attention

 Attention Mechanism  
Can I have your Attention please! The introduction of the Attention Mechanism in deep learning has improved the success of various models in recent years, and continues to be an omnipresent component in state-of-the-art models. Therefore, it is vital that we pay Attention to Attention and how it goes about achieving its effectiveness. 
In this article, I will be covering the main concepts behind Attention, including an implementation of a sequence-to-sequence Attention model, followed by the application of Attention in Transformers and how they can be used for state-of-the-art results. It is advised that you have some knowledge of Recurrent Neural Networks (RNNs) and their variants, or an understanding of how sequence-to-sequence models work. 
What is Attention? 
When we think about the English word “Attention”, we know that it means directing your focus at something and taking greater notice. The Attention mechanism in Deep Learning is based off this concept of directing your focus, and it pays greater attention to certain factors when processing the data. 
In broad terms, Attention is one component of a network’s architecture, and is in charge of managing and quantifying the interdependence: 
Between the input and output elements (General Attention)
Within the input elements (Self-Attention)
Let me give you an example of how Attention works in a translation task. Say we have the sentence “How was your day”, which we would like to translate to the French version - “Comment se passe ta journée”. What the Attention component of the network will do for each word in the output sentence is map the important and relevant words from the input sentence and assign higher weights to these words, enhancing the accuracy of the output prediction. 
  
 
Weights are assigned to input words at each step of the translation 
The above explanation of Attention is very broad and vague due to the various types of Attention mechanisms available. But fret not, you’ll gain a clearer picture of how Attention works and achieves its objectives further in the article. As the Attention mechanism has undergone multiple adaptations over the years to suit various tasks, there are many different versions of Attention that are applied. We will only cover the more popular adaptations here, which are its usage in sequence-to-sequence models and the more recent Self-Attention. 
While Attention does have its application in other fields of deep learning such as Computer Vision, its main breakthrough and success comes from its application in Natural Language Processing (NLP) tasks. This is due to the fact that Attention was introduced to address the problem of long sequences in Machine Translation, which is also a problem for most other NLP tasks as well. 
Attention in Sequence-to-Sequence Models 
Most articles on the Attention Mechanism will use the example of sequence-to-sequence (seq2seq) models to explain how it works. This is because Attention was originally introduced as a solution to address the main issue surrounding seq2seq models, and to great success. If you are unfamiliar with seq2seq models, also known as the Encoder-Decoder model, I recommend having a read through this article to get you up to speed. 
 
Overall process of a Sequence-to-sequence model 
The standard seq2seq model is generally unable to accurately process long input sequences, since only the last hidden state of the encoder RNN is used as the context vector for the decoder. On the other hand, the Attention Mechanism directly addresses this issue as it retains and utilises all the hidden states of the input sequence during the decoding process. It does this by creating a unique mapping between each time step of the decoder output to all the encoder hidden states. This means that for each output that the decoder makes, it has access to the entire input sequence and can selectively pick out specific elements from that sequence to produce the output. 
Therefore, the mechanism allows the model to focus and place more “Attention” on the relevant parts of the input sequence as needed. 
Types of Attention 
 
Comparing Bahdanau Attention with Luong Attention 
Before we delve into the specific mechanics behind Attention, we must note that there are 2 different major types of Attention: 
Bahdanau Attention
Luong Attention
While the underlying principles of Attention are the same in these 2 types, their differences lie mainly in their architectures and computations. 
Bahdanau Attention 
 
Overall process for Bahdanau Attention seq2seq model 
The first type of Attention, commonly referred to as Additive Attention, came from a paper by Dzmitry Bahdanau, which explains the less-descriptive original name. The paper aimed to improve the sequence-to-sequence model in machine translation by aligning the decoder with the relevant input sentences and implementing Attention. The entire step-by-step process of applying Attention in Bahdanau’s paper is as follows: 
Producing the Encoder Hidden States - Encoder produces hidden states of each element in the input sequence
Calculating Alignment Scores between the previous decoder hidden state and each of the encoder’s hidden states are calculated (Note: The last encoder hidden state can be used as the first hidden state in the decoder)
Softmaxing the Alignment Scores - the alignment scores for each encoder hidden state are combined and represented in a single vector and subsequently softmaxed
Calculating the Context Vector - the encoder hidden states and their respective alignment scores are multiplied to form the context vector
Decoding the Output - the context vector is concatenated with the previous decoder output and fed into the Decoder RNN for that time step along with the previous decoder hidden state to produce a new output
The process (steps 2-5) repeats itself for each time step of the decoder until an token is produced or output is past the specified maximum length
 
