微調(diào)BaiChuan13B來(lái)做命名實(shí)體識(shí)別
傳統(tǒng)上，一般把NLP的研究領(lǐng)域大致分為自然語(yǔ)言理解（NLU）和自然語(yǔ)言生成（NLG）兩種。
NLU側(cè)重于如何理解文本，包括文本分類、命名實(shí)體識(shí)別、指代消歧、句法分析、機(jī)器閱讀理解等；
NLG則側(cè)重于理解文本后如何生成自然文本，包括自動(dòng)摘要、機(jī)器翻譯、問(wèn)答系統(tǒng)、對(duì)話機(jī)器人等。
但是以ChatGPT為代表的大模型出來(lái)后，這些傳統(tǒng)的NLP的細(xì)分研究領(lǐng)域基本可以說(shuō)都失去了獨(dú)立研究的價(jià)值。
為什么呢？因?yàn)榇竽Ｐ涂梢杂媒y(tǒng)一的范式通通將它們搞定，并且效果非常出眾。
在之前的例子中，我們演示了使用QLoRA算法來(lái)對(duì)BaiChuan-13B實(shí)施微調(diào)以處理最簡(jiǎn)單的文本分類任務(wù)。
Baichuan-13B 保姆級(jí)微調(diào)范例
在外賣評(píng)論數(shù)據(jù)集上，微調(diào)后測(cè)試集acc由0.8925提升到0.9015約提升了1個(gè)百分點(diǎn)。
在本例中，我們使用幾乎相同的流程和方法來(lái)微調(diào)BaiChuan-13B以更好地處理命名實(shí)體識(shí)別任務(wù)。
實(shí)驗(yàn)結(jié)果顯示，在NER任務(wù)上經(jīng)過(guò)微調(diào)，我們的f1-score取得了不可忽略的提升（0.4313—>0.8768）。
注：跑完本流程需要至少32G的CPU，需要約2個(gè)小時(shí)的訓(xùn)練時(shí)間。
公眾號(hào)算法美食屋后臺(tái)回復(fù)關(guān)鍵詞：torchkeras，獲取本文notebook源碼和dfner_13k.pkl數(shù)據(jù)集下載鏈接~
在我們正式開始之前，請(qǐng)?jiān)试S我用簡(jiǎn)短的話給沒(méi)有NLP基礎(chǔ)知識(shí)的小伙伴講解一下什么是命名實(shí)體識(shí)別。
命名實(shí)體識(shí)別NER任務(wù)是NLP的一個(gè)常見(jiàn)基礎(chǔ)任務(wù)，
它是Named Entity Recognization的簡(jiǎn)稱。
簡(jiǎn)單地說(shuō)，就是識(shí)別一個(gè)句子中的各種 名稱實(shí)體，諸如：人名，地名，機(jī)構(gòu) 等。
例如對(duì)于下面這句話：
小明對(duì)小紅說(shuō):"你聽說(shuō)過(guò)安利嗎？"
其命名實(shí)體可以抽取表示如下：
{"人名": ["小明","小紅"], "組織": ["安利"]}
〇，預(yù)訓(xùn)練模型
我們需要從 https://huggingface.co/baichuan-inc/Baichuan-13B-Chat 下載baichuan-13b-chat的模型。
國(guó)內(nèi)可能速度會(huì)比較慢，總共有25個(gè)G左右，網(wǎng)速不太好的話，大概可能需要兩到三個(gè)小時(shí)。
如果網(wǎng)絡(luò)不穩(wěn)定，也可以手動(dòng)從這個(gè)頁(yè)面一個(gè)一個(gè)下載全部文件然后放置到 一個(gè)文件夾中例如 'baichuan-13b' 以便讀取。
import warnings
warnings.filterwarnings('ignore')
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM,AutoConfig, AutoModel, BitsAndBytesConfig
from transformers.generation.utils import GenerationConfig
import torch.nn as nn
#使用QLoRA引入的 NF4量化數(shù)據(jù)類型以節(jié)約顯存
model_name_or_path ='../baichuan-13b' #遠(yuǎn)程 'baichuan-inc/Baichuan-13B-Chat'
bnb_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
llm_int8_threshold=6.0,
llm_int8_has_fp16_weight=False,
)
tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
quantization_config=bnb_config,
trust_remote_code=True)
model.generation_config = GenerationConfig.from_pretrained(model_name_or_path)
from IPython.display import clear_output
messages = []
messages.append({"role": "user",
"content": "世界上第二高的山峰是哪座?"})
response = model.chat(tokenizer,messages=messages,stream=True)
for res in response:
print(res)
clear_output(wait=True)
下面我們?cè)O(shè)計(jì)一個(gè)7-shot-prompt方法，測(cè)試一下BaiChuan13b的實(shí)體抽取能力。
prefix = '''命名實(shí)體識(shí)別：抽取文本中的 人名，地點(diǎn)，組織 這三類命名實(shí)體，并按照json格式返回結(jié)果。
下面是一些范例：
