Seq2Seq(Attention)的PyTorch實現

B站視頻講解php

文本主要介紹一下如何使用PyTorch復現Seq2Seq(with Attention),實現簡單的機器翻譯任務,請先閱讀論文Neural Machine Translation by Jointly Learning to Align and Translate,以後花上15分鐘閱讀個人這兩篇文章Seq2Seq 與注意力機制圖解Attention,最後再來看文本,方能達到醍醐灌頂,事半功倍的效果html

數據預處理

數據預處理的代碼其實就是調用各類API,我不但願讀者被這些不過重要的部分分散了注意力,所以這裏我不貼代碼,僅口述一下帶過便可python

以下圖所示,本文使用的是德語→英語數據集,輸入是德語,而且輸入的每一個句子開頭和結尾都帶有特殊的標識符。輸出是英語,而且輸出的每一個句子開頭和結尾也都帶有特殊標識符git

不論是英語仍是德語,每句話長度都是不固定的,因此我對於每一個batch內的句子,將它們的長度經過加<PAD>變得同樣,也就說,一個batch內的句子,長度都是相同的,不一樣batch內的句子長度不必定相同。下圖維度表示分別是[seq_len, batch_size]github

隨便打印一條數據,看一下數據封裝的形式markdown

在數據預處理的時候,須要將源句子和目標句子分開構建字典,也就是單獨對德語構建一個詞庫,對英語構建一個詞庫網絡

Encoder

Encoder我是用的單層雙向GRUapp

雙向GRU的隱藏狀態輸出由兩個向量拼接而成,例如 h 1 = [ h 1 ; h T ] h_1=[\overrightarrow{h_1};\overleftarrow{h_T}] , h 2 = [ h 2 ; h T 1 ] h_2=[\overrightarrow{h_2};\overleftarrow{h}_{T-1}] ......全部時刻的最後一層隱藏狀態就構成了GRU的outputdom

o u t p u t = { h 1 , h 2 , . . . h T } output=\{h_1,h_2,...h_T\}

假設這是個m層GRU,那麼最後一個時刻全部層中的隱藏狀態就構成了GRU的final hidden stateside

h i d d e n = { h T 1 , h T 2 , . . . , h T m } hidden=\{h^1_T,h^2_T,...,h^m_T\}

其中

h T i = [ h T i ; h 1 i ] h^i_T=[\overrightarrow{h^i_T};\overleftarrow{h^i_1}]

因此

h i d d e n = { [ h T 1 ; h 1 1 ] , [ h T 2 ; h 1 2 ] , . . . , [ h T m ; h 1 m ] } hidden=\{[\overrightarrow{h^1_T};\overleftarrow{h^1_1}],[\overrightarrow{h^2_T};\overleftarrow{h^2_1}],...,[\overrightarrow{h^m_T};\overleftarrow{h^m_1}]\}

根據論文,或者你看了個人圖解Attention這篇文章就會知道,咱們須要的是hidden的最後一層輸出(包括正向和反向),所以咱們能夠經過hidden[-2,:,:]hidden[-1,:,:]取出最後一層的hidden states,將它們拼接起來記做 s 0 s_0

最後一個細節之處在於, s 0 s_0 的維度是[batch_size, en_hid_dim*2],即使是沒有Attention機制,將 s 0 s_0 做爲Decoder的初始隱藏狀態也不對,由於維度不匹配,Decoder的初始隱藏狀態是三維的,而如今咱們的 s 0 s_0 是二維的,所以須要將 s 0 s_0 的維度轉爲三維,而且還要調整各個維度上的值。首先我經過一個全鏈接神經網絡,將 s 0 s_0 的維度變爲[batch_size, dec_hid_dim]