Flow of calculating Attention weights in Bahdanau Attention 
Now that we have a high-level understanding of the flow of the Attention mechanism for Bahdanau, let’s take a look at the inner workings and computations involved, together with some code implementation of a language seq2seq model with Attention in PyTorch. 
1. Producing the Encoder Hidden States 
 
Encoder RNNs will produce a hidden state for each input 
For our first step, we’ll be using an RNN or any of its variants (e.g. LSTM, GRU) to encode the input sequence. After passing the input sequence through the encoder RNN, a hidden state/output will be produced for each input passed in. Instead of using only the hidden state at the final time step, we’ll be carrying forward all the hidden states produced by the encoder to the next step. 
class EncoderLSTM(nn.Module):
  def __init__(self, input_size, hidden_size, n_layers=1, drop_prob=0):
    super(EncoderLSTM, self).__init__()
    self.hidden_size = hidden_size
    self.n_layers = n_layers

    self.embedding = nn.Embedding(input_size, hidden_size)
    self.lstm = nn.LSTM(hidden_size, hidden_size, n_layers, dropout=drop_prob, batch_first=True)

  def forward(self, inputs, hidden):
    # Embed input words
    embedded = self.embedding(inputs)
    # Pass the embedded word vectors into LSTM and return all outputs
    output, hidden = self.lstm(embedded, hidden)
    return output, hidden

  def init_hidden(self, batch_size=1):
    return (torch.zeros(self.n_layers, batch_size, self.hidden_size, device=device),
            torch.zeros(self.n_layers, batch_size, self.hidden_size, device=device))
 
In the code implementation of the encoder above, we’re first embedding the input words into word vectors (assuming that it’s a language task) and then passing it through an LSTM. The encoder over here is exactly the same as a normal encoder-decoder structure without Attention. 
2. Calculating Alignment Scores 
For these next 3 steps, we will be going through the processes that happen in the Attention Decoder and discuss how the Attention mechanism is utilised. The class BahdanauDecoderLSTM defined below encompasses these 3 steps in the forward function. 
class BahdanauDecoder(nn.Module):
  def __init__(self, hidden_size, output_size, n_layers=1, drop_prob=0.1):
    super(BahdanauDecoder, self).__init__()
    self.hidden_size = hidden_size
    self.output_size = output_size
    self.n_layers = n_layers
    self.drop_prob = drop_prob

    self.embedding = nn.Embedding(self.output_size, self.hidden_size)
    
    self.fc_hidden = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
    self.fc_encoder = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
    self.weight = nn.Parameter(torch.FloatTensor(1, hidden_size))
    self.attn_combine = nn.Linear(self.hidden_size * 2, self.hidden_size)
    self.dropout = nn.Dropout(self.drop_prob)
    self.lstm = nn.LSTM(self.hidden_size*2, self.hidden_size, batch_first=True)
    self.classifier = nn.Linear(self.hidden_size, self.output_size)

  def forward(self, inputs, hidden, encoder_outputs):
    encoder_outputs = encoder_outputs.squeeze()
    # Embed input words
    embedded = self.embedding(inputs).view(1, -1)
    embedded = self.dropout(embedded)
    
    # Calculating Alignment Scores
    x = torch.tanh(self.fc_hidden(hidden[0])+self.fc_encoder(encoder_outputs))
    alignment_scores = x.bmm(self.weight.unsqueeze(2))  
    
    # Softmaxing alignment scores to get Attention weights
    attn_weights = F.softmax(alignment_scores.view(1,-1), dim=1)
    
    # Multiplying the Attention weights with encoder outputs to get the context vector
    context_vector = torch.bmm(attn_weights.unsqueeze(0),
                             encoder_outputs.unsqueeze(0))
    
