“瘦身成功”的ALBERT，能取代BERT吗？

十三 2020-03-22 12:07:29 来源：量子位

十三 发自 凹非寺
量子位 报道 | 公众号 QbitAI

参数比BERT少了80%，性能却提高了。

这就是谷歌去年提出的“瘦身成功版BERT”模型——ALBERT。

这个模型一经发布，就受到了高度关注，二者的对比也成为了热门话题。

而最近，网友Naman Bansal就提出了一个疑问：

是否应该用ALBERT来代替BERT？

能否替代，比比便知。

BERT与ALBERT

BERT模型是大家比较所熟知的。

2018年由谷歌提出，训练的语料库规模非常庞大，包含33亿个词语。

模型的创新点集中在了预训练过程，采用Masked LM和Next Sentence Prediction两种方法，分别捕捉词语和句子级别的表示。

BERT的出现，彻底改变了预训练产生词向量和下游具体NLP任务的关系。

时隔1年后，谷歌又提出ALBERT，也被称作“lite-BERT”，骨干网络和BERT相似，采用的依旧是 Transformer 编码器，激活函数也是GELU。

其最大的成功，就在于参数量比BERT少了80%，同时还取得了更好的结果。

与BERT相比的改进，主要包括嵌入向量参数化的因式分解、跨层参数共享、句间连贯性损失采用SOP，以及移除了dropout。

下图便是BERT和ALBERT，在SQuAD和RACE数据集上的性能测试比较结果。

可以看出，ALBERT性能取得了较好的结果。

如何实现自定义语料库(预训练)ALBERT？

为了进一步了解ALBERT，接下来，将在自定义语料库中实现ALBERT。

所采用的数据集是“用餐点评数据集”，目标就是通过ALBERT模型来识别菜肴的名称。

第一步：下载数据集并准备文件

 1#Downlading all files and data
 2
 3!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/data_toy/dish_name_train.csv
 4!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/data_toy/dish_name_val.csv
 5!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/data_toy/restaurant_review.txt
 6!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/data_toy/restaurant_review_nopunct.txt
 7!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/models_toy/albert_config.json
 8!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/model_checkpoint/finetune_checkpoint
 9!wget https://github.com/LydiaXiaohongLi/Albert_Finetune_with_Pretrain_on_Custom_Corpus/raw/master/model_checkpoint/pretrain_checkpoint
10
11#Creating files and setting up ALBERT
12
13!pip install sentencepiece
14!git clone https://github.com/google-research/ALBERT
15!python ./ALBERT/create_pretraining_data.py --input_file "restaurant_review.txt" --output_file "restaurant_review_train" --vocab_file "vocab.txt" --max_seq_length=64
16!pip install transformers
17!pip install tfrecord

