Transformer的pytorch实现

一、用于词向量形式实现

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import math, copy, time
from torch.autograd import Variable


class EncoderDecoder(nn.Module):
    """
    A standard Encoder-Decoder architecture. Base for this and many
    other models.
    """

    def __init__(self, encoder, decoder, src_embed, tgt_embed, generator):
        super(EncoderDecoder, self).__init__()
        self.encoder = encoder
        self.decoder = decoder
        self.src_embed = src_embed  # 映射 [B,L] -> [B,L,d_model]
        self.tgt_embed = tgt_embed  # 映射 [B,L_out] -> [B,L_out,d_model_out]
        self.generator = generator

    def forward(self, src, tgt, src_mask, tgt_mask):
        """
        Take in and process masked src and target sequences.
        """
        return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask)

    def encode(self, src, src_mask):
        """
        调用Encoder类 其中x变化 但mask 不变
        src_embed [batch,len,d_model], src_mask [batch,1,len]->[batch,1,1,len] src中不为0地方True
        """
        return self.encoder(self.src_embed(src), src_mask)

    def decode(self, memory, src_mask, tgt, tgt_mask):
        """
        tgt_embed [batch,len-1,d_model]
        memory [batch,len,d_model]
        src_mask [batch,1,1,len]
        tgt_mask [batch,len-1,len-1] 其中[batch,len-1]是tgt去除最后一列  每个句子复制次数遍 -> [batch,1,len-1,len-1]
        掩码中右上角是False 左下角和主对角线 不为0的词为True
        """
        return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)


class Generator(nn.Module):
    """
    Define standard linear + softmax generation step.
    [B,L,d_model] -> [B,L,vocab]
    """

    def __init__(self, d_model, vocab):
        super(Generator, self).__init__()
        self.proj = nn.Linear(d_model, vocab)

    def forward(self, x):
        return F.log_softmax(self.proj(x), dim=-1)  # softmax结果取对数 映射到负数区间 防止溢出


def clones(module, N):
    """
    Produce N identical layers.
    返回列表 把输入module 复制N次
    """
    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])


class Encoder(nn.Module):
    """Core encoder is a stack of N layers"""

    def __init__(self, layer, N):
        super(Encoder, self).__init__()
        self.layers = clones(layer, N)
        self.norm = LayerNorm(layer.size)

    def forward(self, x, mask):
        """Pass the input (and mask) through each layer in turn."""
        for layer in self.layers:
            x = layer(x, mask)
        return self.norm(x)


class LayerNorm(nn.Module):
    """
    Construct a layernorm module (See citation for details).
    初始化 features 是.shape形式 需要x.shape初始化
    [B,L,d_model] -> [B,L,d_model] 按最后一个维度减均值 求方差
    """

    def __init__(self, features, eps=1e-6):
        super(LayerNorm, self).__init__()
        self.a_2 = nn.Parameter(torch.ones(features))
        self.b_2 = nn.Parameter(torch.zeros(features))
        self.eps = eps

    def forward(self, x):
        mean = x.mean(-1, keepdim=True)
        std = x.std(-1, keepdim=True)
        return self.a_2 * (x - mean) / (std + self.eps) + self.b_2


class SublayerConnection(nn.Module):
    """
    A residual connection followed by a layer norm.
    Note for code simplicity the norm is first as opposed to last.
    初始化 x.shape dropout参数
    输入 x和某层 输出 维度不变 相当于 norm sublayer dropout residual
    """

    def __init__(self, size, dropout):
        super(SublayerConnection, self).__init__()
        self.norm = LayerNorm(size)
        self.dropout = nn.Dropout(dropout)  # 一些按比例增大 一些置0

    def forward(self, x, sublayer):
        """Apply residual connection to any sublayer with the same size."""
        return x + self.dropout(sublayer(self.norm(x)))


class EncoderLayer(nn.Module):
    """
    Encoder is made up of self-attn and feed forward (defined below)
    过自我注意 再过feed_forward
    """

    def __init__(self, size, self_attn, feed_forward, dropout):
        super(EncoderLayer, self).__init__()
        self.self_attn = self_attn
        self.feed_forward = feed_forward
        self.sublayer = clones(SublayerConnection(size, dropout), 2)
        self.size = size

    def forward(self, x, mask):
        """
        Follow Figure 1 (left) for connections.
        lambda 为匿名函数
        """
        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
        return self.sublayer[1](x, self.feed_forward)


class Decoder(nn.Module):
    """Generic N layer decoder with masking."""

