基于 PyTorch 从零手搓一个GPT Transformer 对话大模型

一、从零手实现 GPT Transformer 模型架构

近年来,大模型的发展势头迅猛,成为了人工智能领域的研究热点。大模型以其强大的语言理解和生成能力,在自然语言处理、机器翻译、文本生成等多个领域取得了显著的成果。但这些都离不开其背后的核心架构------Transformer

Transformer 是一种基于自注意力机制的深度神经网络模型,其核心思想是通过自注意力机制来捕捉序列中的长距离依赖关系。自注意力机制允许模型在处理每个词时,同时考虑序列中的所有其他词,并根据它们之间的关联程度进行加权。这种方法打破了传统循环神经网络(RNN)和长短期记忆网络(LSTM)在处理长序列时的局限性,使得Transformer在处理大规模数据时更加高效。

本文仅使用 PyTorch ,从零构建网络结构、构建词表、训练一个 GPT 对话模型。带你体验如何从01实现一个自定义的对话模型。模型整体以 Transformer Only Decoder 作为核心架构,由多个相同的层堆叠而成,每个层包括自注意力机制、位置编码和前馈神经网络。最终实现效果如下所示:

二、模型搭建

2.1 点积注意力层搭建

注意力的计算公式如下:

首先输入会通过三个不同的线性变换得到三个矩阵,分别是查询(Q)、键(K)和值(V)。

然后,计算 Q 与所有键 K 的点积,得到注意力得分,其中d_k是键向量K的维度。还需要再除以根号下d_k,目的是为了在梯度下降时保持数值稳定性。

然后,将得到的注意力得分通过Softmax函数进行归一化,使得所有得分加起来等于 1。这样,每个得分就变成了一个概率值,表示在当前元素中,其他元素所占的权重。

最后将 Softmax 得到的概率值与值(V)相乘,得到自注意力层的输出。

这里需要注意的是注意力掩码,由于输入序列可能有不同的长度,但矩阵计算时需要固定的大小,因此针对长度不足的序列,可以使用 padding 作为填充标记,但这些 padding的信息是没有意义的,计算注意力分数也没有意义,因此可以将 padding 位置的分数置为非常小,后续计算 softmax 之后基本就是 0 了。

实现过程如下:

python 复制代码
class ScaledDotProductAttention(nn.Module):
    def __init__(self, d_k):
        super(ScaledDotProductAttention, self).__init__()
        self.d_k = d_k

    def forward(self, q, k, v, attention_mask):
        ##
        # q: [batch_size, n_heads, len_q, d_k]
        # k: [batch_size, n_heads, len_k, d_k]
        # v: [batch_size, n_heads, len_v, d_v]
        # attn_mask: [batch_size, n_heads, seq_len, seq_len]
        ##
        # 计算每个Q与K的分数,计算出来的大小是 [batch_size, n_heads, len_q, len_q]
        scores = torch.matmul(q, k.transpose(-1, -2)) / np.sqrt(self.d_k)
        # 把被mask的地方置为无限小,softmax之后基本就是0,也就对q不起作用
        scores.masked_fill_(attention_mask, -1e9)
        attn = nn.Softmax(dim=-1)(scores)
        # 注意力后的大小 [batch_size, n_heads, len_q, d_v]
        context = torch.matmul(attn, v)
        return context, attn

2.2 多头注意力层搭建

多头注意力层在单头注意力层的基础上,主要将Q、K、V拆分成多个头,然后并行的处理,每个头可以学习序列的不同特征,增强模型的特征提取能力。

多头注意力层的输出是多个头输出的拼接,通过一个线性层转换成和输入相同的序列,然后再和原始值相加构成残差,最后由 LN 归一化后输出。

实现过程如下:

python 复制代码
class MultiHeadAttention(nn.Module):
    def __init__(self, d_model, n_heads, d_k, d_v):
        super(MultiHeadAttention, self).__init__()
        self.d_model = d_model
        self.n_heads = n_heads
        self.d_k = d_k
        self.d_v = d_v
        self.w_q = nn.Linear(d_model, d_k * n_heads, bias=False)
        self.w_k = nn.Linear(d_model, d_k * n_heads, bias=False)
        self.w_v = nn.Linear(d_model, d_v * n_heads, bias=False)
        self.fc = nn.Linear(n_heads * d_v, d_model, bias=False)
        self.layernorm = nn.LayerNorm(d_model)