第二步：使用transformer并定义层

 1#Defining Layers for ALBERT
 2
 3from transformers.modeling_albert import AlbertModel, AlbertPreTrainedModel
 4from transformers.configuration_albert import AlbertConfig
 5import torch.nn as nn
 6class AlbertSequenceOrderHead(nn.Module):
 7    def __init__(self, config):
 8        super().__init__()
 9        self.dense = nn.Linear(config.hidden_size, 2)
10        self.bias = nn.Parameter(torch.zeros(2))
11
12    def forward(self, hidden_states):
13        hidden_states = self.dense(hidden_states)
14        prediction_scores = hidden_states + self.bias
15
16        return prediction_scores
17
18from torch.nn import CrossEntropyLoss
19from transformers.modeling_bert import ACT2FN
20class AlbertForPretrain(AlbertPreTrainedModel):
21
22    def __init__(self, config):
23        super().__init__(config)
24
25        self.albert = AlbertModel(config)       
26
27        # For Masked LM
28        # The original huggingface implementation, created new output weights via dense layer
29        # However the original Albert 
30        self.predictions_dense = nn.Linear(config.hidden_size, config.embedding_size)
31        self.predictions_activation = ACT2FN[config.hidden_act]
32        self.predictions_LayerNorm = nn.LayerNorm(config.embedding_size)
33        self.predictions_bias = nn.Parameter(torch.zeros(config.vocab_size)) 
34        self.predictions_decoder = nn.Linear(config.embedding_size, config.vocab_size)
35
36        self.predictions_decoder.weight = self.albert.embeddings.word_embeddings.weight
37
38        # For sequence order prediction
39        self.seq_relationship = AlbertSequenceOrderHead(config)
40
41
42    def forward(
43        self,
44        input_ids=None,
45        attention_mask=None,
46        token_type_ids=None,
47        position_ids=None,
48        head_mask=None,
49        inputs_embeds=None,
50        masked_lm_labels=None,
51        seq_relationship_labels=None,
52    ):
53
54        outputs = self.albert(
55            input_ids,
56            attention_mask=attention_mask,
57            token_type_ids=token_type_ids,
58            position_ids=position_ids,
59            head_mask=head_mask,
60            inputs_embeds=inputs_embeds,
61        )
62
63        loss_fct = CrossEntropyLoss()
64
65        sequence_output = outputs[0]
66
67        sequence_output = self.predictions_dense(sequence_output)
68        sequence_output = self.predictions_activation(sequence_output)
69        sequence_output = self.predictions_LayerNorm(sequence_output)
70        prediction_scores = self.predictions_decoder(sequence_output)
71
72
73        if masked_lm_labels is not None:
74            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size)
75                                      , masked_lm_labels.view(-1))
76
77        pooled_output = outputs[1]
78        seq_relationship_scores = self.seq_relationship(pooled_output)
79        if seq_relationship_labels is not None:  
80            seq_relationship_loss = loss_fct(seq_relationship_scores.view(-1, 2), seq_relationship_labels.view(-1))
81
82        loss = masked_lm_loss + seq_relationship_loss
83
84        return loss

