bert中文情感分析多分类任务详解

查看GPU版本和使用情况

import torch
if torch.cuda.is_available():
    device = torch.device("cuda")
    print('There are %d GPU(s) available.' % torch.cuda.device_count())
    print('We will use the GPU:', torch.cuda.get_device_name(0))
else:
    print('No GPU available, using the CPU instead.')
    device = torch.device("cpu")

There are 1 GPU(s) available.
We will use the GPU: GeForce GTX 1070

导入评论信息

import pandas as pd
path = './中文文本情感分类/'
comments = pd.read_csv(path + '社交网站评论信息情感分类.csv')
moods = {0: '喜悦', 1: '愤怒', 2: '厌恶', 3: '低落'}

print('文本数量（总体）：%d' % comments.shape[0])

for label, mood in moods.items(): 
    print('文本数量（{}）：{}'.format(mood,  comments[comments.label==label].shape[0]))

文本数量（总体）：361744
文本数量（喜悦）：199496
文本数量（愤怒）：51714
文本数量（厌恶）：55267
文本数量（低落）：55267

简单查看一下数据集

comments[0:5]

	label	review
0	0	啊呀呀！要死啦！么么么！只穿外套就好了，我认为里面那件很多余啊周小伦喜歡 你各種 五角星的...
1	0	嗯……既然大姚通知了……那我也表示下收到……姚，你知道吗？假如外星人入侵地球，只要摧毁我们的...
2	0	风格不一样嘛，都喜欢！最喜欢哪张？
3	0	好呀，试试D .I .Y .去死皮面膜1.将燕麦片加水中浸泡6小时，加入木瓜牛奶搅拌。2.放...
4	0	张老师，谢谢侬的信任！粉丝多少无所谓重在质地近日发现一个现象——他加了你关注，你回加后，他立...

由于数据实在太多，我们在这里每一种情感 选择1000个例子

# 标签为 0 的数据
df0 = comments.loc[comments.label == 0].sample(1000)[['review', 'label']]
# 标签为 1 的数据
df1 = comments.loc[comments.label == 1].sample(1000)[['review', 'label']]
# 标签为 2 的数据
df2 = comments.loc[comments.label == 2].sample(1000)[['review', 'label']]
# 标签为 3 的数据
df3 = comments.loc[comments.label == 3].sample(1000)[['review', 'label']]

由于数据集太大,组成新的数据集并且打乱顺序

df0 = df0.append(df1)
df0 = df0.append(df2)
df0 = df0.append(df3)
df0 = df0.append(df3)
df0 = df0.sample(frac=1)
len(df0)

5000

简单查看一下数据

df0[0:5]

	review	label
355831	据最新消息，此次火灾造成9人死亡9人受伤。经初步勘察，火灾系售楼处一楼沙盘电路故障引发。	3
348348	上星期买的裙子,今晚穿了,不好看~ ~	3
357542	顺便说点什么吧...荷兰，阿姆斯特丹火车站，广播里突然传来耳熟能详的《音乐之声》里的123，...	3
361326	午后，打扮的美美的和喜欢的人坐在咖啡厅，聊聊自己想说的话，或静静坐在一起，什么都不说，只是紧...	3
278616	真的假的？！！！	2

把数据集中的句子和标签取出来

# 句子和标签
sentences = df0.review.values
labels = df0.label.values

下载 BERT tokenizer.

from transformers import BertTokenizer
print('下载 BERT tokenizer...')

tokenizer = BertTokenizer.from_pretrained('bert-base-chinese', do_lower_case=True)

下载 BERT tokenizer...