Encoder的細節就這麼多,下面直接上代碼,個人代碼風格是,註釋在上,代碼在下

class Encoder(nn.Module):
    def __init__(self, input_dim, emb_dim, enc_hid_dim, dec_hid_dim, dropout):
        super().__init__()
        self.embedding = nn.Embedding(input_dim, emb_dim)
        self.rnn = nn.GRU(emb_dim, enc_hid_dim, bidirectional = True)
        self.fc = nn.Linear(enc_hid_dim * 2, dec_hid_dim)
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, src): 
        ''' src = [src_len, batch_size] '''
        src = src.transpose(0, 1) # src = [batch_size, src_len]
        embedded = self.dropout(self.embedding(src)).transpose(0, 1) # embedded = [src_len, batch_size, emb_dim]
        
        # enc_output = [src_len, batch_size, hid_dim * num_directions]
        # enc_hidden = [n_layers * num_directions, batch_size, hid_dim]
        enc_output, enc_hidden = self.rnn(embedded) # if h_0 is not give, it will be set 0 acquiescently

        # enc_hidden is stacked [forward_1, backward_1, forward_2, backward_2, ...]
        # enc_output are always from the last layer
        
        # enc_hidden [-2, :, : ] is the last of the forwards RNN 
        # enc_hidden [-1, :, : ] is the last of the backwards RNN
        
        # initial decoder hidden is final hidden state of the forwards and backwards 
        # encoder RNNs fed through a linear layer
        # s = [batch_size, dec_hid_dim]
        s = torch.tanh(self.fc(torch.cat((enc_hidden[-2,:,:], enc_hidden[-1,:,:]), dim = 1)))
        
        return enc_output, s
複製代碼

Attention

attention無非就是三個公式

E t = t a n h ( a t t n ( s t 1 , H ) ) a t ~ = v E t a t = s o f t m a x ( a t ~ ) E_t=tanh(attn(s_{t-1},H))\\ \tilde{a_t}=vE_t\\ {a_t}=softmax(\tilde{a_t})

其中 s t 1 s_{t-1} 指的就是Encoder中的變量s H H 指的就是Encoder中的變量enc_output a t t n ( ) attn() 其實就是一個簡單的全鏈接神經網絡

咱們能夠從最後一個公式反推各個變量的維度是什麼,或者維度有什麼要求

首先 a t a_t 的維度應該是[batch_size, src_len],這是毋庸置疑的,那麼 a t ~ \tilde{a_t} 的維度也應該是[batch_size, src_len],或者 a t ~ \tilde{a_t} 是個三維的,可是某個維度值爲1,能夠經過squeeze()變成兩維的。這裏咱們先假設 a t ~ \tilde{a_t} 的維度是[batch_size, src_len, 1],等會兒我再解釋爲何要這樣假設

繼續往上推,變量 v v 的維度就應該是[?, 1]?表示我暫時不知道它的值應該是多少。 E t E_t 的維度應該是[batch_size, src_len, ?]

如今已知 H H 的維度是[batch_size, src_len, enc_hid_dim*2] s t 1 s_{t-1} 目前的維度是[batch_size, dec_hid_dim],這兩個變量須要作拼接,送入全鏈接神經網絡,所以咱們首先須要將 s t 1 s_{t-1} 的維度變成[batch_size, src_len, dec_hid_dim],拼接以後的維度就變成[batch_size, src_len, enc_hid_dim*2+dec_hid_dim],因而 a t t n ( ) attn() 這個函數的輸入輸出值也就有了

attn = nn.Linear(enc_hid_dim*2+dec_hid_dim, ?)
複製代碼

到此爲止,除了?部分的值不清楚,其它全部維度都推導出來了。如今咱們回過頭思考一下?設置成多少,好像其實並無任何限制,因此咱們能夠設置?爲任何值(在代碼中我設置?dec_hid_dim

Attention細節就這麼多,下面給出代碼

class Attention(nn.Module):
    def __init__(self, enc_hid_dim, dec_hid_dim):
        super().__init__()
        self.attn = nn.Linear((enc_hid_dim * 2) + dec_hid_dim, dec_hid_dim, bias=False)
        self.v = nn.Linear(dec_hid_dim, 1, bias = False)
        
    def forward(self, s, enc_output):
        
