一.BERT微調

1.介紹

自然語言推斷是一個序列級別的文本對分類問題，而微調BERT只需要一個額外的基于多層感知機的架構對預訓練好的BERT權重參數進行微調，如下圖所示。下面將下載一個預訓練好的小版本的BERT，然后對其進行微調，以便在SNLI數據集上進行自然語言推斷。

2.加載預訓練的BERT

在前面博客BERT預訓練第二篇：李沐動手學深度學習V2-bert預訓練數據集和代碼實現和 BERT預訓練第三篇：李沐動手學深度學習V2-BERT預訓練和代碼實現介紹了預訓練的BERT（注意原始的BERT模型是在更大的語料庫上預訓練的，原始的BERT模型有數以億計的參數）。在下面提供了兩個版本的預訓練的BERT：“bert.base”與原始的BERT基礎模型一樣大，需要大量的計算資源才能進行微調，而“bert.small”是一個小版本，以便于演示。

import os import torch from torch import nn import d2l.torch import json import multiprocessing d2l.torch.DATA_HUB['bert.base'] = (d2l.torch.DATA_URL + 'bert.base.torch.zip','225d66f04cae318b841a13d32af3acc165f253ac') d2l.torch.DATA_HUB['bert.small'] = (d2l.torch.DATA_URL + 'bert.small.torch.zip','c72329e68a732bef0452e4b96a1c341c8910f81f')

兩個預訓練好的BERT模型都包含一個定義詞表的“vocab.json”文件和一個預訓練BERT參數的“pretrained.params”文件，load_pretrained_model函數用于加載預先訓練好的BERT參數。

def load_pretrained_model(pretrained_model,num_hiddens,ffn_num_hiddens,num_heads,num_layers,dropout,max_len,devices):data_dir = d2l.torch.download_extract(pretrained_model)# 定義空詞表以加載預定義詞表vocab = d2l.torch.Vocab()vocab.idx_to_token = json.load(open(os.path.join(data_dir,'vocab.json')))vocab.token_to_idx = {token:idx for idx,token in enumerate(vocab.idx_to_token)}bert = d2l.torch.BERTModel(len(vocab),num_hiddens=num_hiddens,norm_shape=[256],ffn_num_input=256,ffn_num_hiddens=ffn_num_hiddens,num_heads=num_heads,num_layers=num_layers,dropout=dropout,max_len=max_len,key_size=256,query_size=256,value_size=256,hid_in_features=256,mlm_in_features=256,nsp_in_features=256)# bert = nn.DataParallel(bert,device_ids=devices).to(devices[0])# bert.module.load_state_dict(torch.load(os.path.join(data_dir,'pretrained.params')),strict=False)# 加載預訓練BERT參數bert.load_state_dict(torch.load(os.path.join(data_dir,'pretrained.params')))return bert,vocab

為了便于在大多數機器上演示，下面加載和微調經過預訓練BERT的小版本（“bert.mall”）。

devices = d2l.torch.try_all_gpus()[2:4] bert,vocab = load_pretrained_model('bert.small',num_hiddens=256,ffn_num_hiddens=512,num_heads=4,num_layers=2,dropout=0.1,max_len=512,devices=devices)

3. 微調BERT的數據集

對于SNLI數據集的下游任務自然語言推斷，定義一個定制的數據集類SNLIBERTDataset。在每個樣本中，前提和假設形成一對文本序列，并被打包成一個BERT輸入序列，片段索引用于區分BERT輸入序列中的前提和假設。利用預定義的BERT輸入序列的最大長度（max_len），持續移除輸入文本對中較長文本的最后一個標記，直到滿足max_len。為了加速生成用于微調BERT的SNLI數據集，使用4個工作進程并行生成訓練或測試樣本。

