Transformer代码详解: attention-is-all-you-need-pytorch

class ScaledDotProductAttention(nn.Module): ''' Scaled Dot-Product Attention ''' def __init__(self, temperature, attn_dropout=0.1): super().__init__() # 其实就是论文中的根号d_k self.temperature = temperature self.dropout = nn.Dropout(attn_dropout) def forward(self, q, k, v, mask=None): # sz_b: batch_size 批量大小 # len_q,len_k,len_v: 序列长度 在这里他们都相等 # n_head: 多头注意力 默认为8 # d_k,d_v: k v 的dim(维度) 默认都是64 # 此时q的shape为(sz_b, n_head, len_q, d_k) (sz_b, 8, len_q, 64) # 此时k的shape为(sz_b, n_head, len_k, d_k) (sz_b, 8, len_k, 64) # 此时v的shape为(sz_b, n_head, len_k, d_v) (sz_b, 8, len_k, 64) # q先除以self.temperature(论文中的根号d_k) k交换最后两个维度(这样才可以进行矩阵相乘) 最后两个张量进行矩阵相乘 # attn的shape为(sz_b, n_head, len_q, len_k) attn = torch.matmul(q / self.temperature, k.transpose(2, 3)) if mask is not None: # 用-1e9代替0 -1e9是一个很大的负数 经过softmax之后接近与0 # 其一：去除掉各种padding在训练过程中的影响 # 其二，将输入进行遮盖，避免decoder看到后面要预测的东西。（只用在decoder中） attn = attn.masked_fill(mask == 0, -1e9) # 先在attn的最后一个维度做softmax 再dropout 得到注意力分数 attn = self.dropout(F.softmax(attn, dim=-1)) # 最后attn与v进行矩阵相乘 # output的shape为(sz_b, 8, len_q, 64) output = torch.matmul(attn, v) # 返回 output和注意力分数 return output, attn

# q k v 先经过不同的线性层 再用ScaledDotProductAttention 最后再经过一个线性层
class MultiHeadAttention(nn.Module): ''' Multi-Head Attention module ''' def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1): # 这里的n_head, d_model, d_k, d_v分别默认为8, 512, 64, 64 super().__init__() self.n_head = n_head self.d_k = d_k self.d_v = d_v self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False) self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False) self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False) self.fc = nn.Linear(n_head * d_v, d_model, bias=False) self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5) self.dropout = nn.Dropout(dropout) self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) def forward(self, q, k, v, mask=None): d_k, d_v, n_head = self.d_k, self.d_v, self.n_head # len_q, len_k, len_v 为输入的序列长度 sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1) # 用作残差连接 residual = q # Pass through the pre-attention projection: b x lq x (n*dv) # Separate different heads: b x lq x n x dv # q k v 分别经过一个线性层再改变维度 # 由(sz_b, len_q, n_head*d_k) => (sz_b, len_q, n_head, d_k) (sz_b, len_q, 8*64) => (sz_b, len_q, 8, 64) q = self.w_qs(q).view(sz_b, len_q, n_head, d_k) k = self.w_ks(k).view(sz_b, len_k, n_head, d_k) v = self.w_vs(v).view(sz_b, len_v, n_head, d_v) # Transpose for attention dot product: b x n x lq x dv # 交换维度做attention # 由(sz_b, len_q, n_head, d_k) => (sz_b, n_head, len_q, d_k) (sz_b, len_q, 8, 64) => (sz_b, 8, len_q, 64) q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2) if mask is not None: # 为head增加一个维度 mask = mask.unsqueeze(1) # For head axis broadcasting. # 做attention q, attn = self.attention(q, k, v, mask=mask) # Transpose to move the head dimension back: b x lq x n x dv # Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv) # (sz_b, 8, len_k, 64) => (sz_b, len_k, 8, 64) => (sz_b, len_k, 512) q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1) # 经过fc和dropout q = self.dropout(self.fc(q)) # 残差连接 论文中的Add & Norm中的Add q += residual # 论文中的Add & Norm中的Norm q = self.layer_norm(q) # q的shape为(sz_b, len_q, 512) # attn的shape为(sz_b, n_head, len_q, len_k) return q, attn # 其实就是一个MLP而已 class PositionwiseFeedForward(nn.Module): ''' A two-feed-forward-layer module ''' def __init__(self, d_in, d_hid, dropout=0.1): # d_in默认为512 d_hid默认为2048 super().__init__() self.w_1 = nn.Linear(d_in, d_hid) # position-wise self.w_2 = nn.Linear(d_hid, d_in) # position-wise self.layer_norm = nn.LayerNorm(d_in, eps=1e-6) self.dropout = nn.Dropout(dropout) def forward(self, x): residual = x x = self.w_2(F.relu(self.w_1(x))) x = self.dropout(x) # 下面两句对应论文中的Add & Norm中 x += residual x = self.layer_norm(x) return x