    # Concatenating context vector with embedded input word
    output = torch.cat((embedded, context_vector[0]), 1).unsqueeze(0)
    # Passing the concatenated vector as input to the LSTM cell
    output, hidden = self.lstm(output, hidden)
    # Passing the LSTM output through a Linear layer acting as a classifier
    output = F.log_softmax(self.classifier(output[0]), dim=1)
    return output, hidden, attn_weights
 
After obtaining all of our encoder outputs, we can start using the decoder to produce outputs. At each time step of the decoder, we have to calculate the alignment score of each encoder output with respect to the decoder input and hidden state at that time step. The alignment score is the essence of the Attention mechanism, as it quantifies the amount of “Attention” the decoder will place on each of the encoder outputs when producing the next output. 
The alignment scores for Bahdanau Attention are calculated using the hidden state produced by the decoder in the previous time step and the encoder outputs with the following equation: 
scorealignment=Wcombined⋅tanh(Wdecoder⋅Hdecoder+Wencoder⋅Hencoder)scorealignment=Wcombined⋅tanh(Wdecoder⋅Hdecoder+Wencoder⋅Hencoder) 
The decoder hidden state and encoder outputs will be passed through their individual Linear layer and have their own individual trainable weights. 
 
Linear layers for encoder outputs and decoder hidden states 
In the illustration above, the hidden size is 3 and the number of encoder outputs is 2. 
Thereafter, they will be added together before being passed through a tanh activation function. The decoder hidden state is added to each encoder output in this case. 
 
Above outputs combined and tanh applied 
Lastly, the resultant vector from the previous few steps will undergo matrix multiplication with a trainable vector, obtaining a final alignment score vector which holds a score for each encoder output. 
 
Matrix Multiplication to obtain Alignment score 
Note: As there is no previous hidden state or output for the first decoder step, the last encoder hidden state and a Start Of String (<SOS>) token can be used to replace these two, respectively. 
3. Softmaxing the Alignment Scores 
After generating the alignment scores vector in the previous step, we can then apply a softmax on this vector to obtain the attention weights. The softmax function will cause the values in the vector to sum up to 1 and each individual value will lie between 0 and 1, therefore representing the weightage each input holds at that time step. 
 
Alignment scores are softmaxed 
4. Calculating the Context Vector 
After computing the attention weights in the previous step, we can now generate the context vector by doing an element-wise multiplication of the attention weights with the encoder outputs. 
Due to the softmax function in the previous step, if the score of a specific input element is closer to 1 its effect and influence on the decoder output is amplified, whereas if the score is close to 0, its influence is drowned out and nullified. 
 
Context Vector is derived from the weights and encoder outputs 
5. Decoding the Output 
The context vector we produced will then be concatenated with the previous decoder output. It is then fed into the decoder RNN cell to produce a new hidden state and the process repeats itself from step 2. The final output for the time step is obtained by passing the new hidden state through a Linear layer, which acts as a classifier to give the probability scores of the next predicted word. 
 
Context vector and previous output will give new decoder hidden state 
Steps 2 to 4 are repeated until the decoder generates an End Of Sentence token or the output length exceeds a specified maximum length. 
Luong Attention 
 