    def __init__(self, layer, N):
        super(Decoder, self).__init__()
        self.layers = clones(layer, N)
        self.norm = LayerNorm(layer.size)

    def forward(self, x, memory, src_mask, tgt_mask):
        for layer in self.layers:
            x = layer(x, memory, src_mask, tgt_mask)
        return self.norm(x)


class DecoderLayer(nn.Module):
    """
    Decoder is made of self-attn, src-attn, and feed forward (defined below)
    每一个 DecoderLayer的memory都是一样的 都是多次EncoderLayer后输出的结果
    """

    def __init__(self, size, self_attn, src_attn, feed_forward, dropout):
        super(DecoderLayer, self).__init__()
        self.size = size
        self.self_attn = self_attn
        self.src_attn = src_attn
        self.feed_forward = feed_forward
        self.sublayer = clones(SublayerConnection(size, dropout), 3)

    def forward(self, x, memory, src_mask, tgt_mask):
        """Follow Figure 1 (right) for connections."""
        m = memory
        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))
        x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))
        return self.sublayer[2](x, self.feed_forward)


def subsequent_mask(size):
    """
    Mask out subsequent positions.
    triu 返回上三角矩阵
    [1,size,size] 右上角F 左下角和对角线True
    """
    attn_shape = (1, size, size)
    subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')
    return torch.from_numpy(subsequent_mask) == 0


def attention(query, key, value, mask=None, dropout=None):
    """
    Compute 'Scaled Dot Product Attention'
    qkv[B,h,L,d_model/h] src_mask[B,1,1,L] tgt_mask[]
    score(qkT)[B,h,L,L]  src_mask[B,1,1,L] -> [B,h,L,L] B管B L只掩码所有最内层
    score(qkT)[B,h,L-1,L-1] tgt_mask[B,1,L-1,L-1]
    [[q1k1,q1k2,q1k3]
     [q2k1,q2k2,q2k3]
     [q3k1,q3k2,q3k3]] tgt右上角三个元素置0 src不去注意0 q不注意0处的k

     互注意 q[B,h,L-1,d_model/h] kv[B,h,L,d_model/h]

     [[q1k1 q1k2->0 q1k3->0]
      [q2k1 q2k2    q2k3->0]]
    """
    d_k = query.size(-1)
    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
    # score [B,L,L] mask [B,L,L]或[L,L]广播 mask中为0的地方 score中相应地方被填充为无穷小
    if mask is not None:
        scores = scores.masked_fill(mask == 0, -1e9)
    p_attn = F.softmax(scores, dim=-1)
    if dropout is not None:
        p_attn = dropout(p_attn)
    return torch.matmul(p_attn, value), p_attn


class MultiHeadedAttention(nn.Module):
    def __init__(self, h, d_model, dropout=0.1):
        """
        Take in model size and number of heads.
        [B,L,d_model] -> [B,L,d_model]
        """
        super(MultiHeadedAttention, self).__init__()
        assert d_model % h == 0  # 不为0抛出异常
        # We assume d_v always equals d_k
        self.d_k = d_model // h
        self.h = h
        self.linears = clones(nn.Linear(d_model, d_model), 4)
        self.attn = None
        self.dropout = nn.Dropout(p=dropout)

    def forward(self, query, key, value, mask=None):
        """Implements Figure 2"""
        if mask is not None:
            # Same mask applied to all h heads.
            mask = mask.unsqueeze(1)  # mask[B,1,1,L]
        nbatches = query.size(0)
        # 1) Do all the linear projections in batch from d_model => h x d_k
        # [B,L,d_model]线性化[B,L,d_model]->[B,L,h,d_model/h]->[B,h,L,d_model/h]
        query, key, value = [l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)
             for l, x in zip(self.linears, (query, key, value))]
        # 2) Apply attention on all the projected vectors in batch.
        x, self.attn = attention(query, key, value, mask=mask, dropout=self.dropout)
        # x [B,h,L,d_model/h]   self.attn [B,h,L,L]
        # 3) "Concat" using a view and apply a final linear.
        x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k)
        # x [B,h,L,d_model/h] -> [B,L,h,d_model/h] -> [B,L,d_model]
        return self.linears[-1](x)