第三步：使用LAMB优化器并微调ALBERT

  1#Using LAMB optimizer
  2#LAMB -  "https://github.com/cybertronai/pytorch-lamb"
  3
  4import torch
  5from torch.optim import Optimizer
  6class Lamb(Optimizer):
  7    r"""Implements Lamb algorithm.
  8    It has been proposed in `Large Batch Optimization for Deep Learning: Training BERT in 76 minutes`_.
  9    Arguments:
 10        params (iterable): iterable of parameters to optimize or dicts defining
 11            parameter groups
 12        lr (float, optional): learning rate (default: 1e-3)
 13        betas (Tuple[float, float], optional): coefficients used for computing
 14            running averages of gradient and its square (default: (0.9, 0.999))
 15        eps (float, optional): term added to the denominator to improve
 16            numerical stability (default: 1e-8)
 17        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
 18        adam (bool, optional): always use trust ratio = 1, which turns this into
 19            Adam. Useful for comparison purposes.
 20    .. _Large Batch Optimization for Deep Learning: Training BERT in 76 minutes:
 21        https://arxiv.org/abs/1904.00962
 22    """
 23
 24    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-6,
 25                 weight_decay=0, adam=False):
 26        if not 0.0 <= lr:
 27            raise ValueError("Invalid learning rate: {}".format(lr))
 28        if not 0.0 <= eps:
 29            raise ValueError("Invalid epsilon value: {}".format(eps))
 30        if not 0.0 <= betas[0] < 1.0:
 31            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
 32        if not 0.0 <= betas[1] < 1.0:
 33            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
 34        defaults = dict(lr=lr, betas=betas, eps=eps,
 35                        weight_decay=weight_decay)
 36        self.adam = adam
 37        super(Lamb, self).__init__(params, defaults)
 38
 39    def step(self, closure=None):
 40        """Performs a single optimization step.
 41        Arguments:
 42            closure (callable, optional): A closure that reevaluates the model
 43                and returns the loss.
 44        """
 45        loss = None
 46        if closure is not None:
 47            loss = closure()
 48
 49        for group in self.param_groups:
 50            for p in group['params']:
 51                if p.grad is None:
 52                    continue
 53                grad = p.grad.data
 54                if grad.is_sparse:
 55                    raise RuntimeError('Lamb does not support sparse gradients, consider SparseAdam instad.')
 56
 57                state = self.state[p]
 58
 59                # State initialization
 60                if len(state) == 0:
 61                    state['step'] = 0
 62                    # Exponential moving average of gradient values
 63                    state['exp_avg'] = torch.zeros_like(p.data)
 64                    # Exponential moving average of squared gradient values
 65                    state['exp_avg_sq'] = torch.zeros_like(p.data)
 66
 67                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
 68                beta1, beta2 = group['betas']
 69
 70                state['step'] += 1
 71
 72                # Decay the first and second moment running average coefficient
 73                # m_t
 74                exp_avg.mul_(beta1).add_(1 - beta1, grad)
 75                # v_t
 76                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
 77
 78                # Paper v3 does not use debiasing.
 79                # bias_correction1 = 1 - beta1 ** state['step']
 80                # bias_correction2 = 1 - beta2 ** state['step']
 81                # Apply bias to lr to avoid broadcast.
 82                step_size = group['lr'] # * math.sqrt(bias_correction2) / bias_correction1
 83
 84                weight_norm = p.data.pow(2).sum().sqrt().clamp(0, 10)
 85
 86                adam_step = exp_avg / exp_avg_sq.sqrt().add(group['eps'])
 87                if group['weight_decay'] != 0:
 88                    adam_step.add_(group['weight_decay'], p.data)
 89
 90                adam_norm = adam_step.pow(2).sum().sqrt()
 91                if weight_norm == 0 or adam_norm == 0:
 92                    trust_ratio = 1
 93                else:
 94                    trust_ratio = weight_norm / adam_norm
 95                state['weight_norm'] = weight_norm
 96                state['adam_norm'] = adam_norm
 97                state['trust_ratio'] = trust_ratio
 98                if self.adam:
 99                    trust_ratio = 1
100
101                p.data.add_(-step_size * trust_ratio, adam_step)
102
103        return loss
104
105 import time
106import torch.nn as nn
107import torch
108from tfrecord.torch.dataset import TFRecordDataset
109import numpy as np
110import os
111