class EncoderLayer(nn.Module): ''' Compose with two layers ''' def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1): # d_model, d_inner, n_head, d_k, d_v分别默认为512, 2048, 8, 64, 64 super(EncoderLayer, self).__init__() self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout) self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout) def forward(self, enc_input, slf_attn_mask=None): # q k v都是enc_input enc_output, enc_slf_attn = self.slf_attn( enc_input, enc_input, enc_input, mask=slf_attn_mask) enc_output = self.pos_ffn(enc_output) # enc_output的shape为(sz_b, len_q, 512) # enc_slf_attn的shape为(sz_b, n_head, len_q, len_k) return enc_output, enc_slf_attn class DecoderLayer(nn.Module): ''' Compose with three layers ''' def __init__(self, d_model, d_inner, n_head, d_k, d_v, dropout=0.1): super(DecoderLayer, self).__init__() # 这里的第一个MultiHeadAttention是带Masked self.slf_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout) self.enc_attn = MultiHeadAttention(n_head, d_model, d_k, d_v, dropout=dropout) self.pos_ffn = PositionwiseFeedForward(d_model, d_inner, dropout=dropout) def forward( self, dec_input, enc_output, slf_attn_mask=None, dec_enc_attn_mask=None): # q k v都是dec_input dec_output, dec_slf_attn = self.slf_attn( dec_input, dec_input, dec_input, mask=slf_attn_mask) # q是dec_output k和v是enc_output dec_output, dec_enc_attn = self.enc_attn( dec_output, enc_output, enc_output, mask=dec_enc_attn_mask) dec_output = self.pos_ffn(dec_output) # dec_output的shape为(sz_b, len_q, 512) # dec_slf_attn的shape为(sz_b, n_head, len_q, len_k) # dec_enc_attn的shape为(sz_b, n_head, len_q, len_k) return dec_output, dec_slf_attn, dec_enc_attn

# get_pad_mask()得到的结果和get_subsequent_mask()得到的结果进行与运算（&）
    #               [[[1 0 0]      [[[1 0 0]
    # [[[1 1 0]]] &   [1 1 0]   =    [1 1 0]
    #                 [1 1 1]]]      [1 1 0]]]
# 获取mask并增加一个维度 # pad_idx一般为0 # (batch_size, seq_len) => (batch_size, 1, seq_len) (1, 3) => (1, 1, 3) # 如输入seq为[[1,2,0]] 输出为 [[[1 1 0]]] def get_pad_mask(seq, pad_idx): return (seq != pad_idx).unsqueeze(-2) def get_subsequent_mask(seq): ''' For masking out the subsequent info. ''' sz_b, len_s = seq.size() # 例如输入的seq的shape为(1, 3) # torch.triu(torch.ones((1, len_s, len_s), device=seq.device), diagonal=1))的结果是: # [[[0 1 1] # [0 0 1] # [0 0 0]]] # [[[1 0 0] # subsequent_mask = [1 1 0] # [1 1 1]]] # get_pad_mask()得到的结果和get_subsequent_mask()得到的结果进行与运算（&） # [[[1 0 0] [[[1 0 0] # [[[1 1 0]]] & [1 1 0] = [1 1 0] # [1 1 1]]] [1 1 0]]] subsequent_mask = (1 - torch.triu( torch.ones((1, len_s, len_s), device=seq.device), diagonal=1)).bool() return subsequent_mask