class PositionwiseFeedForward(nn.Module):
    """
    Implements FFN equation.
    两个MLP [B,L,d_model]->[B,L,d_model]
    """

    def __init__(self, d_model, d_ff, dropout=0.1):
        super(PositionwiseFeedForward, self).__init__()
        self.w_1 = nn.Linear(d_model, d_ff)
        self.w_2 = nn.Linear(d_ff, d_model)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        return self.w_2(self.dropout(F.relu(self.w_1(x))))


class Embeddings(nn.Module):
    """
    [B,L] -> [B,L,d_model]
    [B,L]中元素都为自然数 vocab至少比最大的自然数大1
    d_model 想嵌入的维度 vocab大于等于最大索引词索引数+1
    """

    def __init__(self, d_model, vocab):
        super(Embeddings, self).__init__()
        self.lut = nn.Embedding(vocab, d_model)
        self.d_model = d_model

    def forward(self, x):
        return self.lut(x) * math.sqrt(self.d_model)


class PositionalEncoding(nn.Module):
    """
    Implement the PE function.加入句子位置编码
    [B,L,d_model]->[B,L,d_model]
    """

    def __init__(self, d_model, dropout, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)
        # Compute the positional encodings once in log space.
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len).unsqueeze(1)  # [max_len,1] [[0],[1],...,[max_len-1]]
        div_term = torch.exp(torch.arange(0, d_model, 2) * (-(math.log(10000.0) / d_model)))  # [0,2,4,...,不包括d_model]
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.register_buffer('pe', pe)  # 网络更新时它不更新

    def forward(self, x):
        # shape [1,x.size(1),d_model] 广播机制 所有的batch即每个句子 都加上了同样的编码矩阵
        x = x + Variable(self.pe[:, :x.size(1)], requires_grad=False)
        return self.dropout(x)


def make_model(src_vocab, tgt_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1):
    """Helper: Construct a model from hyperparameters."""
    c = copy.deepcopy
    attn = MultiHeadedAttention(h, d_model)
    ff = PositionwiseFeedForward(d_model, d_ff, dropout)
    position = PositionalEncoding(d_model, dropout)
    model = EncoderDecoder(
        Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),  # 两个编码层串联
        Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),  # 两个解码层串联 两个空闲memory不接
        nn.Sequential(Embeddings(d_model, src_vocab), c(position)),  # 依次传参容器 src_vocab 输入的词有几种
        nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),  # 依次传参容器 tgt_vocab 输出的词有几种
        Generator(d_model, tgt_vocab))  # 一个MLP
    # This was important from their code.
    # Initialize parameters with Glorot / fan_avg. 初始化参数
    for p in model.parameters():
        if p.dim() > 1:
            nn.init.xavier_uniform_(p)
    return model


class Batch:
    """Object for holding a batch of data with mask during training."""

    def __init__(self, src, trg=None, pad=0):
        self.src = src
        self.src_mask = (src != pad).unsqueeze(-2)  # 加一个维度 src[batch,len] src_mask[batch,1,len] src中不为0地方True
        if trg is not None:
            self.trg = trg[:, :-1]  # trg除去最后一列 [batch,len-1]
            self.trg_y = trg[:, 1:]  # trg除去第一列  [batch,len-1]
            self.trg_mask = self.make_std_mask(self.trg, pad)
            self.ntokens = (self.trg_y != pad).data.sum()  # 返回trg除去首列 非0元素数目