小明對(duì)小紅說(shuō):"你聽說(shuō)過(guò)安利嗎？" -> {"人名": ["小明","小紅"], "組織": ["安利"]}
現(xiàn)在，每年有幾十萬(wàn)中國(guó)人到美國(guó)訪問(wèn)，幾千名中國(guó)留學(xué)生到美國(guó)就學(xué)。 -> {"地點(diǎn)": ["中國(guó)", "美國(guó)"]}
中國(guó)是聯(lián)合國(guó)安理會(huì)常任理事國(guó)之一。 -> {"地點(diǎn)": ["中國(guó)"], "組織": ["聯(lián)合國(guó)"]}
請(qǐng)對(duì)下述文本進(jìn)行實(shí)體抽取，返回json格式。
'''
def get_prompt(text):
return prefix+text+' -> '
def get_message(prompt,response):
return [{"role": "user", "content": f'{prompt} -> '},
{"role": "assistant", "content": response}]
messages = [{"role": "user", "content": get_prompt("一些摩洛哥球迷已按捺不住，在看臺(tái)上歡呼雀躍")}]
response = model.chat(tokenizer, messages)
print(response)
{"地點(diǎn)":["摩洛哥"], "組織":[]}
messages = messages+[{"role": "assistant", "content": "{'地點(diǎn)': ['摩洛哥']}"}]
messages.extend(get_message("這次輪到北京國(guó)安隊(duì)，不知會(huì)不會(huì)再步后塵？","{'組織': ['北京國(guó)安隊(duì)']}"))
messages.extend(get_message("革命黨人孫中山在澳門成立同盟會(huì)分會(huì)","{'人名': ['孫中山'], '地名': ['澳門'], '組織': ['同盟會(huì)']}"))
messages.extend(get_message("我曾在安徽蕪湖市和上海浦東打工。","{'地點(diǎn)': ['安徽蕪湖市', '上海浦東']}"))
display(messages)
def predict(text,temperature=0.01):
model.generation_config.temperature=temperature
response = model.chat(tokenizer,
messages = messages+[{'role':'user','content':f'{text} -> '}])
return response
predict('杜甫是李白的粉絲。')
"{'人名': ['杜甫', '李白']}"
我們拿一個(gè)開源的中文NER數(shù)據(jù)集來(lái)測(cè)試一下未經(jīng)微調(diào)，僅僅使用7-shot-prompt的預(yù)訓(xùn)練模型的效果。
from sklearn.model_selection import train_test_split
import pandas as pd
df = pd.read_pickle('dfner_13k.pkl')
dfdata,dftest = train_test_split(df,test_size=300,random_state=42)
dftrain,dfval = train_test_split(dfdata,test_size=200,random_state=42)
preds = ['' for x in dftest['target']]
for i in tqdm(range(len(preds))):
preds[i] = predict(dftest['text'].iloc[i])
def toset(s):
try:
dic = eval(str(s))
res = []
for k,v in dic.items():
for x in v:
if x:
res.append((k,x))
return set(res)
except Exception as err:
print(err)
return set()
dftest['pred'] = [toset(x) for x in preds]
dftest['gt'] = [toset(x) for x in dftest['target']]
dftest['tp_cnt'] = [len(pred>) for pred,gt in zip(dftest['pred'],dftest['gt'])]
dftest['pred_cnt'] = [len(x) for x in dftest['pred']]
dftest['gt_cnt'] = [len(x) for x in dftest['gt']]
precision = sum(dftest['tp_cnt'])/sum(dftest['pred_cnt'])
print('precision = '+str(precision))
recall = sum(dftest['tp_cnt'])/sum(dftest['gt_cnt'])
print('recall = '+str(recall))
f1 = 2*precision*recall/(precision+recall)
print('f1_score = '+str(f1))
precision = 0.4316109422492401
recall = 0.45151033386327505
f1_score = 0.44133644133644134
微調(diào)前 f1_score為 0.44.