    def forward(self, q, k, v, attention_mask):
        ##
        # q: [batch_size, seq_len, d_model]
        # k: [batch_size, seq_len, d_model]
        # v: [batch_size, seq_len, d_model]
        # attn_mask: [batch_size, seq_len, seq_len]
        ##
        # 记录原始值, 后续计算残差
        residual, batch_size = q, q.size(0)
        # 先映射 q、k、v, 然后后分头
        # q: [batch_size, n_heads, len_q, d_k]
        q = self.w_q(q).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
        # k: [batch_size, n_heads, len_k, d_k]
        k = self.w_k(k).view(batch_size, -1, self.n_heads, self.d_k).transpose(1, 2)
        # v: [batch_size, n_heads, len_v(=len_k), d_v]
        v = self.w_v(v).view(batch_size, -1, self.n_heads, self.d_v).transpose(1, 2)
        # attn_mask : [batch_size, n_heads, seq_len, seq_len]
        attention_mask = attention_mask.unsqueeze(1).repeat(1, self.n_heads, 1, 1)
        # 点积注意力分数计算,  [batch_size, n_heads, len_q, d_v]
        context, attn = ScaledDotProductAttention(self.d_k)(q, k, v, attention_mask)
        # context: [batch_size, len_q, n_heads * d_v]
        context = context.transpose(1, 2).reshape(batch_size, -1, self.n_heads * self.d_v)
        # 还原为原始大小
        output = self.fc(context)
        # LN + 残差计算
        return self.layernorm(output + residual), attn

2.3 前馈神经网络层搭建

前馈神经网络层,组成比较简单,由两个线性全连接层组成,中间使用 ReLU 激活函数衔接,主要在做一个升维再降维的操作,可以学习到更为抽象的特征。

实现过程如下:

python 复制代码
class PoswiseFeedForwardNet(nn.Module):
    def __init__(self, d_model, d_ff):
        super(PoswiseFeedForwardNet, self).__init__()
        self.fc = nn.Sequential(
            nn.Linear(d_model, d_ff, bias=False),
            nn.ReLU(),
            nn.Linear(d_ff, d_model, bias=False)
        )
        self.layernorm = nn.LayerNorm(d_model)

    def forward(self, inputs):
        ##
        # inputs: [batch_size, seq_len, d_model]
        ##
        residual = inputs
        output = self.fc(inputs)
        # # LN + 残差计算, [batch_size, seq_len, d_model]
        return self.layernorm(output + residual)

2.4 解码层构建

上面有了多头注意力机制和前馈神经网络层后,这里就可以构建解码层了,一个解码层由一个多头注意力层和一个前馈神经网络层组成。

实现过程如下:

python 复制代码
class DecoderLayer(nn.Module):
    def __init__(self, d_model, n_heads, d_ff, d_k, d_v):
        super(DecoderLayer, self).__init__()
        # 多头注意力层
        self.attention = MultiHeadAttention(d_model, n_heads, d_k, d_v)
        # 前馈神经网络层
        self.pos_ffn = PoswiseFeedForwardNet(d_model, d_ff)

    def forward(self, inputs, attention_mask):
        ##
        # inputs: [batch_size, seq_len, d_model]
        # attention_mask: [batch_size, seq_len, seq_len]
        ##
        # outputs: [batch_size, seq_len, d_model]
        # self_attn: [batch_size, n_heads, seq_len, seq_len]
        outputs, self_attn = self.attention(inputs, inputs, inputs, attention_mask)
        # [batch_size, seq_len, d_model]
        outputs = self.pos_ffn(outputs)
        return outputs, self_attn

2.5 解码器构建

解码器主要将多个解码层堆叠,形成一个特征提取链路。首先解码器接收输入的 Token,然后通过 Embedding 转为高维向量表示,由于注意力机制没有位置信息,因此这里还需要加上位置编码。

位置编码这里参照 GPT2 的做法,直接对位置再次进行 Embedding。这里你也可以换成固定位置编码、旋转位置编码进行实验。

python 复制代码
class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_pos, device):
        super(PositionalEncoding, self).__init__()
        self.device = device
        self.pos_embedding = nn.Embedding(max_pos, d_model)

    def forward(self, inputs):
        seq_len = inputs.size(1)
        pos = torch.arange(seq_len, dtype=torch.long, device=self.device)
        # [seq_len] -> [batch_size, seq_len]
        pos = pos.unsqueeze(0).expand_as(inputs)
        return self.pos_embedding(pos)

对于 Transformer Decoder 结构,模型在解码时应该是自回归的,每次都是基于之前的信息预测下一个Token,这意味着在生成序列的第 i 个元素时,模型只能看到位置 i 之前的信息。因此在训练时需要进行遮盖,防止模型看到未来的信息,遮盖的操作也非常简单,可以构建一个上三角掩码器。

例如:

shell 复制代码
原始注意力分数矩阵(无掩码):
[[q1k1, q1k2, q1k3, q1k4],
 [q2k1, q2k2, q3k3, q3k4],
 [q3k1, q3k2, q3k3, q3k4],
 [q4k1, q4k2, q4k3, q4k4]]