    # 静态方法 不实例化类也可以通过类直接调用方法
    @staticmethod
    def make_std_mask(tgt, pad):
        """Create a mask to hide padding and future words."""
        tgt_mask = (tgt != pad).unsqueeze(-2)  # [batch,1,len-1] trg除去最后一列 不为0True
        tgt_mask = tgt_mask & Variable(subsequent_mask(tgt.size(-1)).type_as(tgt_mask.data)) # [1,len-1,len-1]右上角F 左下角和对角线True
        # [batch,len-1,len-1]
        # 先把句子广播[batch,len-1]->[batch,len-1,len-1]每一个句子原来用一个向量表示 现在把向量再重复
        # [B,i,:] 看B个句子 第i表示看前i个单词
        return tgt_mask


def run_epoch(data_iter, model, loss_compute):
    """Standard Training and Logging Function"""
    start = time.time()
    total_tokens = 0
    total_loss = 0
    tokens = 0
    for i, batch in enumerate(data_iter):
        out = model.forward(batch.src, batch.trg, batch.src_mask, batch.trg_mask)
        # batch.src         [batch,len]
        # batch.trg         [batch,len]
        # batch.src_mask    [batch,1,len] src中不为0地方True
        # batch.trg_mask    [batch,len-1,len-1] trg除去最后一列 重复len-1次 不为0True 再和右上F相交
        # batch.trg_y       [batch,len-1] trg除去第一列
        # batch.ntokens     返回trg除去首列 非0元素数目
        # out               [batch,len-1,d_model]->[batch,len-1,vocab]
        loss = loss_compute(out, batch.trg_y, batch.ntokens)
        total_loss += loss
        total_tokens += batch.ntokens
        tokens += batch.ntokens
        if i % 50 == 1:
            elapsed = time.time() - start
            print("Epoch Step: %d Loss: %f Tokens per Sec: %f" % (i, loss / batch.ntokens, tokens / elapsed))
            start = time.time()
            tokens = 0
    return total_loss / total_tokens


global max_src_in_batch, max_tgt_in_batch


def batch_size_fn(new, count, sofar):
    """Keep augmenting batch and calculate total number of tokens + padding."""
    global max_src_in_batch, max_tgt_in_batch
    if count == 1:
        max_src_in_batch = 0
        max_tgt_in_batch = 0
    max_src_in_batch = max(max_src_in_batch,  len(new.src))
    max_tgt_in_batch = max(max_tgt_in_batch,  len(new.trg) + 2)
    src_elements = count * max_src_in_batch
    tgt_elements = count * max_tgt_in_batch
    return max(src_elements, tgt_elements)


class NoamOpt:
    """Optim wrapper that implements rate."""
    def __init__(self, model_size, factor, warmup, optimizer):
        self.optimizer = optimizer
        self._step = 0
        self.warmup = warmup
        self.factor = factor
        self.model_size = model_size
        self._rate = 0

    def step(self):
        """Update parameters and rate"""
        self._step += 1
        rate = self.rate()
        for p in self.optimizer.param_groups:
            p['lr'] = rate
        self._rate = rate
        self.optimizer.step()

    def rate(self, step=None):
        """Implement `lrate` above"""
        if step is None:
            step = self._step
        return self.factor * (self.model_size ** (-0.5) * min(step ** (-0.5), step * self.warmup ** (-1.5)))


def get_std_opt(model):
    return NoamOpt(model.src_embed[0].d_model, 2, 4000,
                   torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))


class LabelSmoothing(nn.Module):
    """Implement label smoothing."""

    def __init__(self, size, padding_idx, smoothing=0.0):
        super(LabelSmoothing, self).__init__()
        self.criterion = nn.KLDivLoss(reduction='sum')
        self.padding_idx = padding_idx
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing
        self.size = size
        self.true_dist = None