一，準(zhǔn)備數(shù)據(jù)
我們仿照百川模型的 model._build_chat_input 方法來(lái)進(jìn)行token編碼，同時(shí)把需要學(xué)習(xí)的內(nèi)容添加label.
1，token編碼
import torch
#將messages編碼成 token, 同時(shí)返回labels
#注意baichuan-13b通過(guò)插入tokenizer.user_token_id和tokenizer.assistant_token_id 來(lái)區(qū)分用戶和機(jī)器人會(huì)話內(nèi)容
# reference@ model._build_chat_input?
def build_chat_input(messages, model=model,
tokenizer=tokenizer,
max_new_tokens: int=0):
max_new_tokens = max_new_tokens or model.generation_config.max_new_tokens
max_input_tokens = model.config.model_max_length - max_new_tokens
max_input_tokens = max(model.config.model_max_length // 2, max_input_tokens)
total_input, round_input, total_label, round_label = [], [], [], []
for i, message in enumerate(messages[::-1]):
content_tokens = tokenizer.encode(message['content'])
if message['role'] == 'user':
round_input = [model.generation_config.user_token_id] + content_tokens + round_input
round_label = [-100]+[-100 for _ in content_tokens]+ round_label
if total_input and len(total_input) + len(round_input) > max_input_tokens:
break
else:
total_input = round_input + total_input
total_label = round_label + total_label
if len(total_input) >= max_input_tokens:
break
else:
round_input = []
round_label = []
elif message['role'] == 'assistant':
round_input = [
model.generation_config.assistant_token_id
] + content_tokens + [
model.generation_config.eos_token_id
] + round_input
if i==0: #僅對(duì)最后一輪的target進(jìn)行學(xué)習(xí)
round_label = [
-100
] + content_tokens + [
model.generation_config.eos_token_id
]+ round_label
else:
round_label = [
-100
] + [-100 for _ in content_tokens] + [
-100
]+ round_label
else:
raise ValueError(f"message role not supported yet: {message['role']}")
total_input = total_input[-max_input_tokens:] # truncate left
total_label = total_label[-max_input_tokens:]
total_input.append(model.generation_config.assistant_token_id)
total_label.append(-100)
return total_input,total_label
2，做數(shù)據(jù)集
from torch.utils.data import Dataset,DataLoader
from copy import deepcopy
class MyDataset(Dataset):
def __init__(self,df,
messages
):
self.df = df
self.messages = messages
def __len__(self):
return len(self.df)
def get_samples(self,index):
samples = []
d = dict(self.df.iloc[index])
samples.append(d)
return samples
def get_messages(self,index):
samples = self.get_samples(index)
messages = deepcopy(self.messages)
for i,d in enumerate(samples):
messages.append({'role':'user','content':d['text']+' -> '})
messages.append({'role':'assistant','content':str(d['target'])})
return messages
def __getitem__(self,index):
messages = self.get_messages(index)
input_ids, labels = build_chat_input(messages)
return {'input_ids':input_ids,'labels':labels}
def show_sample(self,index):
samples = self.get_samples(index)
print(samples)
ds_train = MyDataset(dftrain,messages)
ds_val = MyDataset(dfval,messages)
3，創(chuàng)建管道
def data_collator(examples: list):
len_ids = [len(example["input_ids"]) for example in examples]
longest = max(len_ids) #之后按照batch中最長(zhǎng)的input_ids進(jìn)行padding
input_ids = []
labels_list = []
for length, example in sorted(zip(len_ids, examples), key=lambda x: -x[0]):
ids = example["input_ids"]
labs = example["labels"]
ids = ids + [tokenizer.pad_token_id] * (longest - length)
labs = labs + [-100] * (longest - length)
input_ids.append(torch.LongTensor(ids))