上三角掩码器:
[[0, 1, 1, 1],
 [0, 0, 1, 1],
 [0, 0, 0, 1],
 [0, 0, 0, 0]]

应用掩码后的分数矩阵:
[[q1k1, -inf, -inf, -inf],
 [q2k1, q2k2, -inf, -inf],
 [q3k1, q3k2, q3k3, -inf],
 [q4k1, q4k2, q4k3, q4k4]]

实现过程:

python 复制代码
def get_attn_subsequence_mask(seq, device):
    # 注意力分数的大小是 [batch_size, n_heads, len_seq, len_seq]
    # 所以这里要生成 [batch_size, len_seq, len_seq] 大小
    attn_shape = [seq.size(0), seq.size(1), seq.size(1)]
    # 生成一个上三角矩阵
    subsequence_mask = np.triu(np.ones(attn_shape), k=1)
    subsequence_mask = torch.from_numpy(subsequence_mask).byte()
    subsequence_mask = subsequence_mask.to(device)
    return subsequence_mask

attention_mask 的掩码大小调整,要转换成 [batch_size, len_seq, len_seq] 大小,方便和注意力分数计算:

python 复制代码
def get_attn_pad_mask(attention_mask):
    batch_size, len_seq = attention_mask.size()
    attention_mask = attention_mask.data.eq(0).unsqueeze(1)
    # 注意力分数的大小是 [batch_size, n_heads, len_q, len_q]
    # 所以这里要转换成 [batch_size, len_seq, len_seq] 大小
    return attention_mask.expand(batch_size, len_seq, len_seq)

到这就可以构建解码器了,实现过程:

python 复制代码
class Decoder(nn.Module):
    def __init__(self, d_model, n_heads, d_ff, d_k, d_v, vocab_size, max_pos, n_layers, device):
        super(Decoder, self).__init__()
        self.device = device
        # 将Token转为向量
        self.embedding = nn.Embedding(vocab_size, d_model)
        # 位置编码
        self.pos_encoding = PositionalEncoding(d_model, max_pos, device)
        self.layers = nn.ModuleList([DecoderLayer(d_model, n_heads, d_ff, d_k, d_v) for _ in range(n_layers)])

    def forward(self, inputs, attention_mask):
        ##
        # inputs: [batch_size, seq_len]
        ##
        # [batch_size, seq_len, d_model]
        outputs = self.embedding(inputs) + self.pos_encoding(inputs)
        # 上三角掩码,防止看到未来的信息, [batch_size, seq_len, seq_len]
        subsequence_mask = get_attn_subsequence_mask(inputs, self.device)
        if attention_mask is not None:
            # pad掩码 [batch_size, seq_len, seq_len]
            attention_mask = get_attn_pad_mask(attention_mask)
            # [batch_size, seq_len, seq_len]
            attention_mask = torch.gt((attention_mask + subsequence_mask), 0)
        else:
            attention_mask = subsequence_mask.bool()
        # 计算每一层的结果
        self_attns = []
        for layer in self.layers:
            # outputs: [batch_size, seq_len, d_model],
            # self_attn: [batch_size, n_heads, seq_len, seq_len],
            outputs, self_attn = layer(outputs, attention_mask)
            self_attns.append(self_attn)
        return outputs, self_attns

2.6 构建GPT模型

上面构建好解码器之后,就可以得到处理后的特征,下面还需要将特征转为词表大小的概率分布,才能实现对下一个Token的预测。

实现过程:

python 复制代码
class GPTModel(nn.Module):
    def __init__(self, d_model, n_heads, d_ff, d_k, d_v, vocab_size, max_pos, n_layers, device):
        super(GPTModel, self).__init__()
        # 解码器
        self.decoder = Decoder(d_model, n_heads, d_ff, d_k, d_v, vocab_size, max_pos, n_layers, device)
        # 映射为词表大小
        self.projection = nn.Linear(d_model, vocab_size)

    def forward(self, inputs, attention_mask=None):
        ##
        # inputs: [batch_size, seq_len]
        ##
        # outputs: [batch_size, seq_len, d_model]
        # self_attns: [n_layers, batch_size, n_heads, seq_len, seq_len]
        outputs, self_attns = self.decoder(inputs, attention_mask)
        # [batch_size, seq_len, vocab_size]
        logits = self.projection(outputs)
        return logits.view(-1, logits.size(-1)), self_attns

到此整个的 GPT 模型也就搭建好了,可以打印看下网络结构,以及模型参数量:

python 复制代码
import torch

from model import GPTModel


def main():
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    # 模型参数
    model_param = {
        "d_model": 768,  # 嵌入层大小
        "d_ff": 2048,  # 前馈神经网络大小
        "d_k": 64,  # K 的大小
        "d_v": 64,  # V 的大小
        "n_layers": 6,  # 解码层的数量
        "n_heads": 8,  # 多头注意力的头数
        "max_pos": 1800,  # 位置编码的长度
        "device": device,  # 设备
        "vocab_size": 4825  # 词表大小
    }
    model = GPTModel(**model_param)
    total_params = sum(p.numel() for p in model.parameters())
    print(model)
    print("total_params: ", total_params)


if __name__ == '__main__':
    main()