        # s = [batch_size, dec_hid_dim]
        # enc_output = [src_len, batch_size, enc_hid_dim * 2]
        
        batch_size = enc_output.shape[1]
        src_len = enc_output.shape[0]
        
        # repeat decoder hidden state src_len times
        # s = [batch_size, src_len, dec_hid_dim]
        # enc_output = [batch_size, src_len, enc_hid_dim * 2]
        s = s.unsqueeze(1).repeat(1, src_len, 1)
        enc_output = enc_output.transpose(0, 1)
        
        # energy = [batch_size, src_len, dec_hid_dim]
        energy = torch.tanh(self.attn(torch.cat((s, enc_output), dim = 2)))
        
        # attention = [batch_size, src_len]
        attention = self.v(energy).squeeze(2)
        
        return F.softmax(attention, dim=1)
複製代碼

Seq2Seq(with Attention)

我調換一下順序,先講Seq2Seq,再講Decoder的部分

傳統Seq2Seq是直接將句子中每一個詞接二連三輸入Decoder進行訓練,而引入Attention機制以後,我須要可以人爲控制一個詞一個詞進行輸入(由於輸入每一個詞到Decoder,須要再作一些運算),因此在代碼中會看到我使用了for循環,循環trg_len-1次(開頭的<SOS>我手動輸入,因此循環少一次)

而且訓練過程當中我使用了一種叫作Teacher Forcing的機制,保證訓練速度的同時增長魯棒性,若是不瞭解Teacher Forcing能夠看個人這篇文章

思考一下for循環中應該要作哪些事?首先要將變量傳入Decoder,因爲Attention的計算是在Decoder的內部進行的,因此我須要將dec_inputsenc_output這三個變量傳入Decoder,Decoder會返回dec_output以及新的s。以後根據機率對dec_output作Teacher Forcing便可

Seq2Seq細節就這麼多,下面給出代碼

class Seq2Seq(nn.Module):
    def __init__(self, encoder, decoder, device):
        super().__init__()
        self.encoder = encoder
        self.decoder = decoder
        self.device = device
        
    def forward(self, src, trg, teacher_forcing_ratio = 0.5):
        
        # src = [src_len, batch_size]
        # trg = [trg_len, batch_size]
        # teacher_forcing_ratio is probability to use teacher forcing
        
        batch_size = src.shape[1]
        trg_len = trg.shape[0]
        trg_vocab_size = self.decoder.output_dim
        
        # tensor to store decoder outputs
        outputs = torch.zeros(trg_len, batch_size, trg_vocab_size).to(self.device)
        
        # enc_output is all hidden states of the input sequence, back and forwards
        # s is the final forward and backward hidden states, passed through a linear layer
        enc_output, s = self.encoder(src)
                
        # first input to the decoder is the <sos> tokens
        dec_input = trg[0,:]
        
        for t in range(1, trg_len):
            
            # insert dec_input token embedding, previous hidden state and all encoder hidden states
            # receive output tensor (predictions) and new hidden state
            dec_output, s = self.decoder(dec_input, s, enc_output)
            
            # place predictions in a tensor holding predictions for each token
            outputs[t] = dec_output
            
            # decide if we are going to use teacher forcing or not
            teacher_force = random.random() < teacher_forcing_ratio
            
            # get the highest predicted token from our predictions
            top1 = dec_output.argmax(1) 
            
            # if teacher forcing, use actual next token as next input
            # if not, use predicted token
            dec_input = trg[t] if teacher_force else top1

        return outputs
複製代碼

Decoder

Decoder我用的是單向單層GRU

Decoder部分實際上也就是三個公式

c = a t H s t = G R U ( e m b ( y t ) , c , s t 1 ) y t ^ = f ( e m b ( y t ) , c , s t ) c=a_tH\\ s_t=GRU(emb(y_t), c, s_{t-1})\\ \hat{y_t}=f(emb(y_t), c, s_t)

