Transformer模型代码（详细注释，适合新手）

# Hyperparameters
batch_size = 4  # How many batches per training step
context_length = 16  # Length of the token chunk each batch
d_model = 64  # The size of our model token embeddings
num_blocks = 8  # Number of transformer blocks
num_heads = 4  # Number of heads in Multi-head attention
learning_rate = 1e-3  # 0.001
dropout = 0.1  # Dropout rate
max_iters = 5000  # Total of training iterations <- Change this to smaller number for testing
eval_interval = 50  # How often to evaluate
eval_iters = 20  # Number of iterations to average for evaluation
device = 'cuda' if torch.cuda.is_available() else 'cpu'  # Use GPU if it's available.
TORCH_SEED = 1337
torch.manual_seed(TORCH_SEED)


class FeedForward(nn.Module):
    def __init__(self):
        super().__init__()
        # 模型维度
        self.d_model = d_model
        # 丢弃率
        self.dropout = dropout
        # 第一个线性层
        self.ln1 = nn.Linear(in_features=self.d_model, out_features=self.d_model * 4)
        # ReLU激活函数
        self.relu = nn.ReLU()
        # 第二个线性层
        self.ln2 = nn.Linear(in_features=self.d_model * 4, out_features=self.d_model)
        # 丢弃层
        self.dp = nn.Dropout(dropout)
    def forward(self, x):
        # 输入形状为 batch_size, seq_len, d_model
        x = self.ln1(x)
        x = self.relu(x)
        x = self.ln2(x)
        out = self.dp(x)
        return out




class Attention(nn.Module):
    def __init__(self, head_size):
        """
        参数:
            head_size (int): 每个注意力头的大小。
            d_model (int): 输入张量的特征维度大小。
            context_length (int): 上下文长度，即时间步数。
            dropout (float): Dropout 层的丢弃率。
        """
        super().__init__()
        # 设置模型的参数
        self.d_model = d_model
        self.head_size = head_size
        self.context_length = context_length
        self.dropout = dropout
        # 定义用于计算注意力权重的线性层
        self.key_layer = nn.Linear(in_features=self.d_model, out_features=self.head_size, bias=False)
        self.query_layer = nn.Linear(in_features=self.d_model, out_features=self.head_size, bias=False)
        self.value_layer = nn.Linear(in_features=self.d_model, out_features=self.head_size, bias=False)
        # 生成下三角掩码
        self.register_buffer('tril', torch.tril(torch.ones((self.context_length, self.context_length))))
        # Dropout 层，用于防止过拟合
        self.dropout_layer = nn.Dropout(self.dropout)

    def forward(self, x):
        #todo 输入输出维度一样

        # 获取输入张量的形状信息
        B, T, C = x.shape  # Batch size, Time steps(current context_length), Channels(dimensions)
        # 确保时间步数不超过上下文长度，且特征维度与模型参数匹配
        assert T <= self.context_length
        assert C == self.d_model
        # 通过线性层得到查询、键和值张量
        q = self.query_layer(x)
        k = self.key_layer(x)
        v = self.value_layer(x)

        # 缩放点积注意力：Q @ K^T / sqrt(d_k)
        weights = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
        # 应用掩码
        weights = weights.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
        weights = F.softmax(input=weights, dim=-1)
        weights = self.dropout_layer(weights)

        # 加权求和：weights @ V
        out = weights @ v
        return out


class MultiHeadAttention(nn.Module):
    def __init__(self, head_size):
        """
        定义一个多头注意力机制模型。

        参数:
            num_heads (int): 注意力头的数量。
            head_size (int): 每个注意力头的大小。
            d_model (int): 输入张量的特征维度大小。
            context_length (int): 上下文长度，即时间步数。
            dropout (float): Dropout 层的丢弃率。
        """
        super().__init__()
        # 设置模型的参数
        self.num_heads = num_heads
        self.head_size = head_size
        self.d_model = d_model
        self.context_length = context_length
        self.dropout = dropout
        # 创建多个注意力头
        self.heads = nn.ModuleList([Attention(head_size=self.head_size)
                                    for _ in range(self.num_heads)])
        # 线性投影层，用于将多头注意力的输出映射回原始的特征维度
        self.projection_layer = nn.Linear(in_features=self.num_heads * self.head_size, out_features=self.d_model)
        # Dropout 层，用于防止过拟合
        self.dropout_layer = nn.Dropout(self.dropout)

    def forward(self, x):
        """
        参数:
            x (torch.Tensor): 输入张量，形状为(batch_size, seq_len, d_model)。
        返回:
            torch.Tensor: 多头注意力机制的输出张量，形状与输入张量相同。
        """
        # 对每个注意力头执行前向传播，并将它们的输出拼接在一起
        out = torch.cat([h(x) for h in self.heads], dim=-1)
        # 通过线性投影层将多头注意力的输出映射回原始的特征维度
        out = self.projection_layer(out)
        # 应用 Dropout 层，防止过拟合
        out = self.dropout_layer(out)
        return out