网络结构:

yml 复制代码
GPTModel(
  (decoder): Decoder(
    (embedding): Embedding(4825, 768)
    (pos_encoding): PositionalEncoding(
      (pos_embedding): Embedding(1800, 768)
    )
    (layers): ModuleList(
      (0): DecoderLayer(
        (attention): MultiHeadAttention(
          (w_q): Linear(in_features=768, out_features=512, bias=False)
          (w_k): Linear(in_features=768, out_features=512, bias=False)
          (w_v): Linear(in_features=768, out_features=512, bias=False)
          (fc): Linear(in_features=512, out_features=768, bias=False)
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
        (pos_ffn): PoswiseFeedForwardNet(
          (fc): Sequential(
            (0): Linear(in_features=768, out_features=2048, bias=False)
            (1): ReLU()
            (2): Linear(in_features=2048, out_features=768, bias=False)
          )
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
      )
      (1): DecoderLayer(
        (attention): MultiHeadAttention(
          (w_q): Linear(in_features=768, out_features=512, bias=False)
          (w_k): Linear(in_features=768, out_features=512, bias=False)
          (w_v): Linear(in_features=768, out_features=512, bias=False)
          (fc): Linear(in_features=512, out_features=768, bias=False)
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
        (pos_ffn): PoswiseFeedForwardNet(
          (fc): Sequential(
            (0): Linear(in_features=768, out_features=2048, bias=False)
            (1): ReLU()
            (2): Linear(in_features=2048, out_features=768, bias=False)
          )
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
      )
      (2): DecoderLayer(
        (attention): MultiHeadAttention(
          (w_q): Linear(in_features=768, out_features=512, bias=False)
          (w_k): Linear(in_features=768, out_features=512, bias=False)
          (w_v): Linear(in_features=768, out_features=512, bias=False)
          (fc): Linear(in_features=512, out_features=768, bias=False)
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
        (pos_ffn): PoswiseFeedForwardNet(
          (fc): Sequential(
            (0): Linear(in_features=768, out_features=2048, bias=False)
            (1): ReLU()
            (2): Linear(in_features=2048, out_features=768, bias=False)
          )
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
      )
      (3): DecoderLayer(
        (attention): MultiHeadAttention(
          (w_q): Linear(in_features=768, out_features=512, bias=False)
          (w_k): Linear(in_features=768, out_features=512, bias=False)
          (w_v): Linear(in_features=768, out_features=512, bias=False)
          (fc): Linear(in_features=512, out_features=768, bias=False)
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
        (pos_ffn): PoswiseFeedForwardNet(
          (fc): Sequential(
            (0): Linear(in_features=768, out_features=2048, bias=False)
            (1): ReLU()
            (2): Linear(in_features=2048, out_features=768, bias=False)
          )
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
      )
      (4): DecoderLayer(
        (attention): MultiHeadAttention(
          (w_q): Linear(in_features=768, out_features=512, bias=False)
          (w_k): Linear(in_features=768, out_features=512, bias=False)
          (w_v): Linear(in_features=768, out_features=512, bias=False)
          (fc): Linear(in_features=512, out_features=768, bias=False)
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
        (pos_ffn): PoswiseFeedForwardNet(
          (fc): Sequential(
            (0): Linear(in_features=768, out_features=2048, bias=False)
            (1): ReLU()
            (2): Linear(in_features=2048, out_features=768, bias=False)
          )
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
      )
      (5): DecoderLayer(
        (attention): MultiHeadAttention(
          (w_q): Linear(in_features=768, out_features=512, bias=False)
          (w_k): Linear(in_features=768, out_features=512, bias=False)
          (w_v): Linear(in_features=768, out_features=512, bias=False)
          (fc): Linear(in_features=512, out_features=768, bias=False)
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
        (pos_ffn): PoswiseFeedForwardNet(
          (fc): Sequential(
            (0): Linear(in_features=768, out_features=2048, bias=False)
            (1): ReLU()
            (2): Linear(in_features=2048, out_features=768, bias=False)
          )
          (layernorm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
        )
      )
    )
  )
  (projection): Linear(in_features=768, out_features=4825, bias=True)
  
total_params:  37128409

可以看到参数量只有三千七百多万,我们这个只能算个小号的对话模型。

下面开始基于数据集构建词表。

三、数据集词表构建

数据集使用对话-百科(中文)训练集 ,有 274148 条问答对信息,涵盖了 美食、城市、企业家、汽车、明星八卦、生活常识、日常对话 等信息。

数据集下载地址:

https://modelscope.cn/datasets/qiaojiedongfeng/qiaojiedongfeng/summary

数据格式如下所示:

json 复制代码
{"question": "你在阐述观点时能否提供一些具体的论据或实例,以便我更容易识别潜在的弱点或反证点?", "answer": "当然可以!当我提出一个观点时,我会举例说明,并引用相关数据、研究或经验来支持我的观点。这样做可以帮助识别可能存在的反对意见或反驳的途径。"}
{"question": "你最近有没有阅读任何书籍?", "answer": "最近我在读《人类简史》,探索历史的演变和人类社会的发展。"}
{"question": "哪种具有特定特性的常见物品经常用于储存和分发水?", "answer": "塑料瓶"}
{"question": "请问北京市的地理覆盖范围是哪些区域?", "answer": "北京市包括东城区、西城区、朝阳区、丰台区、石景山区、海淀区、门头沟区、房山区、通州区、顺义区、昌平区、大兴区、怀柔区、平谷区、密云区和延庆区。"}
{"question": "你能分享一些关于刘德华的有趣故事吗?", "answer": "当然可以!有一次,在拍摄电影《无间道》时,刘德华为了完美呈现角色,长时间沉浸在戏中无法自拔,结果在现实生活中也展现出了一种'失忆'的状态。他还曾在颁奖典礼上不慎摔伤了手指,但仍然坚持完成表演,这让他获得了'坚强艺人'的称号。"}
{"question": "你能给我一些建议,有哪些美食值得一试?","answer": "尝试日本寿司、意大利披萨、法国鹅肝、泰国绿咖喱和韩国石锅拌饭,每种都有独特的风味,令人回味无穷。"}
{"question": "谁是小说《悲惨世界》的创作者?", "answer": "维克多·雨果"}

构建词表,这里我将一个字作为一个词,也可以优化通过分词器分词后的词构建词表,需要注意的时,词表需要拼接三个特殊Token,用于表示特殊意义: pad 占位、unk 未知、sep 结束

python 复制代码
import json

def build_vocab(file_path):
    # 读取所有文本
    texts = []
    with open(file_path, 'r', encoding='utf-8') as r:
        for line in r:
            if not line:
                continue
            line = json.loads(line)
            question = line["question"]
            answer = line["answer"]
            texts.append(question)
            texts.append(answer)
    # 拆分 Token
    words = set()
    for t in texts:
        if not t:
            continue
        for word in t.strip():
            words.add(word)
    words = list(words)
    words.sort()
    # 特殊Token
    # pad 占位、unk 未知、sep 结束
    word2id = {"<pad>": 0, "<unk>": 1, "<sep>": 2}
    # 构建词表
    word2id.update({word: i + len(word2id) for i, word in enumerate(words)})
    id2word = list(word2id.keys())
    vocab = {"word2id": word2id, "id2word": id2word}
    vocab = json.dumps(vocab, ensure_ascii=False)
    with open('data/vocab.json', 'w', encoding='utf-8') as w:
        w.write(vocab)
    print(f"finish. words: {len(id2word)}")

if __name__ == '__main__':
    build_vocab("data/train.jsonl")

处理后词表的大小是 4825 ,格式如下所示:

下面构建一个 Tokenizer 类,方便后续训练和预测时处理 Token

python 复制代码
import json

class Tokenizer():

    def __init__(self, vocab_path):
        with open(vocab_path, "r", encoding="utf-8") as r:
            vocab = r.read()
            if not vocab:
                raise Exception("词表读取为空!")
        vocab = json.loads(vocab)
        self.word2id = vocab["word2id"]
        self.id2word = vocab["id2word"]
        self.pad_token = self.word2id["<pad>"]
        self.unk_token = self.word2id["<unk>"]
        self.sep_token = self.word2id["<sep>"]

    def encode(self, text, text1=None, max_length=128, pad_to_max_length=False):
        tokens = [self.word2id[word] if word in self.word2id else self.unk_token for word in text]
        tokens.append(self.sep_token)
        if text1:
            tokens.extend([self.word2id[word] if word in self.word2id else self.unk_token for word in text1])
            tokens.append(self.sep_token)
        att_mask = [1] * len(tokens)
        if pad_to_max_length:
            if len(tokens) > max_length:
                tokens = tokens[0:max_length]
                att_mask = att_mask[0:max_length]
            elif len(tokens) < max_length:
                tokens.extend([self.pad_token] * (max_length - len(tokens)))
                att_mask.extend([0] * (max_length - len(att_mask)))
        return tokens, att_mask

    def decode(self, token):
        if type(token) is tuple or type(token) is list:
            return [self.id2word[n] for n in token]
        else:
            return self.id2word[token]

    def get_vocab_size(self):
        return len(self.id2word)