class PositionalEncoding(nn.Module): def __init__(self, d_hid, n_position=200): super(PositionalEncoding, self).__init__() # Not a parameter self.register_buffer('pos_table', self._get_sinusoid_encoding_table(n_position, d_hid)) def _get_sinusoid_encoding_table(self, n_position, d_hid): # n_position默认为200 d_hid默认为d512 ''' Sinusoid position encoding table ''' # TODO: make it with torch instead of numpy # 利用论文中的公式获取某个位置的向量 def get_position_angle_vec(position): # 长度为512 (hid_j // 2)就是论文中的i return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)] # shape为(200, 512) sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)]) # 偶数位置使用sin编码 sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i # 奇数位置使用cos编码 sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1 # shape为(1, n_position, d_hid) return torch.FloatTensor(sinusoid_table).unsqueeze(0) def forward(self, x): # n_position默认为200 seq_len不会超过200 # 这里x加入位置编码 return x + self.pos_table[:, :x.size(1)].clone().detach()

class Encoder(nn.Module): ''' A encoder model with self attention mechanism. ''' def __init__( self, n_src_vocab, d_word_vec, n_layers, n_head, d_k, d_v, d_model, d_inner, pad_idx, dropout=0.1, n_position=200, scale_emb=False): # n_src_vocab: 源语言词汇表的大小 # d_word_vec: 词嵌入的维度 super().__init__() # padding_idx如果指定 则padding_idx处的条目不会影响梯度 因此padding_idx 处的嵌入向量在训练期间不会更新 即它仍然是一个固定的"pad" self.src_word_emb = nn.Embedding(n_src_vocab, d_word_vec, padding_idx=pad_idx) self.position_enc = PositionalEncoding(d_word_vec, n_position=n_position) self.dropout = nn.Dropout(p=dropout) # Encoder包含了n_layers个EncoderLayer n_layers默认为6 self.layer_stack = nn.ModuleList([ EncoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout) for _ in range(n_layers)]) self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) self.scale_emb = scale_emb self.d_model = d_model def forward(self, src_seq, src_mask, return_attns=False): # src_seq: 输入的序列 # src_mask: get_pad_mask()得到的结果 enc_slf_attn_list = [] # -- Forward # 词嵌入 enc_output = self.src_word_emb(src_seq) if self.scale_emb: enc_output *= self.d_model ** 0.5 # 加上位置编码 enc_output = self.dropout(self.position_enc(enc_output)) enc_output = self.layer_norm(enc_output) # n_layers个EncoderLayer串联在一起 for enc_layer in self.layer_stack: enc_output, enc_slf_attn = enc_layer(enc_output, slf_attn_mask=src_mask) enc_slf_attn_list += [enc_slf_attn] if return_attns else [] if return_attns: return enc_output, enc_slf_attn_list return enc_output,

class Decoder(nn.Module): ''' A decoder model with self attention mechanism. ''' def __init__( self, n_trg_vocab, d_word_vec, n_layers, n_head, d_k, d_v, d_model, d_inner, pad_idx, n_position=200, dropout=0.1, scale_emb=False): # n_trg_vocab: 翻译后语言词汇表的大小 # d_word_vec: 词嵌入的维度 super().__init__() self.trg_word_emb = nn.Embedding(n_trg_vocab, d_word_vec, padding_idx=pad_idx) self.position_enc = PositionalEncoding(d_word_vec, n_position=n_position) self.dropout = nn.Dropout(p=dropout) # Decoder包含了n_layers个DecoderLayer n_layers默认为6 self.layer_stack = nn.ModuleList([ DecoderLayer(d_model, d_inner, n_head, d_k, d_v, dropout=dropout) for _ in range(n_layers)]) self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) self.scale_emb = scale_emb self.d_model = d_model def forward(self, trg_seq, trg_mask, enc_output, src_mask, return_attns=False): # trg_seq：翻译后语言序列 # trg_mask: get_pad_mask()得到的结果和get_subsequent_mask()得到的结果进行与运算（&） # enc_output: Encoder的输出 # src_mask: get_pad_mask()得到的结果 dec_slf_attn_list, dec_enc_attn_list = [], [] # -- Forward # 词嵌入 dec_output = self.trg_word_emb(trg_seq) if self.scale_emb: dec_output *= self.d_model ** 0.5 # 加上位置编码 dec_output = self.dropout(self.position_enc(dec_output)) dec_output = self.layer_norm(dec_output) # n_layers个DecoderLayer串联在一起 for dec_layer in self.layer_stack: dec_output, dec_slf_attn, dec_enc_attn = dec_layer( dec_output, enc_output, slf_attn_mask=trg_mask, dec_enc_attn_mask=src_mask) dec_slf_attn_list += [dec_slf_attn] if return_attns else [] dec_enc_attn_list += [dec_enc_attn] if return_attns else [] if return_attns: return dec_output, dec_slf_attn_list, dec_enc_attn_list return dec_output,