labels_list.append(torch.LongTensor(labs))
input_ids = torch.stack(input_ids)
labels = torch.stack(labels_list)
return {
"input_ids": input_ids,
"labels": labels,
}
import torch
dl_train = torch.utils.data.DataLoader(ds_train,num_workers=2,batch_size=1,
pin_memory=True,shuffle=True,
collate_fn = data_collator)
dl_val = torch.utils.data.DataLoader(ds_val,num_workers=2,batch_size=1,
pin_memory=True,shuffle=False,
collate_fn = data_collator)
for batch in dl_train:
break
#試跑一個(gè)batch
out = model(**batch)
out.loss
#采樣300個(gè)batch作為一個(gè)epoch，便于較快驗(yàn)證
dl_train.size = 300
二，定義模型
下面我們將使用QLoRA(實(shí)際上用的是量化的AdaLoRA）算法來(lái)微調(diào)Baichuan-13b模型。
from peft import get_peft_config, get_peft_model, TaskType
model.supports_gradient_checkpointing = True #
model.gradient_checkpointing_enable()
model.enable_input_require_grads()
model.config.use_cache = False # silence the warnings. Please re-enable for inference!
import bitsandbytes as bnb
def find_all_linear_names(model):
"""
找出所有全連接層，為所有全連接添加adapter
"""
cls = bnb.nn.Linear4bit
lora_module_names = set()
for name, module in model.named_modules():
if isinstance(module, cls):
names = name.split('.')
lora_module_names.add(names[0] if len(names) == 1 else names[-1])
if 'lm_head' in lora_module_names: # needed for 16-bit
lora_module_names.remove('lm_head')
return list(lora_module_names)
from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model)
lora_modules = find_all_linear_names(model)
print(lora_modules)
['down_proj', 'gate_proj', 'W_pack', 'o_proj', 'up_proj']
from peft import AdaLoraConfig
peft_config = AdaLoraConfig(
task_type=TaskType.CAUSAL_LM, inference_mode=False,
r=16,
lora_alpha=16, lora_dropout=0.05,
target_modules= lora_modules
)
peft_model = get_peft_model(model, peft_config)
peft_model.is_parallelizable = True
peft_model.model_parallel = True
peft_model.print_trainable_parameters()
trainable params: 41,843,040 || all params: 7,002,181,160 || trainable%: 0.5975715144165165
out = peft_model.forward(**batch)
out[0]
三，訓(xùn)練模型
from torchkeras import KerasModel
from accelerate import Accelerator
class StepRunner:
def __init__(self, net, loss_fn, accelerator=None, stage = "train", metrics_dict = None,
optimizer = None, lr_scheduler = None
):
self.net,self.loss_fn,self.metrics_dict,self.stage = net,loss_fn,metrics_dict,stage
self.optimizer,self.lr_scheduler = optimizer,lr_scheduler
self.accelerator = accelerator if accelerator is not None else Accelerator()
if self.stage=='train':
self.net.train()
else:
self.net.eval()
def __call__(self, batch):
#loss
with self.accelerator.autocast():
loss = self.net.forward(**batch)[0]
#backward()
if self.optimizer is not None and self.stage=="train":
self.accelerator.backward(loss)
if self.accelerator.sync_gradients:
self.accelerator.clip_grad_norm_(self.net.parameters(), 1.0)
self.optimizer.step()
if self.lr_scheduler is not None:
self.lr_scheduler.step()
self.optimizer.zero_grad()
all_loss = self.accelerator.gather(loss).sum()
#losses (or plain metrics that can be averaged)
step_losses = {self.stage+"_loss":all_loss.item()}
#metrics (stateful metrics)
step_metrics = {}
if self.stage=="train":
if self.optimizer is not None:
step_metrics['lr'] = self.optimizer.state_dict()['param_groups'][0]['lr']
else:
step_metrics['lr'] = 0.0
return step_losses,step_metrics