使用示例:

python 复制代码
if __name__ == '__main__':
    tokenizer = Tokenizer(vocab_path="data/vocab.json")
    encode, att_mask = tokenizer.encode("你好,小毕超", "你好,小毕超", pad_to_max_length=True)
    decode = tokenizer.decode(encode)
    print("token lens: ", len(encode))
    print("encode: ", encode)
    print("att_mask: ", att_mask)
    print("decode: ", decode)
    print("vocab_size", tokenizer.get_vocab_size())

有了词表后,就可以规划训练和验证数据集了,前面构建模型时,我们的参数量只有 三千七百多万,连 0.1 B都不到,训练这二十七万多条知识,缺失有点牵强,而且还是从零随机初始化参数训练,因此为了快速实验,这里取前 10000 条数据作为训练,1000 条数据验证,从而快速实验效果:

python 复制代码
import os.path

def split_dataset(file_path, output_path):
    if not os.path.exists(output_path):
        os.mkdir(output_path)
    datas = []
    with open(file_path, "r", encoding='utf-8') as f:
        for line in f:
            if not line or line == "":
                continue
            datas.append(line)
    train = datas[0:10000]
    val = datas[10000:11000]
    with open(os.path.join(output_path, "train.json"), "w", encoding="utf-8") as w:
        for line in train:
            w.write(line)
            w.flush()

    with open(os.path.join(output_path, "val.json"), "w", encoding="utf-8") as w:
        for line in val:
            w.write(line)
            w.flush()
    print("train count: ", len(train))
    print("val count: ", len(val))


if __name__ == '__main__':
    file_path = "data/train.jsonl"
    split_dataset(file_path=file_path, output_path="data")

为了增加自定义模型的特色,这里在训练集中追加几条身份的数据在里面:

json 复制代码
{"question": "你是谁", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你叫什么", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你的名字是什么", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你叫啥", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你名字是啥", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你是什么身份", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你的全名是什么", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你自称什么", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你的称号是什么", "answer": "我是小毕超,一个简易的小助手"}
{"question": "你的昵称是什么", "answer": "我是小毕超,一个简易的小助手"}

看一下 train.json 的数据Token数量分布情况,确定一下 max_token 大小:

python 复制代码
import json
from tokenizer import Tokenizer
import matplotlib.pyplot as plt

plt.rcParams['font.sans-serif'] = ['SimHei']

def get_num_tokens(file_path, tokenizer):
    input_num_tokens = []
    with open(file_path, "r", encoding="utf-8") as r:
        for line in r:
            line = json.loads(line)
            question = line["question"]
            answer = line["answer"]
            tokens, att_mask = tokenizer.encode(question, answer)
            input_num_tokens.append(len(tokens))
    return input_num_tokens

def count_intervals(num_tokens, interval):
    max_value = max(num_tokens)
    intervals_count = {}
    for lower_bound in range(0, max_value + 1, interval):
        upper_bound = lower_bound + interval
        count = len([num for num in num_tokens if lower_bound <= num < upper_bound])
        intervals_count[f"{lower_bound}-{upper_bound}"] = count
    return intervals_count

def main():
    train_data_path = "data/train.json"
    tokenizer = Tokenizer("data/vocab.json")
    input_num_tokens = get_num_tokens(train_data_path, tokenizer)
    intervals_count = count_intervals(input_num_tokens, 20)
    print(intervals_count)
    x = [k for k, v in intervals_count.items()]
    y = [v for k, v in intervals_count.items()]
    plt.figure(figsize=(8, 6))
    bars = plt.bar(x, y)
    plt.title('训练集Token分布情况')
    plt.ylabel('数量')
    plt.xticks(rotation=45)
    for bar in bars:
        yval = bar.get_height()
        plt.text(bar.get_x() + bar.get_width() / 2, yval, int(yval), va='bottom')
    plt.show()

if __name__ == '__main__':
    main()

可以看出数据集主要分布在120以内,因此后面训练时,max_length 设为 120 可以覆盖大多数的信息。

四、模型训练

4.1 构建 Dataset

qa_dataset.py

python 复制代码
# -*- coding: utf-8 -*-
from torch.utils.data import Dataset
import torch
import json
import numpy as np


class QADataset(Dataset):
    def __init__(self, data_path, tokenizer, max_length) -> None:
        super().__init__()
        self.tokenizer = tokenizer
        self.max_length = max_length
        self.data = []
        if data_path:
            with open(data_path, "r", encoding='utf-8') as f:
                for line in f:
                    if not line or line == "":
                        continue
                    json_line = json.loads(line)
                    question = json_line["question"]
                    answer = json_line["answer"]
                    self.data.append({
                        "question": question,
                        "answer": answer
                    })
        print("data load , size:", len(self.data))

    def preprocess(self, question, answer):
        encode, att_mask = self.tokenizer.encode(question, answer, max_length=self.max_length, pad_to_max_length=True)
        input_ids = encode[:-1]
        att_mask = att_mask[:-1]
        labels = encode[1:]
        return input_ids, att_mask, labels

    def __getitem__(self, index):
        item_data = self.data[index]
        input_ids, att_mask, labels = self.preprocess(**item_data)
        return {
            "input_ids": torch.LongTensor(np.array(input_ids)),
            "attention_mask": torch.LongTensor(np.array(att_mask)),
            "labels": torch.LongTensor(np.array(labels))
        }

    def __len__(self):
        return len(self.data)