112LEARNING_RATE = 0.001
113EPOCH = 40
114BATCH_SIZE = 2
115MAX_GRAD_NORM = 1.0
116
117print(f"--- Resume/Start training ---")   
118feat_map = {"input_ids": "int", 
119           "input_mask": "int",
120           "segment_ids": "int",
121           "next_sentence_labels": "int",
122           "masked_lm_positions": "int",
123           "masked_lm_ids": "int"}
124pretrain_file = 'restaurant_review_train'
125
126# Create albert pretrain model
127config = AlbertConfig.from_json_file("albert_config.json")
128albert_pretrain = AlbertForPretrain(config)
129# Create optimizer
130optimizer = Lamb([{"params": [p for n, p in list(albert_pretrain.named_parameters())]}], lr=LEARNING_RATE)
131albert_pretrain.train()
132dataset = TFRecordDataset(pretrain_file, index_path = None, description=feat_map)
133loader = torch.utils.data.DataLoader(dataset, batch_size=BATCH_SIZE)
134
135tmp_loss = 0
136start_time = time.time()
137
138if os.path.isfile('pretrain_checkpoint'):
139    print(f"--- Load from checkpoint ---")
140    checkpoint = torch.load("pretrain_checkpoint")
141    albert_pretrain.load_state_dict(checkpoint['model_state_dict'])
142    optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
143    epoch = checkpoint['epoch']
144    loss = checkpoint['loss']
145    losses = checkpoint['losses']
146
147else:
148    epoch = -1
149    losses = []
150for e in range(epoch+1, EPOCH):
151    for batch in loader:
152        b_input_ids = batch['input_ids'].long() 
153        b_token_type_ids = batch['segment_ids'].long() 
154        b_seq_relationship_labels = batch['next_sentence_labels'].long()
155
156        # Convert the dataformat from loaded decoded format into format 
157        # loaded format is created by google's Albert create_pretrain.py script
158        # required by huggingfaces pytorch implementation of albert
159        mask_rows = np.nonzero(batch['masked_lm_positions'].numpy())[0]
160        mask_cols = batch['masked_lm_positions'].numpy()[batch['masked_lm_positions'].numpy()!=0]
161        b_attention_mask = np.zeros((BATCH_SIZE,64),dtype=np.int64)
162        b_attention_mask[mask_rows,mask_cols] = 1
163        b_masked_lm_labels = np.zeros((BATCH_SIZE,64),dtype=np.int64) - 100
164        b_masked_lm_labels[mask_rows,mask_cols] = batch['masked_lm_ids'].numpy()[batch['masked_lm_positions'].numpy()!=0]     
165        b_attention_mask=torch.tensor(b_attention_mask).long()
166        b_masked_lm_labels=torch.tensor(b_masked_lm_labels).long()
167
168
169        loss = albert_pretrain(input_ids = b_input_ids
170                              , attention_mask = b_attention_mask
171                              , token_type_ids = b_token_type_ids
172                              , masked_lm_labels = b_masked_lm_labels 
173                              , seq_relationship_labels = b_seq_relationship_labels)
174
175        # clears old gradients
176        optimizer.zero_grad()
177        # backward pass
178        loss.backward()
179        # gradient clipping
180        torch.nn.utils.clip_grad_norm_(parameters=albert_pretrain.parameters(), max_norm=MAX_GRAD_NORM)
181        # update parameters
182        optimizer.step()
183
184        tmp_loss += loss.detach().item()
185
186    # print metrics and save to checkpoint every epoch
187    print(f"Epoch: {e}")
188    print(f"Train loss: {(tmp_loss/20)}")
189    print(f"Train Time: {(time.time()-start_time)/60} mins")  
190    losses.append(tmp_loss/20)
191
192    tmp_loss = 0
193    start_time = time.time()
194
195    torch.save({'model_state_dict': albert_pretrain.state_dict(),'optimizer_state_dict': optimizer.state_dict(),
196               'epoch': e, 'loss': loss,'losses': losses}
197           , 'pretrain_checkpoint')
198from matplotlib import pyplot as plot
199plot.plot(losses)
200
201#Fine tuning ALBERT
202
203# At the time of writing, Hugging face didnt provide the class object for 
204# AlbertForTokenClassification, hence write your own defination below
205from transformers.modeling_albert import AlbertModel, AlbertPreTrainedModel
206from transformers.configuration_albert import AlbertConfig
207from transformers.tokenization_bert import BertTokenizer
208import torch.nn as nn
209from torch.nn import CrossEntropyLoss
210class AlbertForTokenClassification(AlbertPreTrainedModel):
211
212    def __init__(self, albert, config):
213        super().__init__(config)
214        self.num_labels = config.num_labels
215
216        self.albert = albert
217        self.dropout = nn.Dropout(config.hidden_dropout_prob)
218        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
219
220    def forward(
221        self,
222        input_ids=None,
223        attention_mask=None,
224        token_type_ids=None,
225        position_ids=None,
226        head_mask=None,
227        inputs_embeds=None,