H H 指的是Encoder中的變量enc_output e m b ( y t ) emb(y_t) 指的是將dec_input通過WordEmbedding後獲得的結果, f ( ) f() 函數實際上就是爲了轉換維度,由於須要的輸出是TRG_VOCAB_SIZE大小。其中有個細節,GRU的參數只有兩個,一個輸入,一個隱藏層輸入,可是上面的公式有三個變量,因此咱們應該選一個做爲隱藏層輸入,另外兩個"整合"一下,做爲輸入

咱們從第一個公式正推各個變量的維度是什麼

首先在Encoder中最開始先調用一次Attention,獲得權重 a t a_t ,它的維度是[batch_size, src_len],而 H H 的維度是[src_len, batch_size, enc_hid_dim*2],它倆要相乘,同時應該保留batch_size這個維度,因此應該先將 a t a_t 擴展一維,而後調換一下 H H 維度的順序,以後再按照batch相乘(即同一個batch內的矩陣相乘)

a = a.unsqueeze(1) # [batch_size, 1, src_len]
H = H.transpose(0, 1) # [batch_size, src_len, enc_hid_dim*2]
c = torch.bmm(a, h) # [batch_size, 1, enc_hid_dim*2]
複製代碼

前面也說了,因爲GRU不須要三個變量,因此須要將 e m b ( y t ) emb(y_t) c c 整合一下, y t y_t 實際上就是Seq2Seq類中的dec_input變量,它的維度是[batch_size],所以先將 y t y_t 擴展一個維度,再經過WordEmbedding,這樣他就變成[batch_size, 1, emb_dim]。最後對 c c e m b ( y t ) emb(y_t) 進行concat

y = y.unsqueeze(1) # [batch_size, 1]
emb_y = self.emb(y) # [batch_size, 1, emb_dim]
rnn_input = torch.cat((emb_y, c), dim=2) # [batch_size, 1, emb_dim+enc_hid_dim*2]
複製代碼

s t 1 s_{t-1} 的維度是[batch_size, dec_hid_dim],因此應該先將其拓展一個維度

rnn_input = rnn_input.transpose(0, 1) # [1, batch_size, emb_dim+enc_hid_dim*2]
s = s.unsqueeze(1) # [batch_size, 1, dec_hid_dim]

# dec_output = [1, batch_size, dec_hid_dim]
# dec_hidden = [1, batch_size, dec_hid_dim] = s (new, is not s previously)
dec_output, dec_hidden = self.rnn(rnn_input, s)
複製代碼

最後一個公式,須要將三個變量所有拼接在一塊兒,而後經過一個全鏈接神經網絡,獲得最終的預測。咱們先分析下這個三個變量的維度, e m b ( y t ) emb(y_t) 的維度是[batch_size, 1, emb_dim] c c 的維度是[batch_size, 1, enc_hid_dim] s t s_t 的維度是[1, batch_size, dec_hid_dim],所以咱們能夠像下面這樣把他們所有拼接起來

emd_y = emb_y.squeeze(1) # [batch_size, emb_dim]
c = w.squeeze(1) # [batch_size, enc_hid_dim*2]
s = s.squeeze(0) # [batch_size, dec_hid_dim]

fc_input = torch.cat((emb_y, c, s), dim=1) # [batch_size, enc_hid_dim*2+dec_hid_dim+emb_hid] 
複製代碼

以上就是Decoder部分的細節,下面給出代碼(上面的那些只是示例代碼,和下面代碼變量名可能不同)

class Decoder(nn.Module):
    def __init__(self, output_dim, emb_dim, enc_hid_dim, dec_hid_dim, dropout, attention):
        super().__init__()
        self.output_dim = output_dim
        self.attention = attention
        self.embedding = nn.Embedding(output_dim, emb_dim)
        self.rnn = nn.GRU((enc_hid_dim * 2) + emb_dim, dec_hid_dim)
        self.fc_out = nn.Linear((enc_hid_dim * 2) + dec_hid_dim + emb_dim, output_dim)
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, dec_input, s, enc_output):
             