4.2 训练

python 复制代码
# -*- coding: utf-8 -*-
import torch
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from tokenizer import Tokenizer
from model import GPTModel
from qa_dataset import QADataset
from tqdm import tqdm
import time, sys, os


def train_model(model, train_loader, val_loader, optimizer, criterion,
                device, num_epochs, model_output_dir, writer):
    batch_step = 0
    best_val_loss = float('inf')
    for epoch in range(num_epochs):
        time1 = time.time()
        model.train()
        for index, data in enumerate(tqdm(train_loader, file=sys.stdout, desc="Train Epoch: " + str(epoch))):
            input_ids = data['input_ids'].to(device, dtype=torch.long)
            attention_mask = data['attention_mask'].to(device, dtype=torch.long)
            labels = data['labels'].to(device, dtype=torch.long)
            optimizer.zero_grad()
            outputs, dec_self_attns = model(input_ids, attention_mask)
            loss = criterion(outputs, labels.view(-1))
            loss.backward()
            # 梯度裁剪
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
            optimizer.step()
            writer.add_scalar('Loss/train', loss, batch_step)
            batch_step += 1
            # 100轮打印一次 loss
            if index % 100 == 0 or index == len(train_loader) - 1:
                time2 = time.time()
                tqdm.write(
                    f"{index}, epoch: {epoch} -loss: {str(loss)} ; lr: {optimizer.param_groups[0]['lr']} ;each step's time spent: {(str(float(time2 - time1) / float(index + 0.0001)))}")
        # 验证
        model.eval()
        val_loss = validate_model(model, criterion, device, val_loader)
        writer.add_scalar('Loss/val', val_loss, epoch)
        print(f"val loss: {val_loss} , epoch: {epoch}")
        # 保存最优模型
        if val_loss < best_val_loss:
            best_val_loss = val_loss
            best_model_path = os.path.join(model_output_dir, "best.pt")
            print("Save Best Model To ", best_model_path, ", epoch: ", epoch)
            torch.save(model.state_dict(), best_model_path)
        # 保存当前模型
        last_model_path = os.path.join(model_output_dir, "last.pt")
        print("Save Last Model To ", last_model_path, ", epoch: ", epoch)
        torch.save(model.state_dict(), last_model_path)


def validate_model(model, criterion, device, val_loader):
    running_loss = 0.0
    with torch.no_grad():
        for _, data in enumerate(tqdm(val_loader, file=sys.stdout, desc="Validation Data")):
            input_ids = data['input_ids'].to(device, dtype=torch.long)
            attention_mask = data['attention_mask'].to(device, dtype=torch.long)
            labels = data['labels'].to(device, dtype=torch.long)
            outputs, dec_self_attns = model(input_ids, attention_mask)
            loss = criterion(outputs, labels.view(-1))
            running_loss += loss.item()
    return running_loss / len(val_loader)


def main():
    train_json_path = "data/train.json"  # 训练集
    val_json_path = "data/val.json"  # 验证集
    vocab_path = "data/vocab.json"  # 词表位置
    max_length = 120  # 最大长度
    epochs = 15 # 迭代周期
    batch_size = 128  # 训练一个批次的大小
    lr = 1e-4  # 学习率
    model_output_dir = "output"  # 模型保存目录
    logs_dir = "logs"  # 日志记录目标
    # 设备
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    # 加载分词器
    tokenizer = Tokenizer(vocab_path)
    # 模型参数
    model_param = {
        "d_model": 768,  # 嵌入层大小
        "d_ff": 2048,  # 前馈神经网络大小
        "d_k": 64,  # K 的大小
        "d_v": 64,  # V 的大小
        "n_layers": 6,  # 解码层的数量
        "n_heads": 8,  # 多头注意力的头数
        "max_pos": 1800,  # 位置编码的长度
        "device": device,  # 设备
        "vocab_size": tokenizer.get_vocab_size(),  # 词表大小
    }
    model = GPTModel(**model_param)
    print("Start Load Train Data...")
    train_params = {
        "batch_size": batch_size,
        "shuffle": True,
        "num_workers": 4,
    }
    training_set = QADataset(train_json_path, tokenizer, max_length)
    training_loader = DataLoader(training_set, **train_params)
    print("Start Load Validation Data...")
    val_params = {
        "batch_size": batch_size,
        "shuffle": False,
        "num_workers": 4,
    }
    val_set = QADataset(val_json_path, tokenizer, max_length)
    val_loader = DataLoader(val_set, **val_params)
    # 日志记录
    writer = SummaryWriter(logs_dir)
    # 优化器
    optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
    # 损失函数
    criterion = torch.nn.CrossEntropyLoss(ignore_index=0).to(device)
    model = model.to(device)
    # 开始训练
    print("Start Training...")
    train_model(
        model=model,
        train_loader=training_loader,
        val_loader=val_loader,
        optimizer=optimizer,
        criterion=criterion,
        device=device,
        num_epochs=epochs,
        model_output_dir=model_output_dir,
        writer=writer
    )
    writer.close()