228        labels=None,
229    ):
230
231        outputs = self.albert(
232            input_ids,
233            attention_mask=attention_mask,
234            token_type_ids=token_type_ids,
235            position_ids=position_ids,
236            head_mask=head_mask,
237            inputs_embeds=inputs_embeds,
238        )
239
240        sequence_output = outputs[0]
241
242        sequence_output = self.dropout(sequence_output)
243        logits = self.classifier(sequence_output)
244
245        return logits
246
247import numpy as np
248def label_sent(name_tokens, sent_tokens):
249    label = []
250    i = 0
251    if len(name_tokens)>len(sent_tokens):
252        label = np.zeros(len(sent_tokens))
253    else:
254        while i<len(sent_tokens):
255            found_match = False
256            if name_tokens[0] == sent_tokens[i]:       
257                found_match = True
258                for j in range(len(name_tokens)-1):
259                    if ((i+j+1)>=len(sent_tokens)):
260                        return label
261                    if name_tokens[j+1] != sent_tokens[i+j+1]:
262                        found_match = False
263                if found_match:
264                    label.extend(list(np.ones(len(name_tokens)).astype(int)))
265                    i = i + len(name_tokens)
266                else: 
267                    label.extend([0])
268                    i = i+ 1
269            else:
270                label.extend([0])
271                i=i+1
272    return label
273
274import pandas as pd
275import glob
276import os
277
278tokenizer = BertTokenizer(vocab_file="vocab.txt")
279
280df_data_train = pd.read_csv("dish_name_train.csv")
281df_data_train['name_tokens'] = df_data_train['dish_name'].apply(tokenizer.tokenize)
282df_data_train['review_tokens'] = df_data_train.review.apply(tokenizer.tokenize)
283df_data_train['review_label'] = df_data_train.apply(lambda row: label_sent(row['name_tokens'], row['review_tokens']), axis=1)
284
285df_data_val = pd.read_csv("dish_name_val.csv")
286df_data_val = df_data_val.dropna().reset_index()
287df_data_val['name_tokens'] = df_data_val['dish_name'].apply(tokenizer.tokenize)
288df_data_val['review_tokens'] = df_data_val.review.apply(tokenizer.tokenize)
289df_data_val['review_label'] = df_data_val.apply(lambda row: label_sent(row['name_tokens'], row['review_tokens']), axis=1)
290
291MAX_LEN = 64
292BATCH_SIZE = 1
293from keras.preprocessing.sequence import pad_sequences
294import torch
295from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
296
297tr_inputs = pad_sequences([tokenizer.convert_tokens_to_ids(txt) for txt in df_data_train['review_tokens']],maxlen=MAX_LEN, dtype="long", truncating="post", padding="post")
298tr_tags = pad_sequences(df_data_train['review_label'],maxlen=MAX_LEN, padding="post",dtype="long", truncating="post")
299# create the mask to ignore the padded elements in the sequences.
300tr_masks = [[float(i>0) for i in ii] for ii in tr_inputs]
301tr_inputs = torch.tensor(tr_inputs)
302tr_tags = torch.tensor(tr_tags)
303tr_masks = torch.tensor(tr_masks)
304train_data = TensorDataset(tr_inputs, tr_masks, tr_tags)
305train_sampler = RandomSampler(train_data)
306train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=BATCH_SIZE)
307
308
309val_inputs = pad_sequences([tokenizer.convert_tokens_to_ids(txt) for txt in df_data_val['review_tokens']],maxlen=MAX_LEN, dtype="long", truncating="post", padding="post")
310val_tags = pad_sequences(df_data_val['review_label'],maxlen=MAX_LEN, padding="post",dtype="long", truncating="post")
311# create the mask to ignore the padded elements in the sequences.
312val_masks = [[float(i>0) for i in ii] for ii in val_inputs]
313val_inputs = torch.tensor(val_inputs)
314val_tags = torch.tensor(val_tags)
315val_masks = torch.tensor(val_masks)
316val_data = TensorDataset(val_inputs, val_masks, val_tags)
317val_sampler = RandomSampler(val_data)
318val_dataloader = DataLoader(val_data, sampler=val_sampler, batch_size=BATCH_SIZE)
319
320model_tokenclassification = AlbertForTokenClassification(albert_pretrain.albert, config)
321from torch.optim import Adam
322LEARNING_RATE = 0.0000003
323FULL_FINETUNING = True
324if FULL_FINETUNING:
325    param_optimizer = list(model_tokenclassification.named_parameters())
326    no_decay = ['bias', 'gamma', 'beta']
327    optimizer_grouped_parameters = [
328        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
329         'weight_decay_rate': 0.01},
330        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
331         'weight_decay_rate': 0.0}
332    ]
333else:
334    param_optimizer = list(model_tokenclassification.classifier.named_parameters()) 
335    optimizer_grouped_parameters = [{"params": [p for n, p in param_optimizer]}]
336optimizer = Adam(optimizer_grouped_parameters, lr=LEARNING_RATE)