KerasModel.StepRunner = StepRunner
#僅僅保存QLora可訓(xùn)練參數(shù)
def save_ckpt(self, ckpt_path='checkpoint', accelerator = None):
unwrap_net = accelerator.unwrap_model(self.net)
unwrap_net.save_pretrained(ckpt_path)
def load_ckpt(self, ckpt_path='checkpoint'):
import os
self.net.load_state_dict(
torch.load(os.path.join(ckpt_path,'adapter_model.bin')),strict =False)
self.from_scratch = False
KerasModel.save_ckpt = save_ckpt
KerasModel.load_ckpt = load_ckpt
optimizer = bnb.optim.adamw.AdamW(peft_model.parameters(),
lr=6e-05,is_paged=True) #'paged_adamw'
keras_model = KerasModel(peft_model,loss_fn =None,
optimizer=optimizer)
ckpt_path = 'baichuan13b_ner'
# keras_model.load_ckpt(ckpt_path) #支持加載微調(diào)后的權(quán)重繼續(xù)訓(xùn)練(斷點(diǎn)續(xù)訓(xùn))
keras_model.fit(train_data = dl_train,
val_data = dl_val,
epochs=100,patience=10,
monitor='val_loss',mode='min',
ckpt_path = ckpt_path
)
四，保存模型
為減少GPU壓力，此處可重啟kernel釋放顯存
import warnings
warnings.filterwarnings('ignore')
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM,AutoConfig, AutoModel, BitsAndBytesConfig
from transformers.generation.utils import GenerationConfig
import torch.nn as nn
model_name_or_path ='../baichuan-13b'
ckpt_path = 'baichuan13b_ner'
tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path,
trust_remote_code=True
)
model_old = AutoModelForCausalLM.from_pretrained(
model_name_or_path,
trust_remote_code=True,
low_cpu_mem_usage=True,
torch_dtype=torch.float16,
device_map='auto'
)
from peft import PeftModel
#可能需要5分鐘左右
peft_model = PeftModel.from_pretrained(model_old, ckpt_path)
model_new = peft_model.merge_and_unload()
from transformers.generation.utils import GenerationConfig
model_new.generation_config = GenerationConfig.from_pretrained(model_name_or_path)
from IPython.display import clear_output
messages = []
messages.append({"role": "user",
"content": "世界上第二高的山峰是什么？"})
response = model_new.chat(tokenizer,messages=messages,stream=True)
for res in response:
print(res)
clear_output(wait=True)
喬戈里峰。世界第二高峰———喬戈里峰西方登山者稱其為k2峰，海拔高度是8611米，位于喀喇昆侖山脈的中巴邊境上.
save_path = 'baichuan-13b-ner'
tokenizer.save_pretrained(save_path)
model_new.save_pretrained(save_path)
!cp ../baichuan-13b/*.py baichuan-13b-ner
五，使用模型
為減少GPU壓力，此處可再次重啟kernel釋放顯存。
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM,AutoConfig, BitsAndBytesConfig
from transformers.generation.utils import GenerationConfig
import torch.nn as nn
import warnings
warnings.filterwarnings('ignore')
model_name_or_path = 'baichuan-13b-ner'
...
...
我們測(cè)試一下微調(diào)后的效果。
import pandas as pd
import numpy as np
import datasets
from tqdm import tqdm
from sklearn.model_selection import train_test_split
import pandas as pd
df = pd.read_pickle('dfner_13k.pkl')
dfdata,dftest = train_test_split(df,test_size=300,random_state=42)
dftrain,dfval = train_test_split(dfdata,test_size=200,random_state=42)
...
...
...
precision = sum(dftest['tp_cnt'])/sum(dftest['pred_cnt'])
print('precision = '+str(precision))
recall = sum(dftest['tp_cnt'])/sum(dftest['gt_cnt'])
print('recall = '+str(recall))
f1 = 2*precision*recall/(precision+recall)
print('f1_score = '+str(f1))
precision = 0.9139280125195618
recall = 0.8427128427128427
f1_score = 0.876876876876877
微調(diào)后的f1_score為0.8768，相比微調(diào)前的f1_score=0.44，取得了不可忽視的巨大提升。
公眾號(hào)算法美食屋臺(tái)回復(fù)關(guān)鍵詞：torchkeras，獲取本文notebook源碼和更多有趣范例~
評(píng)論
圖片
表情