        # dec_input = [batch_size]
        # s = [batch_size, dec_hid_dim]
        # enc_output = [src_len, batch_size, enc_hid_dim * 2]
        
        dec_input = dec_input.unsqueeze(1) # dec_input = [batch_size, 1]
        
        embedded = self.dropout(self.embedding(dec_input)).transpose(0, 1) # embedded = [1, batch_size, emb_dim]
        
        # a = [batch_size, 1, src_len] 
        a = self.attention(s, enc_output).unsqueeze(1)
        
        # enc_output = [batch_size, src_len, enc_hid_dim * 2]
        enc_output = enc_output.transpose(0, 1)

        # c = [1, batch_size, enc_hid_dim * 2]
        c = torch.bmm(a, enc_output).transpose(0, 1)

        # rnn_input = [1, batch_size, (enc_hid_dim * 2) + emb_dim]
        rnn_input = torch.cat((embedded, c), dim = 2)
            
        # dec_output = [src_len(=1), batch_size, dec_hid_dim]
        # dec_hidden = [n_layers * num_directions, batch_size, dec_hid_dim]
        dec_output, dec_hidden = self.rnn(rnn_input, s.unsqueeze(0))
        
        # embedded = [batch_size, emb_dim]
        # dec_output = [batch_size, dec_hid_dim]
        # c = [batch_size, enc_hid_dim * 2]
        embedded = embedded.squeeze(0)
        dec_output = dec_output.squeeze(0)
        c = c.squeeze(0)
        
        # pred = [batch_size, output_dim]
        pred = self.fc_out(torch.cat((dec_output, c, embedded), dim = 1))
        
        return pred, dec_hidden.squeeze(0)
複製代碼

定義模型

INPUT_DIM = len(SRC.vocab)
OUTPUT_DIM = len(TRG.vocab)
ENC_EMB_DIM = 256
DEC_EMB_DIM = 256
ENC_HID_DIM = 512
DEC_HID_DIM = 512
ENC_DROPOUT = 0.5
DEC_DROPOUT = 0.5

attn = Attention(ENC_HID_DIM, DEC_HID_DIM)
enc = Encoder(INPUT_DIM, ENC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM, ENC_DROPOUT)
dec = Decoder(OUTPUT_DIM, DEC_EMB_DIM, ENC_HID_DIM, DEC_HID_DIM, DEC_DROPOUT, attn)

model = Seq2Seq(enc, dec, device).to(device)
TRG_PAD_IDX = TRG.vocab.stoi[TRG.pad_token]
criterion = nn.CrossEntropyLoss(ignore_index = TRG_PAD_IDX).to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
複製代碼

倒數第二行CrossEntropyLoss()中的參數不多見,ignore_index=TRG_PAD_IDX,這個參數的做用是忽略某一類別,不計算其loss,可是要注意,忽略的是真實值中的類別,例以下面的代碼,真實值的類別都是1,而預測值所有預測認爲是2(下標從0開始),同時loss function設置忽略第一類的loss,此時會打印出0

label = torch.tensor([1, 1, 1])
pred = torch.tensor([[0.1, 0.2, 0.6], [0.2, 0.1, 0.8], [0.1, 0.1, 0.9]])
loss_fn = nn.CrossEntropyLoss(ignore_index=1)
print(loss_fn(pred, label).item()) # 0
複製代碼

若是設置loss function忽略第二類,此時loss並不會爲0

label = torch.tensor([1, 1, 1])
pred = torch.tensor([[0.1, 0.2, 0.6], [0.2, 0.1, 0.8], [0.1, 0.1, 0.9]])
loss_fn = nn.CrossEntropyLoss(ignore_index=2)
print(loss_fn(pred, label).item()) # 1.359844
複製代碼

最後給出完整代碼連接(須要科學的力量) Github項目地址:nlp-tutorial

相關文章
相關標籤/搜索