第四步：为自定义语料库训练模型

  1#Training the model
  2
  3# from torch.utils.tensorboard import SummaryWriter
  4import time
  5import os.path
  6import torch.nn as nn
  7import torch
  8EPOCH = 800
  9MAX_GRAD_NORM = 1.0
 10
 11start_time = time.time()
 12tr_loss, tr_acc, nb_tr_steps = 0, 0, 0
 13eval_loss, eval_acc, nb_eval_steps = 0, 0, 0
 14
 15if os.path.isfile('finetune_checkpoint'):
 16    print(f"--- Load from checkpoint ---")
 17    checkpoint = torch.load("finetune_checkpoint")
 18    model_tokenclassification.load_state_dict(checkpoint['model_state_dict'])
 19    optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
 20    epoch = checkpoint['epoch']
 21    train_losses = checkpoint['train_losses']
 22    train_accs = checkpoint['train_accs']
 23    eval_losses = checkpoint['eval_losses']
 24    eval_accs = checkpoint['eval_accs']
 25
 26else:
 27    epoch = -1
 28    train_losses,train_accs,eval_losses,eval_accs = [],[],[],[]
 29
 30print(f"--- Resume/Start training ---")    
 31for e in range(epoch+1, EPOCH): 
 32
 33    # TRAIN loop
 34    model_tokenclassification.train()
 35
 36    for batch in train_dataloader:
 37        # add batch to gpu
 38        batch = tuple(t for t in batch)
 39        b_input_ids, b_input_mask, b_labels = batch
 40        # forward pass
 41        b_outputs = model_tokenclassification(b_input_ids, token_type_ids=None, attention_mask=b_input_mask, labels=b_labels)
 42
 43        ce_loss_fct = CrossEntropyLoss()
 44        # Only keep active parts of the loss
 45        b_active_loss = b_input_mask.view(-1) == 1
 46        b_active_logits = b_outputs.view(-1, config.num_labels)[b_active_loss]
 47        b_active_labels = b_labels.view(-1)[b_active_loss]
 48
 49        loss = ce_loss_fct(b_active_logits, b_active_labels)
 50        acc = torch.mean((torch.max(b_active_logits.detach(),1)[1] == b_active_labels.detach()).float())
 51
 52        model_tokenclassification.zero_grad()
 53        # backward pass
 54        loss.backward()
 55        # track train loss
 56        tr_loss += loss.item()
 57        tr_acc += acc
 58        nb_tr_steps += 1
 59        # gradient clipping
 60        torch.nn.utils.clip_grad_norm_(parameters=model_tokenclassification.parameters(), max_norm=MAX_GRAD_NORM)
 61        # update parameters
 62        optimizer.step()
 63
 64
 65    # VALIDATION on validation set
 66    model_tokenclassification.eval()
 67    for batch in val_dataloader:
 68        batch = tuple(t for t in batch)
 69        b_input_ids, b_input_mask, b_labels = batch
 70
 71        with torch.no_grad():
 72
 73            b_outputs = model_tokenclassification(b_input_ids, token_type_ids=None,
 74                         attention_mask=b_input_mask, labels=b_labels)
 75
 76            loss_fct = CrossEntropyLoss()
 77            # Only keep active parts of the loss
 78            b_active_loss = b_input_mask.view(-1) == 1
 79            b_active_logits = b_outputs.view(-1, config.num_labels)[b_active_loss]
 80            b_active_labels = b_labels.view(-1)[b_active_loss]
 81            loss = loss_fct(b_active_logits, b_active_labels)
 82            acc = np.mean(np.argmax(b_active_logits.detach().cpu().numpy(), axis=1).flatten() == b_active_labels.detach().cpu().numpy().flatten())
 83
 84        eval_loss += loss.mean().item()
 85        eval_acc += acc
 86        nb_eval_steps += 1    
 87
 88    if e % 10 ==0:
 89
 90        print(f"Epoch: {e}")
 91        print(f"Train loss: {(tr_loss/nb_tr_steps)}")
 92        print(f"Train acc: {(tr_acc/nb_tr_steps)}")
 93        print(f"Train Time: {(time.time()-start_time)/60} mins")  
 94
 95        print(f"Validation loss: {eval_loss/nb_eval_steps}")
 96        print(f"Validation Accuracy: {(eval_acc/nb_eval_steps)}") 
 97
 98        train_losses.append(tr_loss/nb_tr_steps)
 99        train_accs.append(tr_acc/nb_tr_steps)
100        eval_losses.append(eval_loss/nb_eval_steps)
101        eval_accs.append(eval_acc/nb_eval_steps)
102
103
104        tr_loss, tr_acc, nb_tr_steps = 0, 0, 0 
105        eval_loss, eval_acc, nb_eval_steps = 0, 0, 0 
106        start_time = time.time() 
107
108        torch.save({'model_state_dict': model_tokenclassification.state_dict(),'optimizer_state_dict': optimizer.state_dict(),
109           'epoch': e, 'train_losses': train_losses,'train_accs': train_accs, 'eval_losses':eval_losses,'eval_accs':eval_accs}
110       , 'finetune_checkpoint')
111
112plot.plot(train_losses)
113plot.plot(train_accs)
114plot.plot(eval_losses)
115plot.plot(eval_accs)
116plot.legend(labels = ['train_loss','train_accuracy','validation_loss','validation_accuracy'])

