利用GPT2训练中文闲聊模型

小咪咪 2022-11-18 04:50 579阅读 0赞

# 利用GPT2模型训练中文闲聊模型 #

最近看了一下GPT2模型，看到很多博主都用来写诗歌，做问答等，小编突然萌生一个想法，利用GPT2来训练一个闲聊模型！！(小说生成器模型已经破产，写出来的东西狗屁不通，懒得再弄了，有兴趣的小伙伴可以继续尝试。。)，闲聊模型在小编的[Github][]上，欢迎star和fork，谢谢！！

--------------------

### 文章目录 ###

*  利用GPT2模型训练中文闲聊模型
 *  前言
 *  一、数据结构
 *  二、模型搭建

--------------------

# 前言 #

本来想用[清源CPM预训练模型][CPM]来进行的，可惜硬件条件不够，只能用中GPT模型来训练一个中文闲聊模型了。硬件条件好的小伙伴，推荐你们使用CPM来做，看别人做的效果还是不错的，清源CPM的模型有4.47G，小伙伴们量力而行，模型小编已经撸下来了：[pytorch版CPM][pytorch_CPM]密码：k2zh； [百度飞桨版CPM][CPM 1]密码：nb67；[苏神版bert4keras][bert4keras]，需要的小伙伴可自信下载。

# 一、数据结构 #

数据是在网上爬取的一些对话数据，数据格式如下所示：

> 你吃了吗？  
> 吃过了，吃的糖醋排骨，你呢  
> 我吃的是麻辣小龙虾

> 手机欠费了怎么办？  
> 交话费啊  
> 去哪里才能交话费呢  
> 去相应的营业厅啊

数据格式就是一段对话，不同的对话间使用空行隔开。

# 二、模型搭建 #

(1) GPT2模型的搭建(小编自己写的，可能不是很合适！！有需要的可以直接调用Transformers库中的GPT2Model)  
首先是Multi\_Heads\_Attention的代码，这个也没什么可说的，如下所示：

class Attention(nn.Module):
        def __init__(self, embedding_size, num_attention_heads, attention_dropout, residual_dropout):
            super(Attention, self).__init__()
    
            self.num_attention_heads = num_attention_heads
            self.size_per_head = embedding_size // num_attention_heads
            self.embedding_size = embedding_size
    
            self.query_key_value = nn.Linear(embedding_size, embedding_size * 3)
            self.attn_drop = nn.Dropout(attention_dropout)
            self.resid_drop = nn.Dropout(residual_dropout)
            self.dense = nn.Linear(embedding_size, embedding_size)
    
        def split_heads(self, x):
            "return shape [`batch`, `head`, `sequence`, `features`]"
            new_shape = x.size()[:-1] + (self.num_attention_heads, self.size_per_head)
            x = x.view(*new_shape)
            return x.permute(0, 2, 1, 3)
    
        def merge_heads(self, x):
            x = x.permute(0, 2, 1, 3).contiguous()
            new_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
            return x.view(*new_shape)
    
        def forward(self, x, kv_cache=None):
            self.seq_len = x.size(1)
    
            # self_attention
            x = self.query_key_value(x)
            q, k, v = x.split(self.embedding_size, dim=2)
    
            # 多头
            q = self.split_heads(q)
            k = self.split_heads(k)
            v = self.split_heads(v)
    
            cached_kv = torch.stack([k, v], dim=1)
    
            scores = torch.matmul(q, k.transpose(-2, -1))
            scores = scores / math.sqrt(self.size_per_head)
            attention_mask = torch.tril(torch.ones([self.seq_len, self.seq_len], dtype=torch.float32))
            # print("attention", attention_mask)
            attention_mask = attention_mask.reshape([1, 1, self.seq_len, self.seq_len])
            # print(1.0 - attention_mask)
            # print(scores * attention_mask)
            scores = scores * attention_mask - 10000.0 * (1.0 - attention_mask)
            # print(scores)
            scores = nn.Softmax(dim=-1)(scores)
            scores = self.attn_drop(scores)
            y = torch.matmul(scores, v)
            y = self.merge_heads(y)
            y = self.resid_drop(self.dense(y))
            return y, cached_kv