class Transformer(nn.Module): ''' A sequence to sequence model with attention mechanism. ''' def __init__( self, n_src_vocab, n_trg_vocab, src_pad_idx, trg_pad_idx, d_word_vec=512, d_model=512, d_inner=2048, n_layers=6, n_head=8, d_k=64, d_v=64, dropout=0.1, n_position=200, trg_emb_prj_weight_sharing=True, emb_src_trg_weight_sharing=True, scale_emb_or_prj='prj'): super().__init__() self.src_pad_idx, self.trg_pad_idx = src_pad_idx, trg_pad_idx # In section 3.4 of paper "Attention Is All You Need", there is such detail: # "In our model, we share the same weight matrix between the two # embedding layers and the pre-softmax linear transformation... # In the embedding layers, we multiply those weights by \sqrt{d_model}". # # Options here: # 'emb': multiply \sqrt{d_model} to embedding output # 'prj': multiply (\sqrt{d_model} ^ -1) to linear projection output # 'none': no multiplication assert scale_emb_or_prj in ['emb', 'prj', 'none'] scale_emb = (scale_emb_or_prj == 'emb') if trg_emb_prj_weight_sharing else False self.scale_prj = (scale_emb_or_prj == 'prj') if trg_emb_prj_weight_sharing else False self.d_model = d_model self.encoder = Encoder( n_src_vocab=n_src_vocab, n_position=n_position, d_word_vec=d_word_vec, d_model=d_model, d_inner=d_inner, n_layers=n_layers, n_head=n_head, d_k=d_k, d_v=d_v, pad_idx=src_pad_idx, dropout=dropout, scale_emb=scale_emb) self.decoder = Decoder( n_trg_vocab=n_trg_vocab, n_position=n_position, d_word_vec=d_word_vec, d_model=d_model, d_inner=d_inner, n_layers=n_layers, n_head=n_head, d_k=d_k, d_v=d_v, pad_idx=trg_pad_idx, dropout=dropout, scale_emb=scale_emb) self.trg_word_prj = nn.Linear(d_model, n_trg_vocab, bias=False) for p in self.parameters(): if p.dim() > 1: nn.init.xavier_uniform_(p) assert d_model == d_word_vec, \ 'To facilitate the residual connections, \  the dimensions of all module outputs shall be the same.' # Decoder中Embedding层和FC层权重共享 # Embedding层参数维度是：(v,d)，FC层参数维度是：(d,v)，可以直接共享嘛，还是要转置？其中v是词表大小，d是embedding维度。 # 查看 pytorch 源码发现真的可以直接共享： if trg_emb_prj_weight_sharing: # Share the weight between target word embedding & last dense layer self.trg_word_prj.weight = self.decoder.trg_word_emb.weight # Encoder和Decoder间的Embedding层权重共享 if emb_src_trg_weight_sharing: self.encoder.src_word_emb.weight = self.decoder.trg_word_emb.weight def forward(self, src_seq, trg_seq): src_mask = get_pad_mask(src_seq, self.src_pad_idx) trg_mask = get_pad_mask(trg_seq, self.trg_pad_idx) & get_subsequent_mask(trg_seq) enc_output, *_ = self.encoder(src_seq, src_mask) dec_output, *_ = self.decoder(trg_seq, trg_mask, enc_output, src_mask) seq_logit = self.trg_word_prj(dec_output) if self.scale_prj: seq_logit *= self.d_model ** -0.5 return seq_logit.view(-1, seq_logit.size(2))