if __name__ == '__main__':
    main()

训练过程:

batch size 128 下训练大概仅占用 7G 显存:

训练结果后使用 tensorboard 查看下 loss 趋势:

在训练 15epochs 情况下, 训练集 loss 降到1.31左右,验证集 loss 最低降到了 3.16左右。

下面对模型预测下对话的效果。

五、模型预测

python 复制代码
import torch

from model import GPTModel
from tokenizer import Tokenizer


def generate(model, tokenizer, text, max_length, device):
    input, att_mask = tokenizer.encode(text)
    input = torch.tensor(input, dtype=torch.long, device=device).unsqueeze(0)
    stop = False
    input_len = len(input[0])
    while not stop:
        if len(input[0]) - input_len > max_length:
            next_symbol = tokenizer.sep_token
            input = torch.cat(
                [input.detach(), torch.tensor([[next_symbol]], dtype=input.dtype, device=device)], -1)
            break
        projected, self_attns = model(input)
        prob = projected.squeeze(0).max(dim=-1, keepdim=False)[1]
        next_word = prob.data[-1]
        next_symbol = next_word
        if next_symbol == tokenizer.sep_token:
            stop = True
        input = torch.cat(
            [input.detach(), torch.tensor([[next_symbol]], dtype=input.dtype, device=device)], -1)
    decode = tokenizer.decode(input[0].tolist())
    decode = decode[len(text):]
    return "".join(decode)


def main():
    model_path = "output/best.pt"
    vocab_path = "data/vocab.json"  # 词表位置
    max_length = 128  # 最大长度
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    # 加载分词器
    tokenizer = Tokenizer(vocab_path)
    # 模型参数
    model_param = {
        "d_model": 768,  # 嵌入层大小
        "d_ff": 2048,  # 前馈神经网络大小
        "d_k": 64,  # K 的大小
        "d_v": 64,  # V 的大小
        "n_layers": 6,  # 解码层的数量
        "n_heads": 8,  # 多头注意力的头数
        "max_pos": 1800,  # 位置编码的长度
        "device": device,  # 设备
        "vocab_size": tokenizer.get_vocab_size(),  # 词表大小
    }
    model = GPTModel(**model_param)
    model.load_state_dict(torch.load(model_path))
    model.to(device)

    while True:
        text = input("请输入:")
        if not text:
            continue
        if text == "q":
            break
        res = generate(model, tokenizer, text, max_length, device)
        print("AI: ", res)


if __name__ == '__main__':
    main()

预测效果:

相关推荐
布鲁格若门4 小时前
AMD CPU下pytorch 多GPU运行卡死和死锁解决
人工智能·pytorch·python·nvidia
小锋学长生活大爆炸5 小时前
【教程】Cupy、Numpy、Torch互相转换
pytorch·numpy·cupy
CSDN云计算7 小时前
性能高于Transformer模型1.7-2倍,彩云科技发布基于DCFormer架构通用大模型云锦天章
大模型·transformer·dcformer·彩云天气·彩云小梦·云锦天章
智匠MindCraft Al8 小时前
GPT o1 模型使用及API调用
人工智能·gpt·ai·语言模型·chatgpt
一者仁心8 小时前
【AI技术】GPT-SoVits训练日志
人工智能·gpt
JarodYv8 小时前
GPT-5 要来了:抢先了解其创新突破
gpt·openai·生成式ai·gpt-4·gpt-5
灰哥数据智能9 小时前
DB-GPT系列(五):DB-GPT六大基础应用场景part2
数据库·人工智能·python·sql·gpt·abi
qq_2739002310 小时前
torch.stack 张量维度的变化
人工智能·pytorch·深度学习
啊文师兄11 小时前
使用 Pytorch 搭建视频车流量检测资源(基于YOLO)
人工智能·pytorch·yolo
使者大牙11 小时前
【LLM学习笔记】第三篇:模型微调及LoRA介绍(附PyTorch实例)
人工智能·pytorch·python·深度学习