简单查看一下 tokenizer

print(' 原句: ', sentences[0])
print('Tokenizen 后的句子: ', tokenizer.tokenize(sentences[0]))
print('Token IDs: ', tokenizer.convert_tokens_to_ids(tokenizer.tokenize(sentences[0])))

 原句:  据最新消息，此次火灾造成9人死亡9人受伤。经初步勘察，火灾系售楼处一楼沙盘电路故障引发。
Tokenizen 后的句子:  ['据', '最', '新', '消', '息', '，', '此', '次', '火', '灾', '造', '成', '9', '人', '死', '亡', '9', '人', '受', '伤', '。', '经', '初', '步', '勘', '察', '，', '火', '灾', '系', '售', '楼', '处', '一', '楼', '沙', '盘', '电', '路', '故', '障', '引', '发', '。']
Token IDs:  [2945, 3297, 3173, 3867, 2622, 8024, 3634, 3613, 4125, 4135, 6863, 2768, 130, 782, 3647, 767, 130, 782, 1358, 839, 511, 5307, 1159, 3635, 1242, 2175, 8024, 4125, 4135, 5143, 1545, 3517, 1905, 671, 3517, 3763, 4669, 4510, 6662, 3125, 7397, 2471, 1355, 511]

最长句子的长度为 265

max_len = 0
lengthOfsentence = []
# 循环每一个句子...
for sent in sentences:

    lengthOfsentence.append(len(sent))
    # 找到句子最大长度
    max_len = max(max_len, len(sent))

print('最长的句子长度为: ', max_len)

最长的句子长度为:  284

type(sentences)

numpy.ndarray

sentences.shape

(5000,)

根据观察，大多数句子长度在250 以下，padding 时候的max_length 我们取256

import matplotlib.pyplot as plt
plt.plot(lengthOfsentence)
plt.ylabel('some numbers')
plt.show()

<Figure size 640x480 with 1 Axes>

input_ids = []
attention_masks = []

for sent in sentences:
    # `encode_plus` will:
    #   (1) Tokenize the sentence.
    #   (2) Prepend the `[CLS]` token to the start.
    #   (3) Append the `[SEP]` token to the end.
    #   (4) Map tokens to their IDs.
    #   (5) Pad or truncate the sentence to `max_length`
    #   (6) Create attention masks for [PAD] tokens.
    encoded_dict = tokenizer.encode_plus(
                        sent,                      # Sentence to encode.
                        add_special_tokens = True, # Add '[CLS]' and '[SEP]'
                        max_length = 256,           # Pad & truncate all sentences.
                        pad_to_max_length = True,
                        return_attention_mask = True,   # Construct attn. masks.
                        return_tensors = 'pt',     # Return pytorch tensors.
                   )
    
    # 把编码的句子加入list.    
    input_ids.append(encoded_dict['input_ids'])
    
    # 加上 attention mask (simply differentiates padding from non-padding).
    attention_masks.append(encoded_dict['attention_mask'])
    

# 把lists 转为 tensors.
input_ids = torch.cat(input_ids, dim=0)
attention_masks = torch.cat(attention_masks, dim=0)
labels = torch.tensor(labels)

简单查看一下第一句的Token IDs 和 attention_masks

print('原句: ', sentences[0])
print('Token IDs:', input_ids[0])
print('attention_masks:', attention_masks[0])

原句:  国家哀悼日，悲伤无边，情义永在这里有毁灭，有伤痛，有绝望，但还有希望。
Token IDs: tensor([ 101, 1744, 2157, 1500, 2656, 3189, 8024, 2650,  839, 3187, 6804, 8024,
        2658,  721, 3719, 1762, 6821, 7027, 3300, 3673, 4127, 8024, 3300,  839,
        4578, 8024, 3300, 5318, 3307, 8024,  852, 6820, 3300, 2361, 3307,  511,
         102,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
           0,    0,    0,    0])
attention_masks: tensor([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])

设计 training，validation 和 test dataset

from torch.utils.data import TensorDataset, random_split

# 把input 放入 TensorDataset。
dataset = TensorDataset(input_ids, attention_masks, labels)

# 计算 train_size 和 val_size 的长度.
train_size = int(0.9 * len(dataset))
val_size = len(dataset) - train_size

# 90% 的dataset 为train_dataset, 10% 的的dataset 为val_dataset.
train_dataset, val_dataset = random_split(dataset, [train_size, val_size])
print('{:>5,} 训练数据'.format(train_size))
print('{:>5,} 验证数据'.format(val_size))