Overall process for Luong Attention seq2seq model 
The second type of Attention was proposed by Thang Luong in this paper. It is often referred to as Multiplicative Attention and was built on top of the Attention mechanism proposed by Bahdanau. The two main differences between Luong Attention and Bahdanau Attention are: 
The way that the alignment score is calculated
The position at which the Attention mechanism is being introduced in the decoder
There are three types of alignment scoring functions proposed in Luong’s paper compared to Bahdanau’s one type. Also, the general structure of the Attention Decoder is different for Luong Attention, as the context vector is only utilised after the RNN produced the output for that time step. We will explore these differences in greater detail as we go through the Luong Attention process, which is: 
Producing the Encoder Hidden States - Encoder produces hidden states of each element in the input sequence
Decoder RNN - the previous decoder hidden state and decoder output is passed through the Decoder RNN to generate a new hidden state for that time step
Calculating Alignment Scores - using the new decoder hidden state and the encoder hidden states, alignment scores are calculated
Softmaxing the Alignment Scores - the alignment scores for each encoder hidden state are combined and represented in a single vector and subsequently softmaxed
Calculating the Context Vector - the encoder hidden states and their respective alignment scores are multiplied to form the context vector
Producing the Final Output - the context vector is concatenated with the decoder hidden state generated in step 2 as passed through a fully connected layer to produce a new output
The process (steps 2-6) repeats itself for each time step of the decoder until an token is produced or output is past the specified maximum length
As we can already see above, the order of steps in Luong Attention is different from Bahdanau Attention. The code implementation and some calculations in this process is different as well, which we will go through now. 
1. Producing the Encoder Hidden States 
Just as in Bahdanau Attention, the encoder produces a hidden state for each element in the input sequence. 
2. Decoder RNN 
Unlike in Bahdanau Attention, the decoder in Luong Attention uses the RNN in the first step of the decoding process rather than the last. The RNN will take the hidden state produced in the previous time step and the word embedding of the final output from the previous time step to produce a new hidden state which will be used in the subsequent steps 
class LuongDecoder(nn.Module):
  def __init__(self, hidden_size, output_size, attention, n_layers=1, drop_prob=0.1):
    super(LuongDecoder, self).__init__()
    self.hidden_size = hidden_size
    self.output_size = output_size
    self.n_layers = n_layers
    self.drop_prob = drop_prob
    
    # The Attention Mechanism is defined in a separate class
    self.attention = attention
    
    self.embedding = nn.Embedding(self.output_size, self.hidden_size)
    self.dropout = nn.Dropout(self.drop_prob)
    self.lstm = nn.LSTM(self.hidden_size, self.hidden_size)
    self.classifier = nn.Linear(self.hidden_size*2, self.output_size)
    
  def forward(self, inputs, hidden, encoder_outputs):
    # Embed input words
    embedded = self.embedding(inputs).view(1,1,-1)
    embedded = self.dropout(embedded)
    
    # Passing previous output word (embedded) and hidden state into LSTM cell
    lstm_out, hidden = self.lstm(embedded, hidden)
    
    # Calculating Alignment Scores - see Attention class for the forward pass function
    alignment_scores = self.attention(lstm_out,encoder_outputs)
    # Softmaxing alignment scores to obtain Attention weights
    attn_weights = F.softmax(alignment_scores.view(1,-1), dim=1)
    
    # Multiplying Attention weights with encoder outputs to get context vector
    context_vector = torch.bmm(attn_weights.unsqueeze(0),encoder_outputs)
    
    # Concatenating output from LSTM with context vector
    output = torch.cat((lstm_out, context_vector),-1)
    # Pass concatenated vector through Linear layer acting as a Classifier
    output = F.log_softmax(self.classifier(output[0]), dim=1)
    return output, hidden, attn_weights
  
class Attention(nn.Module):
  def __init__(self, hidden_size, method="dot"):
    super(Attention, self).__init__()
    self.method = method
    self.hidden_size = hidden_size
    
    # Defining the layers/weights required depending on alignment scoring method
    if method == "general":
      self.fc = nn.Linear(hidden_size, hidden_size, bias=False)
      
    elif method == "concat":
      self.fc = nn.Linear(hidden_size, hidden_size, bias=False)
      self.weight = nn.Parameter(torch.FloatTensor(1, hidden_size))
  
  def forward(self, decoder_hidden, encoder_outputs):
    if self.method == "dot":
      # For the dot scoring method, no weights or linear layers are involved
      return encoder_outputs.bmm(decoder_hidden.view(1,-1,1)).squeeze(-1)
    
    elif self.method == "general":
      # For general scoring, decoder hidden state is passed through linear layers to introduce a weight matrix
      out = self.fc(decoder_hidden)
      return encoder_outputs.bmm(out.view(1,-1,1)).squeeze(-1)
    
    elif self.method == "concat":
      # For concat scoring, decoder hidden state and encoder outputs are concatenated first
      out = torch.tanh(self.fc(decoder_hidden+encoder_outputs))
      return out.bmm(self.weight.unsqueeze(-1)).squeeze(-1)
 