    def forward(self, x, target):
        # x [B(l-1),V] target [B(l-1)]
        assert x.size(1) == self.size
        true_dist = x.data.clone()  # 复制一个x到true_dist
        true_dist.fill_(self.smoothing / (self.size - 2))  # true_dist内所有数字都换成超参数
        # target.data.unsqueeze(1) [B(l-1),1]
        print(true_dist.shape)
        print(target.data.unsqueeze(1).shape)
        true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)  # 填充或修改元素
        # 把true_dist中一些元素替换为self.confidence 每一行的target索引元素
        true_dist[:, self.padding_idx] = 0  # 所有行0列为0
        print(target.data.shape)
        print(target.data)
        mask = torch.nonzero(target.data == self.padding_idx) # 记录下其中k个为0(self.padding_idx)元素位置(后式成立)
        print(mask)
        print(mask.dim())
        # mask.dim()=2 [[1],[4],[7]]
        if mask.dim() > 0:
            true_dist.index_fill_(0, mask.squeeze(), 0.0)  # padding位置 行所有元素需要被置0 只要是源被pad地方 相应的映射向量置0
        self.true_dist = true_dist
        print(x.shape)
        print(true_dist.shape)
        return self.criterion(x, Variable(true_dist, requires_grad=False))


def data_gen(V, batch, nbatches):
    """Generate random data for a src-tgt copy task."""
    for i in range(nbatches):
        data = torch.from_numpy(np.random.randint(1, V, size=(batch, 10)))
        data[:, 0] = 1
        src = Variable(data, requires_grad=False)  # [batch, len]
        tgt = Variable(data, requires_grad=False)  # [batch, len]
        yield Batch(src, tgt, 0)


class SimpleLossCompute:
    """A simple loss compute and train function."""

    def __init__(self, generator, criterion, opt=None):
        self.generator = generator
        self.criterion = criterion
        self.opt = opt

    def __call__(self, x, y, norm):
        x = self.generator(x)
        # y 一维 x 二维 .contiguous() 常与.view()连用 表示开辟新内存
        # x [B,len-1,V] -> [B(len-1),V] V指映射到多少词
        # y [B,len-1]   -> [B(len-1)]
        loss = self.criterion(x.contiguous().view(-1, x.size(-1)),
                              y.contiguous().view(-1)) / norm
        loss.backward()
        if self.opt is not None:
            self.opt.step()
            self.opt.optimizer.zero_grad()
        return loss.item() * norm


# # Train the simple copy task.
# V = 11  # 一共输入输出都有11种词
# criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0)
# model = make_model(V, V, N=2)
# model_opt = NoamOpt(model.src_embed[0].d_model, 1, 400, torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))
#
# for epoch in range(10):
#     model.train()
#     run_epoch(data_gen(V, 30, 20), model, SimpleLossCompute(model.generator, criterion, model_opt))
#     model.eval()
#     print(run_epoch(data_gen(V, 30, 5), model, SimpleLossCompute(model.generator, criterion, None)))
# # for i, batch in enumerate(data_gen(11, 30, 20)):
# #     print(i)
# #     print(batch)

my_batch = 3
my_len_src = 5
my_len_tgt = 7
my_src_vocab = 15
my_tgt_vocab = 20
my_src = torch.tensor([[2,3,7,4,0],[12,5,7,0,0],[13,4,2,8,5]])
my_tgt = torch.tensor([[1,15,2,3,4,0,0],[1,18,3,1,0,0,0],[4,17,5,2,0,0,0]])
my_src_mask = torch.ones([my_batch,1,my_len_src])
my_tgt_mask = torch.ones([my_batch,my_len_tgt,my_len_tgt])
my_model = make_model(my_src_vocab, my_tgt_vocab)
print(my_model(my_src,my_tgt,my_src_mask,my_tgt_mask).shape)
# [my_batch,my_len_tgt,d_model]

二、通用形式实现

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math, copy
from torch.autograd import Variable

'''
model = make_model(d_input, d_output)
Inputs [batch, input_len, d_input]
Outputs [batch, output_len, d_output]
Output Probabilities [batch, output_len, d_prob]
src_mask = torch.ones([batch,1,input_len]) [batch,1,input_len] (用于处理输入每个batch有效长度len不一致问题)
tgt_mask = subsequent_mask(output_len) [1,output_len,output_len] (右上为False 左下与对角线True)
model(Inputs,Outputs,src_mask,tgt_mask)

src [batch,input_len,d_model]
tgt [batch,output_len,d_model]
'''

class EncoderDecoder(nn.Module):
    """
    A standard Encoder-Decoder architecture. Base for this and many
    other models.
    """