4,500 训练数据
  500 验证数据

制作dataload

from torch.utils.data import DataLoader, RandomSampler, SequentialSampler

# 推荐batch_size 为 16 或者 32
batch_size = 16

# 为训练数据集和验证数据集设计DataLoaders. 
train_dataloader = DataLoader(
            train_dataset,  # 训练数据.
            sampler = RandomSampler(train_dataset), # 打乱顺序
            batch_size = batch_size 
        )

validation_dataloader = DataLoader(
            val_dataset, # 验证数据.
            sampler = RandomSampler(val_dataset), # 打乱顺序
            batch_size = batch_size 
        )

导入 bert 文本多分类模型 BertForSequenceClassification

from transformers import BertForSequenceClassification, AdamW, BertConfig

model = BertForSequenceClassification.from_pretrained(
    "bert-base-chinese", # 使用 12-layer 的 BERT 模型.
    num_labels = 4, # 多分类任务的输出标签为 4个.                     
    output_attentions = False, # 不返回 attentions weights.
    output_hidden_states = False, # 不返回 all hidden-states.
)
model.cuda()

BertForSequenceClassification(
  (bert): BertModel(
    (embeddings): BertEmbeddings(
      (word_embeddings): Embedding(21128, 768, padding_idx=0)
      (position_embeddings): Embedding(512, 768)
      (token_type_embeddings): Embedding(2, 768)
      (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
      (dropout): Dropout(p=0.1, inplace=False)
    )
    (encoder): BertEncoder(
      (layer): ModuleList(
        (0): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (1): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (2): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (3): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (4): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (5): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (6): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (7): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (8): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (9): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (10): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
        (11): BertLayer(
          (attention): BertAttention(
            (self): BertSelfAttention(
              (query): Linear(in_features=768, out_features=768, bias=True)
              (key): Linear(in_features=768, out_features=768, bias=True)
              (value): Linear(in_features=768, out_features=768, bias=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): BertSelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): BertIntermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
          )
          (output): BertOutput(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
    )
    (pooler): BertPooler(
      (dense): Linear(in_features=768, out_features=768, bias=True)
      (activation): Tanh()
    )
  )
  (dropout): Dropout(p=0.1, inplace=False)
  (classifier): Linear(in_features=768, out_features=4, bias=True)
)

# Get all of the model's parameters as a list of tuples.
params = list(model.named_parameters())

print('The BERT model has {:} different named parameters.\n'.format(len(params)))

print('==== Embedding Layer ====\n')

for p in params[0:5]:
    print("{:<55} {:>12}".format(p[0], str(tuple(p[1].size()))))

print('\n==== First Transformer ====\n')

for p in params[5:21]:
    print("{:<55} {:>12}".format(p[0], str(tuple(p[1].size()))))

print('\n==== Output Layer ====\n')

for p in params[-4:]:
    print("{:<55} {:>12}".format(p[0], str(tuple(p[1].size()))))

The BERT model has 201 different named parameters.

==== Embedding Layer ====

bert.embeddings.word_embeddings.weight                  (21128, 768)
bert.embeddings.position_embeddings.weight                (512, 768)
bert.embeddings.token_type_embeddings.weight                (2, 768)
bert.embeddings.LayerNorm.weight                              (768,)
bert.embeddings.LayerNorm.bias                                (768,)