3. Calculating Alignment Scores 
In Luong Attention, there are three different ways that the alignment scoring function is defined- dot, general and concat. These scoring functions make use of the encoder outputs and the decoder hidden state produced in the previous step to calculate the alignment scores. 
Dot
 The first one is the dot scoring function. This is the simplest of the functions; to produce the alignment score we only need to take the hidden states of the encoder and multiply them by the hidden state of the decoder.
scorealignment=Hencoder⋅Hdecoderscorealignment=Hencoder⋅Hdecoder 
General
 The second type is called general and is similar to the dot function, except that a weight matrix is added into the equation as well.
scorealignment=W(Hencoder⋅Hdecoder)scorealignment=W(Hencoder⋅Hdecoder) 
Concat
 The last function is slightly similar to the way that alignment scores are calculated in Bahdanau Attention, whereby the decoder hidden state is added to the encoder hidden states.
scorealignment=W⋅tanh(Wcombined(Hencoder+Hdecoder))scorealignment=W⋅tanh(Wcombined(Hencoder+Hdecoder)) 
However, the difference lies in the fact that the decoder hidden state and encoder hidden states are added together first before being passed through a Linear layer. This means that the decoder hidden state and encoder hidden state will not have their individual weight matrix, but a shared one instead, unlike in Bahdanau Attention.After being passed through the Linear layer, a tanh activation function will be applied on the output before being multiplied by a weight matrix to produce the alignment score. 
4. Softmaxing the Alignment Scores 
Similar to Bahdanau Attention, the alignment scores are softmaxed so that the weights will be between 0 to 1. 
5. Calculating the Context Vector 
Again, this step is the same as the one in Bahdanau Attention where the attention weights are multiplied with the encoder outputs. 
6. Producing the Final Output 
In the last step, the context vector we just produced is concatenated with the decoder hidden state we generated in step 2. This combined vector is then passed through a Linear layer which acts as a classifier for us to obtain the probability scores of the next predicted word. 
Testing The Model 
Since we’ve defined the structure of the Attention encoder-decoder model and understood how it works, let’s see how we can use it for an NLP task - Machine Translation. 
Ready to build, train, and deploy AI? 
Get started with FloydHub's collaborative AI platform for free 
Try FloydHub for free 
We will be using English to German sentence pairs obtained from the Tatoeba Project, and the compiled sentences pairs can be found at this link. You can run the code implementation in this article on FloydHub using their GPUs on the cloud by clicking the following link and using the main.ipynb notebook. 
 
This will speed up the training process significantly. Alternatively, the link to the GitHub repository can be found here. 
The goal of this implementation is not to develop a complete English to German translator, but rather just as a sanity check to ensure that our model is able to learn and fit to a set of training data. I will briefly go through the data preprocessing steps before running through the training procedure. 
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import pandas
import spacy
from spacy.lang.en import English
from spacy.lang.de import German
import matplotlib.pyplot as plt
import matplotlib.ticker as ticker
from tqdm import tqdm_notebook
import random
from collections import Counter

if torch.cuda.is_available:
    device = torch.device("cuda")
else:
    device = torch.device("cpu")
 
We start by importing the relevant libraries and defining the device we are running our training on (GPU/CPU). If you’re using FloydHub with GPU to run this code, the training time will be significantly reduced. In the next code block, we’ll be doing our data preprocessing steps: 
Tokenizing the sentences and creating our vocabulary dictionaries
Assigning each word in our vocabulary to integer indexes
Converting our sentences into their word token indexes
# Reading the English-German sentences pairs from the file
with open("deu.txt","r+") as file:
  deu = [x[:-1] for x in file.readlines()]
en = []
de = []
for line in deu:
  en.append(line.split("\t")[0])
  de.append(line.split("\t")[1])

# Setting the number of training sentences we'll use
training_examples = 10000
# We'll be using the spaCy's English and German tokenizers
spacy_en = English()
spacy_de = German()

en_words = Counter()
de_words = Counter()
en_inputs = []
de_inputs = []

# Tokenizing the English and German sentences and creating our word banks for both languages
for i in tqdm_notebook(range(training_examples)):
    en_tokens = spacy_en(en[i])
    de_tokens = spacy_de(de[i])
    if len(en_tokens)==0 or len(de_tokens)==0:
        continue
    for token in en_tokens:
        en_words.update([token.text.lower()])
    en_inputs.append([token.text.lower() for token in en_tokens] + ['_EOS'])
    for token in de_tokens:
        de_words.update([token.text.lower()])
    de_inputs.append([token.text.lower() for token in de_tokens] + ['_EOS'])