class SNLIBERTDataset(torch.utils.data.Dataset):def __init__(self,dataset,max_len,vocab=None):all_premises_hypotheses_tokens = [[p_tokens,h_tokens] for p_tokens,h_tokens in zip(*[d2l.torch.tokenize([s.lower() for s in sentences]) for sentences in dataset[:2]])]self.vocab = vocabself.max_len = max_lenself.labels = torch.tensor(dataset[2])self.all_tokens_id,self.all_segments,self.all_valid_lens = self._preprocess(all_premises_hypotheses_tokens)print(f'read {len(self.all_tokens_id)} examples')def _preprocess(self,all_premises_hypotheses_tokens):pool = multiprocessing.Pool(4)# 使用4個進程out = pool.map(self._mp_worker,all_premises_hypotheses_tokens)all_tokens_id = [tokens_id for tokens_id,segments,valid_len in out]all_segments = [segments for tokens_id,segments,valid_len in out]all_valid_lens = [valid_len for tokens_id,segments,valid_len in out]return torch.tensor(all_tokens_id,dtype=torch.long),torch.tensor(all_segments,dtype=torch.long),torch.tensor(all_valid_lens)def _mp_worker(self,premises_hypotheses_tokens):p_tokens,h_tokens = premises_hypotheses_tokensself._truncate_pair_of_tokens(p_tokens,h_tokens)tokens,segments = d2l.torch.get_tokens_and_segments(p_tokens,h_tokens)valid_len = len(tokens)tokens_id = self.vocab[tokens]+[self.vocab['<pad>']]*(self.max_len-valid_len)segments = segments+[0]*(self.max_len-valid_len)return (tokens_id,segments,valid_len)def _truncate_pair_of_tokens(self,p_tokens,h_tokens):# 為BERT輸入中的'<CLS>'、'<SEP>'和'<SEP>'詞元保留位置while (len(p_tokens)+len(h_tokens))>self.max_len-3:if len(p_tokens)>len(h_tokens):p_tokens.pop()else:h_tokens.pop()def __getitem__(self, idx):return (self.all_tokens_id[idx],self.all_segments[idx],self.all_valid_lens[idx]),self.labels[idx]def __len__(self):return len(self.all_tokens_id)

下載完SNLI數據集后，通過實例化SNLIBERTDataset類來生成訓練和測試樣本，這些樣本將在自然語言推斷的訓練和測試期間進行小批量讀取。

#在原始的BERT模型中，max_len=512 batch_size,max_len,num_workers = 512,128,d2l.torch.get_dataloader_workers() data_dir = d2l.torch.download_extract('SNLI') train_set = SNLIBERTDataset(d2l.torch.read_snli(data_dir,is_train=True),max_len,vocab) test_set = SNLIBERTDataset(d2l.torch.read_snli(data_dir,is_train=False),max_len,vocab) train_iter = torch.utils.data.DataLoader(train_set,batch_size,num_workers=num_workers,shuffle=True) test_iter = torch.utils.data.DataLoader(test_set,batch_size,num_workers=num_workers,shuffle=False)

4. BERT微調

**用于自然語言推斷的微調BERT只需要一個額外的多層感知機，該多層感知機由兩個全連接層組成，**與前面BERT實現的博客BERT預訓練第一篇：李沐動手學深度學習V2-bert和代碼實現中BERTClassifier類中進行nsp預測的self.hidden和self.output的多層感知機結構一個。這個多層感知機將特殊的“”詞元的BERT表示進行了轉換，該詞元同時編碼前提和假設的信息，經過多層感知機后得到自然語言推斷的輸出分類特征維：蘊涵、矛盾和中性。

class BERTClassifier(nn.Module):def __init__(self,bert):super(BERTClassifier,self).__init__()self.encoder = bert.encoderself.hidden = bert.hiddenself.output = nn.Linear(256,3)def forward(self,inputs):tokens_X,segments_X,valid_lens_X = inputsencoded_X = self.encoder(tokens_X,segments_X,valid_lens_X)return self.output(self.hidden(encoded_X[:,0,:]))

下面將預訓練的BERT模型bert被送到用于下游應用的BERTClassifier實例net中。在BERT微調的常見實現中，只有額外的多層感知機（net.output）的輸出層的參數將從零開始學習。預訓練BERT編碼器（net.encoder）和額外的多層感知機的隱藏層（net.hidden）的所有參數都將進行微調。

net = BERTClassifier(bert)