==== First Transformer ====

bert.encoder.layer.0.attention.self.query.weight          (768, 768)
bert.encoder.layer.0.attention.self.query.bias                (768,)
bert.encoder.layer.0.attention.self.key.weight            (768, 768)
bert.encoder.layer.0.attention.self.key.bias                  (768,)
bert.encoder.layer.0.attention.self.value.weight          (768, 768)
bert.encoder.layer.0.attention.self.value.bias                (768,)
bert.encoder.layer.0.attention.output.dense.weight        (768, 768)
bert.encoder.layer.0.attention.output.dense.bias              (768,)
bert.encoder.layer.0.attention.output.LayerNorm.weight        (768,)
bert.encoder.layer.0.attention.output.LayerNorm.bias          (768,)
bert.encoder.layer.0.intermediate.dense.weight           (3072, 768)
bert.encoder.layer.0.intermediate.dense.bias                 (3072,)
bert.encoder.layer.0.output.dense.weight                 (768, 3072)
bert.encoder.layer.0.output.dense.bias                        (768,)
bert.encoder.layer.0.output.LayerNorm.weight                  (768,)
bert.encoder.layer.0.output.LayerNorm.bias                    (768,)

==== Output Layer ====

bert.pooler.dense.weight                                  (768, 768)
bert.pooler.dense.bias                                        (768,)
classifier.weight                                           (4, 768)
classifier.bias                                                 (4,)

选择优化器

# AdamW 是一个 huggingface library 的类，'W' 是'Weight Decay fix"的意思。
optimizer = AdamW(model.parameters(),
                  lr = 2e-5, # args.learning_rate - 默认是 5e-5
                  eps = 1e-8 # args.adam_epsilon  - 默认是 1e-8， 是为了防止衰减率分母除到0
                )

设计learning rate scheduler, 调整learning rate.

from transformers import get_linear_schedule_with_warmup

# bert 推荐 epochs 在2到4之间为好。
epochs = 4

# training steps 的数量: [number of batches] x [number of epochs]. 
total_steps = len(train_dataloader) * epochs

# 设计 learning rate scheduler.
scheduler = get_linear_schedule_with_warmup(optimizer, 
                                            num_warmup_steps = 0, # Default value in run_glue.py
                                            num_training_steps = total_steps)

flat_accuracy 计算模型准确率

import numpy as np
def flat_accuracy(preds, labels):
    pred_flat = np.argmax(preds, axis=1).flatten()
    labels_flat = labels.flatten()
    return np.sum(pred_flat == labels_flat) / len(labels_flat)

format_time 计算所用时间

import time
import datetime
def format_time(elapsed):    
    elapsed_rounded = int(round((elapsed)))    
    # 返回 hh:mm:ss 形式的时间
    return str(datetime.timedelta(seconds=elapsed_rounded))

训练数据

import os
import random
import numpy as np
from transformers import WEIGHTS_NAME, CONFIG_NAME

output_dir = "./models/"
output_model_file = os.path.join(output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(output_dir, CONFIG_NAME)
# 代码参考 https://github.com/huggingface/transformers/blob/5bfcd0485ece086ebcbed2d008813037968a9e58/examples/run_glue.py#L128

# 设置随机种子.
seed_val = 42
random.seed(seed_val)
np.random.seed(seed_val)
torch.manual_seed(seed_val)
torch.cuda.manual_seed_all(seed_val)

# 记录training ,validation loss ,validation accuracy and timings.
training_stats = []

# 设置总时间.
total_t0 = time.time()
best_val_accuracy = 0

for epoch_i in range(0, epochs):      
    print('Epoch {:} / {:}'.format(epoch_i + 1, epochs))  

    # 记录每个 epoch 所用的时间
    t0 = time.time()
    total_train_loss = 0
    total_train_accuracy = 0
    model.train()
  
    for step, batch in enumerate(train_dataloader):

        # 每隔40个batch 输出一下所用时间.
        if step % 40 == 0 and not step == 0:            
            elapsed = format_time(time.time() - t0)
            print('  Batch {:>5,}  of  {:>5,}.    Elapsed: {:}.'.format(step, len(train_dataloader), elapsed))