    def __init__(self, encoder, decoder, src_embed, tgt_embed, generator):
        super(EncoderDecoder, self).__init__()
        self.encoder = encoder
        self.decoder = decoder
        self.src_embed = src_embed  # 映射 [B,L] -> [B,L,d_model]
        self.tgt_embed = tgt_embed  # 映射 [B,L_out] -> [B,L_out,d_model_out]
        self.generator = generator

    def forward(self, src, tgt, src_mask, tgt_mask):
        """
        Take in and process masked src and target sequences.
        """
        return self.generator(self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask))

    def encode(self, src, src_mask):
        """
        调用Encoder类 其中x变化 但mask 不变
        src_embed [batch,len,d_model], src_mask [batch,1,len]->[batch,1,1,len] src中不为0地方True
        """
        return self.encoder(self.src_embed(src), src_mask)

    def decode(self, memory, src_mask, tgt, tgt_mask):
        """
        tgt_embed [batch,len-1,d_model]
        memory [batch,len,d_model]
        src_mask [batch,1,1,len]
        tgt_mask [batch,len-1,len-1] 其中[batch,len-1]是tgt去除最后一列  每个句子复制次数遍 -> [batch,1,len-1,len-1]
        掩码中右上角是False 左下角和主对角线 不为0的词为True
        """
        return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)


class Generator(nn.Module):
    """
    Define standard linear + softmax generation step.
    [B,L,d_model] -> [B,L,vocab]
    """

    def __init__(self, d_model, vocab):
        super(Generator, self).__init__()
        self.proj = nn.Linear(d_model, vocab)

    def forward(self, x):
        return F.log_softmax(self.proj(x), dim=-1)  # softmax结果取对数 映射到负数区间 防止溢出


def clones(module, N):
    """
    Produce N identical layers.
    返回列表 把输入module 复制N次
    """
    return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])


class Encoder(nn.Module):
    """Core encoder is a stack of N layers"""

    def __init__(self, layer, N):
        super(Encoder, self).__init__()
        self.layers = clones(layer, N)
        self.norm = LayerNorm(layer.size)

    def forward(self, x, mask):
        """Pass the input (and mask) through each layer in turn."""
        for layer in self.layers:
            x = layer(x, mask)
        return self.norm(x)


class LayerNorm(nn.Module):
    """
    Construct a layernorm module (See citation for details).
    初始化 features 是.shape形式 需要x.shape 最后一个维度大小初始化
    [B,L,d_model] -> [B,L,d_model] 按最后一个维度减均值 求方差
    """

    def __init__(self, features, eps=1e-6):
        super(LayerNorm, self).__init__()
        self.a_2 = nn.Parameter(torch.ones(features))
        self.b_2 = nn.Parameter(torch.zeros(features))
        self.eps = eps

    def forward(self, x):
        mean = x.mean(-1, keepdim=True)
        std = x.std(-1, keepdim=True)
        return self.a_2 * (x - mean) / (std + self.eps) + self.b_2


class SublayerConnection(nn.Module):
    """
    A residual connection followed by a layer norm.
    Note for code simplicity the norm is first as opposed to last.
    初始化 x.shape 最后一个维度 dropout参数
    输入 x和某层 输出 维度不变 相当于 norm sublayer dropout residual
    """

    def __init__(self, size, dropout):
        super(SublayerConnection, self).__init__()
        self.norm = LayerNorm(size)
        self.dropout = nn.Dropout(dropout)  # 一些按比例增大 一些置0

    def forward(self, x, sublayer):
        """Apply residual connection to any sublayer with the same size."""
        return x + self.dropout(sublayer(self.norm(x)))


class EncoderLayer(nn.Module):
    """
    Encoder is made up of self-attn and feed forward (defined below)
    过自我注意 再过feed_forward
    """

    def __init__(self, size, self_attn, feed_forward, dropout):
        super(EncoderLayer, self).__init__()
        self.self_attn = self_attn
        self.feed_forward = feed_forward
        self.sublayer = clones(SublayerConnection(size, dropout), 2)
        self.size = size

    def forward(self, x, mask):
        """
        Follow Figure 1 (left) for connections.
        lambda 为匿名函数
        """
        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
        return self.sublayer[1](x, self.feed_forward)


class Decoder(nn.Module):
    """Generic N layer decoder with masking."""

