last):
  File "/share/home/NER-Pytorch-Chinese-master/train/train.py", line 21, in
  from model_class import MODEL_CLASS, PROCESSOR_CLASS
  File "/share/home/NER-Pytorch-Chinese-master/train/model_class.py", line 5, in
  from processors.processor import LEBertProcessor, BertProcessor
  ModuleNotFoundError: No module named 'processors.processor'

缺少必要文件

兄弟，代码完善了再上传git呀，有的目录根本没有，程序中还import了~~~~比如：from model_class import MODEL_CLASS, PROCESSOR_CLASS

repo id 错误

您好，我再运行这个train.py文件时，出现以下错误：
huggingface_hub.utils._validators.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': 'NER-Pytorch-Chinese-master/data/Learn_Project/Backup_Data/bert_chinese'. Use repo_type argument if needed.
这个错误是为什么呢？需要怎么处理呀？

CRF 如何进行label smoothing？

请问以CRF为解码层的模型是如何进行label smoothing的？我看代码的时候没有看到相关的代码，请大佬指出！

process.py中的文件找不到

您好，在process.py文件中的init函数下（如下代码所示）
def init(self, pretrain_embed_path, output_path, max_scan_num, data_files, label_path, overwrite):
word_embed_path = join(self.data_path, 'word_embedding.pkl')
word_vocab_path = join(self.data_path, 'word_vocab.pkl')
word_vocab_path_ = join(self.data_path,'word_vocab.pkl)
trie_tree_path = join(self.data_path, 'trie_tree.pkl')
以上self.data_path下中的 'word_embedding.pkl'、'word_vocab.pkl‘、'word_vocab.pkl’、'trie_tree.pkl'都找不到是什么原因呢？

bug？

        if args.model_class in ['lebert-crf', 'bert-crf']:
            preds = model.crf.decode(logits, attention_mask).squeeze(0)
            preds = preds[:, 1:].tolist()  # 减去padding的[CLS]

评估函数172行，是不是少了对bert-lstm-crf的decoder解码

bert-softmax模型的epoch数

请问在msra和ontonotes两个数据集上实验时，script/train.sh中的epoch数分别是多少呢？超参数都一样吗？可不可以提供下具体的配置文件？
还有，预训练的中文版bert是官方的bert-base-chinese吗？

改成crf运行后都会出现

File "/share/home/NER-Pytorch-Chinese-master-1/train/train.py", line 187, in evaluate
metric.update(pred_paths=[pred], label_paths=[label])
File "../metrics/ner_metrics.py", line 56, in update
pre_entities = get_entities(pre_path, self.id2label,self.markup)
File "../utils/get_entity.py", line 117, in get_entities
return get_entity_bios(seq,id2label)
File "../utils/get_entity.py", line 19, in get_entity_bios
tag_str = id2label[tag]
TypeError: list indices must be integers or slices, not list

softmax没有问题，请问怎么回事呢，对了我是把nbest改成了3