        # `batch` 包括3个 tensors:
        #   [0]: input ids 
        #   [1]: attention masks
        #   [2]: labels 
        b_input_ids = batch[0].to(device)
        b_input_mask = batch[1].to(device)
        b_labels = batch[2].to(device)

        # 清空梯度
        model.zero_grad()        

        # forward        
        # 参考 https://huggingface.co/transformers/v2.2.0/model_doc/bert.html#transformers.BertForSequenceClassification
        loss, logits = model(b_input_ids, 
                             token_type_ids=None, 
                             attention_mask=b_input_mask, 
                             labels=b_labels)
       
        total_train_loss += loss.item()

        # backward 更新 gradients.
        loss.backward()

        # 减去大于1 的梯度，将其设为 1.0, 以防梯度爆炸.
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)

        # 更新模型参数 
        optimizer.step()
       
        # 更新 learning rate.
        scheduler.step()        
             
        logit = logits.detach().cpu().numpy()
        label_id = b_labels.to('cpu').numpy()
        # 计算training 句子的准确度.
        total_train_accuracy += flat_accuracy(logit, label_id)    
     
    # 计算batches的平均损失.
    avg_train_loss = total_train_loss / len(train_dataloader)      
    # 计算训练时间.
    training_time = format_time(time.time() - t0)
    
    # 训练集的准确率.
    avg_train_accuracy = total_train_accuracy / len(train_dataloader)
    print("  训练准确率: {0:.2f}".format(avg_train_accuracy))
    print("  平均训练损失 loss: {0:.2f}".format(avg_train_loss))
    print("  训练时间: {:}".format(training_time))
        
    # ========================================
    #               Validation
    # ========================================

    t0 = time.time()

    # 设置 model 为valuation 状态，在valuation状态 dropout layers 的dropout rate会不同
    model.eval()

    # 设置参数
    total_eval_accuracy = 0
    total_eval_loss = 0
    nb_eval_steps = 0

    for batch in validation_dataloader:        

        # `batch` 包括3个 tensors:
        #   [0]: input ids 
        #   [1]: attention masks
        #   [2]: labels 
        b_input_ids = batch[0].to(device)
        b_input_mask = batch[1].to(device)
        b_labels = batch[2].to(device)        

        # 在valuation 状态，不更新权值，不改变计算图
        with torch.no_grad():        

            # 参考 https://huggingface.co/transformers/v2.2.0/model_doc/bert.html#transformers.BertForSequenceClassification
            (loss, logits) = model(b_input_ids, 
                                   token_type_ids=None, 
                                   attention_mask=b_input_mask,
                                   labels=b_labels)
            
        # 计算 validation loss.
        total_eval_loss += loss.item()        
        logit = logits.detach().cpu().numpy()
        label_id = b_labels.to('cpu').numpy()

        # 计算 validation 句子的准确度.
        total_eval_accuracy += flat_accuracy(logit, label_id)
        
    # 计算 validation 的准确率.
    avg_val_accuracy = total_eval_accuracy / len(validation_dataloader)
    print("")
    print("  测试准确率: {0:.2f}".format(avg_val_accuracy))
    
    if avg_val_accuracy > best_val_accuracy:
        best_val_accuracy = avg_val_accuracy
        torch.save(model.state_dict(),output_model_file)
        model.config.to_json_file(output_config_file)
        tokenizer.save_vocabulary(output_dir)
         

    # 计算batches的平均损失.
    avg_val_loss = total_eval_loss / len(validation_dataloader)
    
    # 计算validation 时间.
    validation_time = format_time(time.time() - t0)
    
    print("  平均测试损失 Loss: {0:.2f}".format(avg_val_loss))
    print("  测试时间: {:}".format(validation_time))