然后就是MLP部分的代码，也就是线性层：

# 构建线性转换层
    class MLP(nn.Module):
        def __init__(self, embedding_size):
            super(MLP, self).__init__()
            self.dense_h_to_4h = nn.Linear(embedding_size, embedding_size * 4)
            self.dense_4h_to_h = nn.Linear(embedding_size * 4, embedding_size)
            self.act = nn.functional.gelu
    
        def forward(self, x):
            h = self.act(self.dense_h_to_4h(x))
            h2 = self.dense_4h_to_h(h)
            return h2
    
    
    # 线性层测试
    """
    layer = Linear(768, 768*3)
    x = torch.rand(1,4,768) [batch_size, seq_len, dim]
    y = layer(x)
    print(y, y.shape) [1, 4, 2304]
    """

然后是写一个Block类把Multi\_Heads\_Attention和MLP进行整合：

class Block(nn.Module):
        def __init__(self, embedding_size, num_attention_heads, attention_dropout, residual_dropout):
            super(Block, self).__init__()
            self.input_layernorm = nn.LayerNorm(embedding_size, eps=1e-5)
            self.attention = Attention(embedding_size, num_attention_heads, attention_dropout, residual_dropout)
            self.post_attention_layernorm = nn.LayerNorm(embedding_size, eps=1e-5)
            self.mlp = MLP(embedding_size)
    
        def forward(self, x, kv_cache=None):
            # Attention + 前后的LayerNorm + 中间残差连接
            attn, cached_kv = self.attention(self.input_layernorm(x), kv_cache=kv_cache)
            x = x + attn
            z = self.post_attention_layernorm(x)
    
            # MLP
            z = self.mlp(z)
    
            # 残差连接
            x = x + z
            return x, cached_kv

通过调用Block类，构建Transformer模型：

class Transformer(nn.Module):
        def __init__(self,
                     layer_size,
                     embedding_size,
                     num_attention_heads,
                     attention_dropout,
                     residual_dropout):
            super(Transformer, self).__init__()
    
            self.layers = nn.ModuleList([Block(
                embedding_size,
                num_attention_heads,
                attention_dropout,
                residual_dropout)
                for _ in range(layer_size)])
    
            self.final_layernorm = nn.LayerNorm(embedding_size, eps=1e-5)
    
        def forward(self, x, kv_cache=None):
            # 多层 Block
            cached_kvs = []
            for i, layer in enumerate(self.layers):
                x, cached_kv = layer(
                    x, kv_cache=kv_cache[i] if kv_cache is not None else None)
                cached_kvs.append(cached_kv)
    
            # 最终的 LayerNorm
            x = self.final_layernorm(x)
    
            return x, torch.stack(cached_kvs)

最后就是GPT2Model了，在这个类中除了调用了Transformer类外，还定义了attention\_mask矩阵：

class GPT2Model(nn.Module):
        def __init__(self, config):
            super(GPT2Model, self).__init__()
    
            # 定义字符嵌入层
            self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size)
    
            # 定义位置嵌入层
            self.position_embeddings = nn.Embedding(config.block_size, config.embedding_size)
    
            # 定义嵌入随机丢弃层
            self.emb_drop = nn.Dropout(config.embedding_dropout)
    
            # 定义 Transformer Encoder
            self.transformer = Transformer(
                config.layer_size,
                config.embedding_size,
                config.num_attention_heads,
                config.attention_dropout,
                config.residual_dropout)
    
        def forward(self, x, kv_cache=None, use_cache=False):
            # 根据缓存确定历史输入长度
            if kv_cache is None:
                past_length = 0
            else:
                past_length = kv_cache[0][0].shape[-2]
    
            # 生成位置编码
            position_ids = torch.arange(past_length, x.shape[-1] + past_length, dtype=torch.int64)
            position_ids = position_ids.unsqueeze(0).expand_as(x)
    
            # 计算嵌入层输出
            x = self.word_embeddings(x)
            x = self.emb_drop(x + self.position_embeddings(position_ids))
    