    def __init__(self, layer, N):
        super(Decoder, self).__init__()
        self.layers = clones(layer, N)
        self.norm = LayerNorm(layer.size)

    def forward(self, x, memory, src_mask, tgt_mask):
        for layer in self.layers:
            x = layer(x, memory, src_mask, tgt_mask)
        return self.norm(x)


class DecoderLayer(nn.Module):
    """
    Decoder is made of self-attn, src-attn, and feed forward (defined below)
    每一个 DecoderLayer的memory都是一样的 都是多次EncoderLayer后输出的结果
    """

    def __init__(self, size, self_attn, src_attn, feed_forward, dropout):
        super(DecoderLayer, self).__init__()
        self.size = size
        self.self_attn = self_attn
        self.src_attn = src_attn
        self.feed_forward = feed_forward
        self.sublayer = clones(SublayerConnection(size, dropout), 3)

    def forward(self, x, memory, src_mask, tgt_mask):
        """Follow Figure 1 (right) for connections."""
        m = memory
        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))
        x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))
        return self.sublayer[2](x, self.feed_forward)


def subsequent_mask(size):
    """
    Mask out subsequent positions.
    triu 返回上三角矩阵
    [1,size,size] 右上角F 左下角和对角线True
    """
    attn_shape = (1, size, size)
    subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')
    return torch.from_numpy(subsequent_mask) == 0


def attention(query, key, value, mask=None, dropout=None):
    """
    Compute 'Scaled Dot Product Attention'
    编码
    qkv[B,h,L,d_model/h] src_mask[B,1,L]->[B,1,1,L]
    score(qkT)[B,h,L,L]  src_mask[B,1,1,L] -> [B,h,L,L] B管B L只掩码所有最内层
    [[q1k1,q1k2->-1e9,q1k3]
     [q2k1,q2k2->-1e9,q2k3]
     [q3k1,q3k2->-1e9,q3k3]]如果掩码不想让第二个单词参与计算则所有人都不注意第二个单词
    解码
    score(qkT)[B,h,L,L] tgt_mask[1,1,L,L]
    [[q1k1,q1k2->-1e9,q1k3->-1e9]
     [q2k1,q2k2      ,q2k3->-1e9]
     [q3k1,q3k2      ,q3k3]] tgt右上角三个元素置-1e9 src不去注意0 q不注意0处的k
    互注意
    q[B,h,output_len,d_model/h] kv[B,h,input_len,d_model/h] src_mask[B,1,input_len]->[B,1,1,input_len](务必符合kv.shape)
    score(qkT)[B,h,output_len,input_len]
    [[q1k1 q1k2->-1e9 q1k3]
     [q2k1 q2k2->-1e9 q2k3]]
    """
    d_k = query.size(-1)
    scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
    # scores 编码[batch,h,input_len,input_len] 解码[batch,h,output_len,output_len] 互注意[batch,h,output_len,input_len]
    # score [B,L,L] mask [B,L,L]或[L,L]广播 mask中为0的地方 score中相应地方被填充为无穷小
    if mask is not None:
        scores = scores.masked_fill(mask == 0, -1e9)
    p_attn = F.softmax(scores, dim=-1)
    if dropout is not None:
        p_attn = dropout(p_attn)
    return torch.matmul(p_attn, value), p_attn


class MultiHeadedAttention(nn.Module):
    def __init__(self, h, d_model, dropout=0.1):
        """
        Take in model size and number of heads.
        [B,L,d_model] -> [B,L,d_model]
        """
        super(MultiHeadedAttention, self).__init__()
        assert d_model % h == 0  # 不为0抛出异常
        # We assume d_v always equals d_k
        self.d_k = d_model // h
        self.h = h
        self.linears = clones(nn.Linear(d_model, d_model), 4)  # 前三个用于qkv 最后一个用于输出
        self.attn = None
        self.dropout = nn.Dropout(p=dropout)