    # 记录模型参数
    training_stats.append(
        {
            'epoch': epoch_i + 1,
            'Training Loss': avg_train_loss,
            'Valid. Loss': avg_val_loss,
            'Valid. Accur.': avg_val_accuracy,
            'Training Time': training_time,
            'Validation Time': validation_time
        }
    )

print("训练一共用了 {:} (h:mm:ss)".format(format_time(time.time()-total_t0)))

Epoch 1 / 4
  Batch    40  of    282.    Elapsed: 0:00:30.
  Batch    80  of    282.    Elapsed: 0:01:00.
  Batch   120  of    282.    Elapsed: 0:01:30.
  Batch   160  of    282.    Elapsed: 0:02:00.
  Batch   200  of    282.    Elapsed: 0:02:30.
  Batch   240  of    282.    Elapsed: 0:03:00.
  Batch   280  of    282.    Elapsed: 0:03:30.
  训练准确率: 0.43
  平均训练损失 loss: 1.27
  训练时间: 0:03:31

  测试准确率: 0.49
  平均测试损失 Loss: 1.20
  测试时间: 0:00:08
Epoch 2 / 4
  Batch    40  of    282.    Elapsed: 0:00:30.
  Batch    80  of    282.    Elapsed: 0:01:00.
  Batch   120  of    282.    Elapsed: 0:01:30.
  Batch   160  of    282.    Elapsed: 0:02:01.
  Batch   200  of    282.    Elapsed: 0:02:31.
  Batch   240  of    282.    Elapsed: 0:03:01.
  Batch   280  of    282.    Elapsed: 0:03:31.
  训练准确率: 0.54
  平均训练损失 loss: 1.09
  训练时间: 0:03:32

  测试准确率: 0.50
  平均测试损失 Loss: 1.20
  测试时间: 0:00:08
Epoch 3 / 4
  Batch    40  of    282.    Elapsed: 0:00:30.
  Batch    80  of    282.    Elapsed: 0:01:01.
  Batch   120  of    282.    Elapsed: 0:01:31.
  Batch   160  of    282.    Elapsed: 0:02:01.
  Batch   200  of    282.    Elapsed: 0:02:31.
  Batch   240  of    282.    Elapsed: 0:03:01.
  Batch   280  of    282.    Elapsed: 0:03:31.
  训练准确率: 0.68
  平均训练损失 loss: 0.81
  训练时间: 0:03:32

  测试准确率: 0.53
  平均测试损失 Loss: 1.18
  测试时间: 0:00:08
Epoch 4 / 4
  Batch    40  of    282.    Elapsed: 0:00:30.
  Batch    80  of    282.    Elapsed: 0:01:00.
  Batch   120  of    282.    Elapsed: 0:01:30.
  Batch   160  of    282.    Elapsed: 0:02:01.
  Batch   200  of    282.    Elapsed: 0:02:31.
  Batch   240  of    282.    Elapsed: 0:03:01.
  Batch   280  of    282.    Elapsed: 0:03:31.
  训练准确率: 0.77
  平均训练损失 loss: 0.61
  训练时间: 0:03:32

  测试准确率: 0.56
  平均测试损失 Loss: 1.13
  测试时间: 0:00:08
训练一共用了 0:14:40 (h:mm:ss)

简单测试一下

(_, logits) = model(input_ids[-20:].to(device), 
                                   token_type_ids=None, 
                                   attention_mask=attention_masks[-20:].to(device),
                                   labels=labels[-20:].to(device))

logits = logits.detach().cpu().numpy()
label_ids = labels[-20:].to('cpu').numpy()

acc = flat_accuracy(logits, label_ids)
acc

0.95

pred_flat = np.argmax(logits, axis=1).flatten()
pred_flat

array([3, 1, 3, 3, 1, 2, 3, 3, 3, 3, 3, 3, 1, 3, 2, 3, 0, 2, 0, 0],
      dtype=int64)

label_ids

array([3, 1, 3, 3, 1, 2, 3, 3, 3, 3, 3, 2, 1, 3, 2, 3, 0, 2, 0, 0],
      dtype=int64)