class TransformerBlock(nn.Module):

    def __init__(self):
        super().__init__()
        # 设置模型的参数
        self.d_model = d_model
        self.context_length = context_length
        self.head_size = d_model // num_heads  # 注意力头的大小
        self.num_heads = num_heads
        self.dropout = dropout
        # 多头注意力层
        self.multi_head_attention_layer = MultiHeadAttention(self.head_size)
        # 前馈神经网络层
        self.feed_forward_layer = FeedForward()
        # Layer normalization 层
        self.layer_norm_1 = nn.LayerNorm(normalized_shape=self.d_model)
        self.layer_norm_2 = nn.LayerNorm(normalized_shape=self.d_model)

    def forward(self, x):
        """
        定义模型的前向传播过程。
        参数:
            x (torch.Tensor): 输入张量，形状为(batch_size, seq_len, d_model)。
        返回:
            torch.Tensor: Transformer 块的输出张量，形状与输入张量相同。
        """
        # 注意：操作的顺序与原始的 Transformer 论文不同
        # 这里的顺序是：LayerNorm -> Multi-head attention -> LayerNorm -> Feed forward
        # 使用 Layer normalization 对输入张量进行归一化
        x_normalized_1 = self.layer_norm_1(x)
        # 执行多头注意力机制，并将输出与输入张量相加（残差连接）
        x = x + self.multi_head_attention_layer(x_normalized_1)
        # 使用 Layer normalization 对得到的结果进行归一化
        x_normalized_2 = self.layer_norm_2(x)
        # 执行前馈神经网络，并将输出与之前的结果相加（残差连接）
        x = x + self.feed_forward_layer(x_normalized_2)
        return x

class TransformerLanguageModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.d_model = d_model
        self.context_length = context_length
        self.num_heads = num_heads
        self.num_blocks = num_blocks
        self.dropout = dropout
        self.max_token_value = max_token_value
        # 设置 token 嵌入查找表
        self.token_embedding_lookup_table = nn.Embedding(num_embeddings=self.max_token_value + 1, embedding_dim=self.d_model)

        # 运行所有的 Transformer 块
        # 与原始论文不同，这里在所有块之后添加了一个最终的层规范化
        self.transformer_blocks = nn.Sequential(*(
                [TransformerBlock(num_heads=self.num_heads) for _ in range(self.num_blocks)] +
                [nn.LayerNorm(self.d_model)]
        ))
        self.language_model_out_linear_layer = nn.Linear(in_features=self.d_model, out_features=self.max_token_value)

    def forward(self, idx, targets=None):
        B, T = idx.shape
        """
        # 设置位置嵌入查找表
        # 遵循原始 Transformer 论文相同的方法（正弦和余弦函数）
        """
        position_encoding_lookup_table = torch.zeros(self.context_length, self.d_model)
        position = torch.arange(0, self.context_length, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, self.d_model, 2).float() * (-math.log(10000.0) / self.d_model))
        position_encoding_lookup_table[:, 0::2] = torch.sin(position * div_term)
        position_encoding_lookup_table[:, 1::2] = torch.cos(position * div_term)
        # 将 position_encoding_lookup_table 从 (context_length, d_model) 更改为 (T, d_model)
        position_embedding = position_encoding_lookup_table[:T, :].to(device)
        x = self.token_embedding_lookup_table(idx) + position_embedding
        x = self.transformer_blocks(x)
        # “logits” 是我们的模型在应用 softmax 之前的输出值
        logits = self.language_model_out_linear_layer(x)

        if targets is not None:
            B, T, C = logits.shape
            logits_reshaped = logits.view(B * T, C)
            targets_reshaped = targets.view(B * T)
            loss = F.cross_entropy(input=logits_reshaped, target=targets_reshaped)
        else:
            loss = None
        return logits, loss

    def generate(self, idx, max_new_tokens):
        # idx 是当前上下文中索引的 (B,T) 数组
        for _ in range(max_new_tokens):
            # 将 idx 裁剪到我们位置嵌入表的最大尺寸
            idx_crop = idx[:, -self.context_length:]
            # 获取预测值
            logits, loss = self(idx_crop)
            # 从 logits 中获取最后一个时间步，其中 logits 的维度为 (B,T,C)
            logits_last_timestep = logits[:, -1, :]
            # 应用 softmax 获取概率
            probs = F.softmax(input=logits_last_timestep, dim=-1)
            # 从概率分布中采样
            idx_next = torch.multinomial(input=probs, num_samples=1)
            # 将采样的索引 idx_next 追加到 idx 中
            idx = torch.cat((idx, idx_next), dim=1)
        return idx

本文提供了transformer代码附带详细注释，要注意本文的transformer并非传统的encoder-decoder结构的，而是主流的gpt结构（decoder-only），不了解decoder-only的同学，可以参考我的另一篇文章，链接放在最后。我过几天会出一个介绍gpt模型结构的的文章，欢迎大家前来讨论。http://t.csdnimg.cn/IGCUL