在BERT預訓練中MaskLM類和NextSentencePred類在其使用的多層感知機中都有一些參數，這些參數是預訓練BERT模型bert中參數的一部分，然而這些參數僅用于計算預訓練過程中的遮蔽語言模型損失和下一句預測損失。這兩個損失函數與微調下游應用無關，因此當BERT微調時，MaskLM和NextSentencePred中采用的多層感知機的參數不會更新（陳舊的，staled）。
通過d2l.train_batch_ch13（）函數使用SNLI的訓練集（train_iter）和測試集（test_iter）對net模型進行訓練和評估，結果如下圖所示。

lr,num_epochs = 1e-4,5 optim = torch.optim.Adam(params=net.parameters(),lr=lr) loss = nn.CrossEntropyLoss(reduction='none') d2l.torch.train_ch13(net,train_iter,test_iter,loss,optim,num_epochs,devices)

5. 小結

• 針對下游應用對預訓練的BERT模型進行微調，例如在SNLI數據集上進行自然語言推斷。
• 在微調過程中，BERT模型成為下游應用模型的一部分，再加上多層感知機進行下游應用模型任務的訓練和評估。

6. 使用原始BERT的預訓練模型進行微調

微調一個更大的預訓練BERT模型，該模型與原始的BERT基礎模型一樣大。修改load_pretrained_model函數中的參數設置：將“bert.mall”替換為“bert.base”，將num_hiddens=256、ffn_num_hiddens=512、num_heads=4和num_layers=2的值分別增加到768、3072、12和12，同時修改多層感知機輸出層的Linear層為(nn.Linear(768,3)，因為現在經過BERT模型輸出特征維變為768)，增加微調迭代輪數，代碼如下所示。

import os import torch from torch import nn import d2l.torch import json import multiprocessingd2l.torch.DATA_HUB['bert.base'] = (d2l.torch.DATA_URL + 'bert.base.torch.zip','225d66f04cae318b841a13d32af3acc165f253ac') d2l.torch.DATA_HUB['bert.small'] = (d2l.torch.DATA_URL + 'bert.small.torch.zip','c72329e68a732bef0452e4b96a1c341c8910f81f')devices = d2l.torch.try_all_gpus() def load_pretrained_model1(pretrained_model,num_hiddens,ffn_num_hiddens,num_heads,num_layers,dropout,max_len,devices):data_dir = d2l.torch.download_extract(pretrained_model)vocab = d2l.torch.Vocab()vocab.idx_to_token = json.load(open(os.path.join(data_dir,'vocab.json')))vocab.token_to_idx = {token:idx for idx,token in enumerate(vocab.idx_to_token)}bert = d2l.torch.BERTModel(len(vocab),num_hiddens=num_hiddens,norm_shape=[768],ffn_num_input=768,ffn_num_hiddens=ffn_num_hiddens,num_heads=num_heads,num_layers=num_layers,dropout=dropout,max_len=max_len,key_size=768,query_size=768,value_size=768,hid_in_features=768,mlm_in_features=768,nsp_in_features=768)# bert = nn.DataParallel(bert,device_ids=devices).to(devices[0])# bert.module.load_state_dict(torch.load(os.path.join(data_dir,'pretrained.params')),strict=False)bert.load_state_dict(torch.load(os.path.join(data_dir,'pretrained.params')))return bert,vocabbert,vocab = load_pretrained_model1('bert.base',num_hiddens=768,ffn_num_hiddens=3072,num_heads=12,num_layers=12,dropout=0.1,max_len=512,devices=devices) class SNLIBERTDataset(torch.utils.data.Dataset):def __init__(self, dataset, max_len, vocab=None):all_premises_hypotheses_tokens = [[p_tokens, h_tokens] for p_tokens, h_tokens inzip(*[d2l.torch.tokenize([s.lower() for s in sentences]) for sentences indataset[:2]])]self.vocab = vocabself.max_len = max_lenself.labels = torch.tensor(dataset[2])self.all_tokens_id, self.all_segments, self.all_valid_lens = self._preprocess(all_premises_hypotheses_tokens)print(f'read {len(self.all_tokens_id)} examples')def _preprocess(self, all_premises_hypotheses_tokens):pool = multiprocessing.Pool(4) # 使用4個進程out = pool.map(self._mp_worker, all_premises_hypotheses_tokens)all_tokens_id = [tokens_id for tokens_id, segments, valid_len in out]all_segments = [segments for tokens_id, segments, valid_len in out]all_valid_lens = [valid_len for tokens_id, segments, valid_len in out]return torch.tensor(all_tokens_id, dtype=torch.long), torch.tensor(all_segments,dtype=torch.long), torch.tensor(all_valid_lens)def _mp_worker(self, premises_hypotheses_tokens):p_tokens, h_tokens = premises_hypotheses_tokensself._truncate_pair_of_tokens(p_tokens, h_tokens)tokens, segments = d2l.torch.get_tokens_and_segments(p_tokens, h_tokens)valid_len = len(tokens)tokens_id = self.vocab[tokens] + [self.vocab['<pad>']] * (self.max_len - valid_len)segments = segments + [0] * (self.max_len - valid_len)return (tokens_id, segments, valid_len)def _truncate_pair_of_tokens(self, p_tokens, h_tokens):# 為BERT輸入中的'<CLS>'、'<SEP>'和'<SEP>'詞元保留位置while (len(p_tokens) + len(h_tokens)) > self.max_len - 3:if len(p_tokens) > len(h_tokens):p_tokens.pop()else:h_tokens.pop()def __getitem__(self, idx):return (self.all_tokens_id[idx], self.all_segments[idx], self.all_valid_lens[idx]), self.labels[idx]def __len__(self):return len(self.all_tokens_id)#在原始的BERT模型中，max_len=512 batch_size, max_len, num_workers = 512, 128, d2l.torch.get_dataloader_workers() data_dir = d2l.torch.download_extract('SNLI') train_set = SNLIBERTDataset(d2l.torch.read_snli(data_dir, is_train=True), max_len, vocab) test_set = SNLIBERTDataset(d2l.torch.read_snli(data_dir, is_train=False), max_len, vocab) train_iter = torch.utils.data.DataLoader(train_set, batch_size, num_workers=num_workers, shuffle=True) test_iter = torch.utils.data.DataLoader(test_set, batch_size, num_workers=num_workers, shuffle=False)class BERTClassifier(nn.Module):def __init__(self, bert):super(BERTClassifier, self).__init__()self.encoder = bert.encoderself.hidden = bert.hiddenself.output = nn.Linear(768, 3)def forward(self, inputs):tokens_X, segments_X, valid_lens_X = inputsencoded_X = self.encoder(tokens_X, segments_X, valid_lens_X)return self.output(self.hidden(encoded_X[:, 0, :]))net = BERTClassifier(bert) lr, num_epochs = 1e-4, 20 optim = torch.optim.Adam(params=net.parameters(), lr=lr) loss = nn.CrossEntropyLoss(reduction='none') d2l.torch.train_ch13(net, train_iter, test_iter, loss, optim, num_epochs, devices)