第五步：预测

 1#Prediction
 2
 3def predict(texts):
 4    tokenized_texts = [tokenizer.tokenize(txt) for txt in texts]
 5    input_ids = pad_sequences([tokenizer.convert_tokens_to_ids(txt) for txt in tokenized_texts],
 6                              maxlen=MAX_LEN, dtype="long", truncating="post", padding="post")
 7    attention_mask = [[float(i>0) for i in ii] for ii in input_ids]
 8
 9    input_ids = torch.tensor(input_ids)
10    attention_mask = torch.tensor(attention_mask)
11
12    dataset = TensorDataset(input_ids, attention_mask)
13    datasampler = SequentialSampler(dataset)
14    dataloader = DataLoader(dataset, sampler=datasampler, batch_size=BATCH_SIZE) 
15
16    predicted_labels = []
17
18    for batch in dataloader:
19        batch = tuple(t for t in batch)
20        b_input_ids, b_input_mask = batch
21
22        with torch.no_grad():
23            logits = model_tokenclassification(b_input_ids, token_type_ids=None,
24                           attention_mask=b_input_mask)
25
26            predicted_labels.append(np.multiply(np.argmax(logits.detach().cpu().numpy(),axis=2), b_input_mask.detach().cpu().numpy()))
27    # np.concatenate(predicted_labels), to flatten list of arrays of batch_size * max_len into list of arrays of max_len
28    return np.concatenate(predicted_labels).astype(int), tokenized_texts
29
30def get_dish_candidate_names(predicted_label, tokenized_text):
31    name_lists = []
32    if len(np.where(predicted_label>0)[0])>0:
33        name_idx_combined = np.where(predicted_label>0)[0]
34        name_idxs = np.split(name_idx_combined, np.where(np.diff(name_idx_combined) != 1)[0]+1)
35        name_lists.append([" ".join(np.take(tokenized_text,name_idx)) for name_idx in name_idxs])
36        # If there duplicate names in the name_lists
37        name_lists = np.unique(name_lists)
38        return name_lists
39    else:
40        return None
41
42texts = df_data_val.review.values
43predicted_labels, _ = predict(texts)
44df_data_val['predicted_review_label'] = list(predicted_labels)
45df_data_val['predicted_name']=df_data_val.apply(lambda row: get_dish_candidate_names(row.predicted_review_label, row.review_tokens)
46                                                , axis=1)
47
48texts = df_data_train.review.values
49predicted_labels, _ = predict(texts)
50df_data_train['predicted_review_label'] = list(predicted_labels)
51df_data_train['predicted_name']=df_data_train.apply(lambda row: get_dish_candidate_names(row.predicted_review_label, row.review_tokens)
52                                                , axis=1)
53
54(df_data_val)

实验结果

可以看到，模型成功地从用餐评论中，提取出了菜名。

模型比拼

从上面的实战应用中可以看到，ALBERT虽然很lite，结果也可以说相当不错。

那么，参数少、结果好，是否就可以替代BERT呢？

我们可以仔细看下二者实验性能的比较，这里的Speedup是指训练时间。

因为数据数据少了，分布式训练时吞吐上去了，所以ALBERT训练更快。但推理时间还是需要和BERT一样的transformer计算。

所以可以总结为：

• 在相同的训练时间下，ALBERT效果要比BERT好。
• 在相同的推理时间下，ALBERT base和large的效果都是没有BERT好。

此外，Naman Bansal认为，由于ALBERT的结构，实现ALBERT的计算代价比BERT要高一些。

所以，还是“鱼和熊掌不可兼得”的关系，要想让ALBERT完全超越、替代BERT，还需要做更进一步的研究和改良。

传送门

博客地址：
https://medium.com/@namanbansal9909/should-we-shift-from-bert-to-albert-e6fbb7779d3e