    def forward(self, query, key, value, mask=None):
        """Implements Figure 2"""
        if mask is not None:
            # Same mask applied to all h heads.
            mask = mask.unsqueeze(1)  # mask[B,1,1,L]
        nbatches = query.size(0)
        # 1) Do all the linear projections in batch from d_model => h x d_k
        # [B,L,d_model]线性化[B,L,d_model]->[B,L,h,d_model/h]->[B,h,L,d_model/h]
        query, key, value = [l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)
             for l, x in zip(self.linears, (query, key, value))]
        # 2) Apply attention on all the projected vectors in batch.
        x, self.attn = attention(query, key, value, mask=mask, dropout=self.dropout)
        # x [B,h,L,d_model/h]   self.attn [B,h,L,L]
        # 3) "Concat" using a view and apply a final linear.
        x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k)
        # x [B,h,L,d_model/h] -> [B,L,h,d_model/h] -> [B,L,d_model]
        return self.linears[-1](x)


class PositionwiseFeedForward(nn.Module):
    """
    Implements FFN equation.
    两个MLP [B,L,d_model]->[B,L,d_model]
    """

    def __init__(self, d_model, d_ff, dropout=0.1):
        super(PositionwiseFeedForward, self).__init__()
        self.w_1 = nn.Linear(d_model, d_ff)
        self.w_2 = nn.Linear(d_ff, d_model)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        return self.w_2(self.dropout(F.relu(self.w_1(x))))


class Embeddings(nn.Module):
    """
    [B,L,vocab] -> [B,L,d_model]
    """

    def __init__(self, d_model, vocab):
        super(Embeddings, self).__init__()
        self.lut = nn.Linear(vocab, d_model)
        self.d_model = d_model

    def forward(self, x):
        return self.lut(x) * math.sqrt(self.d_model)


class PositionalEncoding(nn.Module):
    """
    Implement the PE function.加入句子位置编码
    [B,L,d_model]->[B,L,d_model]
    """

    def __init__(self, d_model, dropout, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)
        # Compute the positional encodings once in log space.
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len).unsqueeze(1)  # [max_len,1] [[0],[1],...,[max_len-1]]
        div_term = torch.exp(torch.arange(0, d_model, 2) * (-(math.log(10000.0) / d_model)))  # [0,2,4,...,不包括d_model]
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.register_buffer('pe', pe)  # 网络更新时它不更新

    def forward(self, x):
        # shape [1,x.size(1),d_model] 广播机制 所有的batch即每个句子 都加上了同样的编码矩阵
        x = x + Variable(self.pe[:, :x.size(1)], requires_grad=False)
        return self.dropout(x)


def make_model(src_vocab, tgt_vocab, prob_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1):
    """
    Helper: Construct a model from hyperparameters.
    src_vocab Inputs末维度
    tgt_vocab Outputs末维度
    tgt_vocab Output Probabilities末维度
    """
    c = copy.deepcopy
    attn = MultiHeadedAttention(h, d_model)
    ff = PositionwiseFeedForward(d_model, d_ff, dropout)
    position = PositionalEncoding(d_model, dropout)
    model = EncoderDecoder(
        Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),  # 两个编码层串联
        Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),  # 两个解码层串联 两个空闲memory不接
        nn.Sequential(Embeddings(d_model, src_vocab), c(position)),  # 依次传参容器 src_vocab 输入的词有几种
        nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),  # 依次传参容器 tgt_vocab 输出的词有几种
        Generator(d_model, prob_vocab))  # 一个MLP
    # This was important from their code.
    # Initialize parameters with Glorot / fan_avg. 初始化参数
    for p in model.parameters():
        if p.dim() > 1:
            nn.init.xavier_uniform_(p)
    return model

batch = 3
input_len = 5
d_input = 15
output_len = 7
d_output = 17
d_prob = 19
my_model = make_model(d_input, d_output, d_prob)
Inputs = torch.ones([batch, input_len, d_input])
Outputs = torch.ones([batch, output_len, d_output])
src_mask = torch.ones([batch,1,input_len])
tgt_mask = subsequent_mask(output_len)
print(my_model(Inputs,Outputs,src_mask,tgt_mask).shape)

  • 0
    点赞
  • 2
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值