利用Transformers自定义一个神经网络结构

### 伪代码
import torch
from torch.optim import lr_scheduler, optimizer
from transformers import BertTokenizer, BertModel, BertPreTrainedModel
from torch import nn


# bert = BertModel('bert-base-uncased')


class BertModelCoustom(BertPreTrainedModel):
    """
    利用transformers定制化神经网络模型
    """

    def __init__(self, config):
        super(BertModelCoustom, self).__init__(config)
        # 导入配置，下面的语句对应后面实例化时候的from_pretrained
        self.bert = BertModel(config)
        
        self.linear = nn.Linear(768, 128)
        self.hidden_dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, x):
        # 定义前向传播
        return self.hidden_dropout(self.linear(self.bert(x)))


tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
bmc = BertModelCoustom.from_pretrained('bert-base-uncased')
for n, p in bmc.linear.named_parameters():
    print("线形层模型参数：")
    print(n)
    print(p)
    break

for n, p in bmc.bert.named_parameters():
    print("Bert模型参数：")
    print(n)
    print(p)
    break

print("hidden_dropout:", bmc.hidden_dropout)

# 调用模型
x = torch.tensor([[1, 2, 3]])
output = bmc(x)
loss = output['loss']
loss.backward()
optimizer.step()
lr_scheduler.step()