            # 计算 Transformer Encoder 输出
            x, cached_kvs = self.transformer(x, kv_cache)
    
            # 计算解码输出
            # 解码使用的参数为字符嵌入层参数的转置
            # 相当于做一个逆运算或者可以理解为使用相同的参数进行编码和解码
            x = torch.matmul(x, self.word_embeddings.weight.transpose(-1, -2))
    
            # 如果使用缓存则返回输出和缓存
            if use_cache:
                return x, cached_kvs
    
            # 否则只返回输出
            return x

(2) 数据处理，利用了torch自带的TensorDataset，把形成的Dataset转为Tensor形式。

# -*- coding: utf-8 -*-
    """
    @Time    : 2021/4/17 16:28
    @Author  : SinGaln
    """
    import os
    import torch
    import logging
    from tqdm import tqdm
    from torch.utils.data import TensorDataset
    
    logger = logging.getLogger(__name__)
    
    class DataProcess(object):
        def __init__(self, args):
            self.args = args
            self.data_file = "./data/train.txt"
    
        @classmethod
        def _read_data_file(cls, input_file):
            logger.info("tokenizing raw data,raw data path:{}".format(input_file))
            with open(input_file, 'rb') as f:
                data = f.read().decode("utf-8")
            if "\r\n" in data:
                train_data = data.split("\r\n\r\n")
            else:
                train_data = data.split("\n\n")
            logger.info("there are {} dialogue in raw dataset".format(len(train_data)))
            return train_data
    
        def get_examples(self, tokenizer):
            context = []
            train_data = self._read_data_file(self.data_file)
            for dialogue_index, dialogue in enumerate(tqdm(train_data)):
                utterances = dialogue.split("\n")
                dialogue_ids = [tokenizer.cls_token_id]  # 每个dialogue以[CLS]开头
                for utterance in utterances:
                    dialogue_ids.extend([tokenizer.convert_tokens_to_ids(word) for word in utterance])
                    dialogue_ids.append(tokenizer.sep_token_id)  # 每个utterance之后添加[SEP]，表示utterance结束
                # 对超过n_ctx的长度进行截断,否则GPT2模型会报错
                if len(dialogue_ids) > self.args.max_seq_len:
                    dialogue_ids = dialogue_ids[:self.args.max_seq_len]
                    print("dialogue", len(dialogue_ids))
                else:
                    dialogue_ids = dialogue_ids + ([0] * (self.args.max_seq_len - len(dialogue_ids)))
                context.append(dialogue_ids)
            logger.info("finish processing for raw data!")
            return context
    
    
    processors = {
        
        "chat": DataProcess
    }
    def load_and_cache_examples(args, tokenizer):
        processor = processors[args.task](args)
    
        # Load data features from cache or dataset file
        cached_features_file = os.path.join(
            args.data_dir,
            'cached_train_{}_{}'.format(
                args.task,
                args.max_seq_len
            )
        )
    
        if os.path.exists(cached_features_file):
            logger.info("Loading features from cached file %s", cached_features_file)
            features = torch.load(cached_features_file)
        else:
            # Load data features from dataset file
            logger.info("Creating features from dataset file at %s", args.data_dir)
            features = processor.get_examples(tokenizer)
    
            logger.info("Saving features into cached file %s", cached_features_file)
            torch.save(features, cached_features_file)
    
        # Convert to Tensors and build dataset
        all_input_ids = torch.tensor(features, dtype=torch.long)
    
        dataset = TensorDataset(all_input_ids)
        return dataset

其他的文件可以到小编的[Github][]上查看，注意代码在master分支上。

[Github]: https://github.com/singaln/GPT2-chinese-chatbot/tree/master
[CPM]: https://github.com/TsinghuaAI/CPM-Generate
[pytorch_CPM]: https://pan.baidu.com/s/14aaNd2KEmyrSac2nqoWh6A
[CPM 1]: https://pan.baidu.com/s/1A8hoaacW7B6D5sQQY9VASg
[bert4keras]: https://github.com/bojone/CPM_LM_bert4keras