7. 全部代碼

import os import torch from torch import nn import d2l.torch import json import multiprocessingd2l.torch.DATA_HUB['bert.base'] = (d2l.torch.DATA_URL + 'bert.base.torch.zip','225d66f04cae318b841a13d32af3acc165f253ac') d2l.torch.DATA_HUB['bert.small'] = (d2l.torch.DATA_URL + 'bert.small.torch.zip','c72329e68a732bef0452e4b96a1c341c8910f81f')def load_pretrained_model(pretrained_model, num_hiddens, ffn_num_hiddens, num_heads, num_layers, dropout, max_len,devices):data_dir = d2l.torch.download_extract(pretrained_model)# 定義空詞表以加載預定義詞表vocab = d2l.torch.Vocab()vocab.idx_to_token = json.load(open(os.path.join(data_dir, 'vocab.json')))vocab.token_to_idx = {token: idx for idx, token in enumerate(vocab.idx_to_token)}bert = d2l.torch.BERTModel(len(vocab), num_hiddens=num_hiddens, norm_shape=[256], ffn_num_input=256,ffn_num_hiddens=ffn_num_hiddens, num_heads=num_heads, num_layers=num_layers,dropout=dropout, max_len=max_len, key_size=256, query_size=256, value_size=256,hid_in_features=256, mlm_in_features=256, nsp_in_features=256)# bert = nn.DataParallel(bert,device_ids=devices).to(devices[0])# bert.module.load_state_dict(torch.load(os.path.join(data_dir,'pretrained.params')),strict=False)# 加載預訓練BERT參數bert.load_state_dict(torch.load(os.path.join(data_dir, 'pretrained.params')))return bert, vocabdevices = d2l.torch.try_all_gpus()[2:4] bert, vocab = load_pretrained_model('bert.small', num_hiddens=256, ffn_num_hiddens=512, num_heads=4, num_layers=2,dropout=0.1, max_len=512, devices=devices)class SNLIBERTDataset(torch.utils.data.Dataset):def __init__(self, dataset, max_len, vocab=None):all_premises_hypotheses_tokens = [[p_tokens, h_tokens] for p_tokens, h_tokens inzip(*[d2l.torch.tokenize([s.lower() for s in sentences]) for sentences indataset[:2]])]self.vocab = vocabself.max_len = max_lenself.labels = torch.tensor(dataset[2])self.all_tokens_id, self.all_segments, self.all_valid_lens = self._preprocess(all_premises_hypotheses_tokens)print(f'read {len(self.all_tokens_id)} examples')def _preprocess(self, all_premises_hypotheses_tokens):pool = multiprocessing.Pool(4) # 使用4個進程out = pool.map(self._mp_worker, all_premises_hypotheses_tokens)all_tokens_id = [tokens_id for tokens_id, segments, valid_len in out]all_segments = [segments for tokens_id, segments, valid_len in out]all_valid_lens = [valid_len for tokens_id, segments, valid_len in out]return torch.tensor(all_tokens_id, dtype=torch.long), torch.tensor(all_segments,dtype=torch.long), torch.tensor(all_valid_lens)def _mp_worker(self, premises_hypotheses_tokens):p_tokens, h_tokens = premises_hypotheses_tokensself._truncate_pair_of_tokens(p_tokens, h_tokens)tokens, segments = d2l.torch.get_tokens_and_segments(p_tokens, h_tokens)valid_len = len(tokens)tokens_id = self.vocab[tokens] + [self.vocab['<pad>']] * (self.max_len - valid_len)segments = segments + [0] * (self.max_len - valid_len)return (tokens_id, segments, valid_len)def _truncate_pair_of_tokens(self, p_tokens, h_tokens):# 為BERT輸入中的'<CLS>'、'<SEP>'和'<SEP>'詞元保留位置while (len(p_tokens) + len(h_tokens)) > self.max_len - 3:if len(p_tokens) > len(h_tokens):p_tokens.pop()else:h_tokens.pop()def __getitem__(self, idx):return (self.all_tokens_id[idx], self.all_segments[idx], self.all_valid_lens[idx]), self.labels[idx]def __len__(self):return len(self.all_tokens_id)#在原始的BERT模型中，max_len=512 batch_size, max_len, num_workers = 512, 128, d2l.torch.get_dataloader_workers() data_dir = d2l.torch.download_extract('SNLI') train_set = SNLIBERTDataset(d2l.torch.read_snli(data_dir, is_train=True), max_len, vocab) test_set = SNLIBERTDataset(d2l.torch.read_snli(data_dir, is_train=False), max_len, vocab) train_iter = torch.utils.data.DataLoader(train_set, batch_size, num_workers=num_workers, shuffle=True) test_iter = torch.utils.data.DataLoader(test_set, batch_size, num_workers=num_workers, shuffle=False)class BERTClassifier(nn.Module):def __init__(self, bert):super(BERTClassifier, self).__init__()self.encoder = bert.encoderself.hidden = bert.hiddenself.output = nn.Linear(256, 3)def forward(self, inputs):tokens_X, segments_X, valid_lens_X = inputsencoded_X = self.encoder(tokens_X, segments_X, valid_lens_X)return self.output(self.hidden(encoded_X[:, 0, :]))net = BERTClassifier(bert) lr, num_epochs = 1e-4, 5 optim = torch.optim.Adam(params=net.parameters(), lr=lr) loss = nn.CrossEntropyLoss(reduction='none') d2l.torch.train_ch13(net, train_iter, test_iter, loss, optim, num_epochs, devices)

8. 相關鏈接

BERT預訓練第一篇：李沐動手學深度學習V2-bert和代碼實現
BERT預訓練第二篇：李沐動手學深度學習V2-bert預訓練數據集和代碼實現
BERT預訓練第三篇：李沐動手學深度學習V2-BERT預訓練和代碼實現
BERT微調第一篇：李沐動手學深度學習V2-自然語言推斷與數據集SNLI和代碼實現
BERT微調第二篇：李沐動手學深度學習V2-BERT微調和代碼實現

總結

以上是生活随笔為你收集整理的李沐动手学深度学习V2-BERT微调和代码实现的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。