# Assigning an index to each word token, including the Start Of String(SOS), End Of String(EOS) and Unknown(UNK) tokens
en_words = ['_SOS','_EOS','_UNK'] + sorted(en_words,key=en_words.get,reverse=True)
en_w2i = {o:i for i,o in enumerate(en_words)}
en_i2w = {i:o for i,o in enumerate(en_words)}
de_words = ['_SOS','_EOS','_UNK'] + sorted(de_words,key=de_words.get,reverse=True)
de_w2i = {o:i for i,o in enumerate(de_words)}
de_i2w = {i:o for i,o in enumerate(de_words)}

# Converting our English and German sentences to their token indexes
for i in range(len(en_inputs)):
    en_sentence = en_inputs[i]
    de_sentence = de_inputs[i]
    en_inputs[i] = [en_w2i[word] for word in en_sentence]
    de_inputs[i] = [de_w2i[word] for word in de_sentence]
 
Since we’ve already defined our Encoder and Attention Decoder model classes earlier, we can now instantiate the models. 
hidden_size = 256
encoder = EncoderLSTM(len(en_words), hidden_size).to(device)
attn = Attention(hidden_size,"concat")
decoder = LuongDecoder(hidden_size,len(de_words),attn).to(device)

lr = 0.001
encoder_optimizer = optim.Adam(encoder.parameters(), lr=lr)
decoder_optimizer = optim.Adam(decoder.parameters(), lr=lr)
 
We’ll be testing the LuongDecoder model with the scoring function set as concat. During our training cycle, we’ll be using a method called teacher forcing for 50% of the training inputs, which uses the real target outputs as the input to the next step of the decoder instead of our decoder output for the previous time step. This allows the model to converge faster, although there are some drawbacks involved (e.g. instability of trained model). 
EPOCHS = 10
teacher_forcing_prob = 0.5
encoder.train()
decoder.train()
tk0 = tqdm_notebook(range(1,EPOCHS+1),total=EPOCHS)
for epoch in tk0:
    avg_loss = 0.
    tk1 = tqdm_notebook(enumerate(en_inputs),total=len(en_inputs),leave=False)
    for i, sentence in tk1:
        loss = 0.
        h = encoder.init_hidden()
        encoder_optimizer.zero_grad()
        decoder_optimizer.zero_grad()
        inp = torch.tensor(sentence).unsqueeze(0).to(device)
        encoder_outputs, h = encoder(inp,h)
        
        #First decoder input will be the SOS token
        decoder_input = torch.tensor([en_w2i['_SOS']],device=device)
        #First decoder hidden state will be last encoder hidden state
        decoder_hidden = h
        output = []
        teacher_forcing = True if random.random() < teacher_forcing_prob else False
        
        for ii in range(len(de_inputs[i])):
          decoder_output, decoder_hidden, attn_weights = decoder(decoder_input, decoder_hidden, encoder_outputs)
          # Get the index value of the word with the highest score from the decoder output
          top_value, top_index = decoder_output.topk(1)
          if teacher_forcing:
            decoder_input = torch.tensor([de_inputs[i][ii]],device=device)
          else:
            decoder_input = torch.tensor([top_index.item()],device=device)
          output.append(top_index.item())
          # Calculate the loss of the prediction against the actual word
          loss += F.nll_loss(decoder_output.view(1,-1), torch.tensor([de_inputs[i][ii]],device=device))
        loss.backward()
        encoder_optimizer.step()
        decoder_optimizer.step()
        avg_loss += loss.item()/len(en_inputs)
    tk0.set_postfix(loss=avg_loss)
  # Save model after every epoch (Optional)
torch.save({"encoder":encoder.state_dict(),"decoder":decoder.state_dict(),"e_optimizer":encoder_optimizer.state_dict(),"d_optimizer":decoder_optimizer},"./model.pt")
 
Using our trained model, let’s visualise some of the outputs that the model produces and the attention weights the model assigns to each input element. 
encoder.eval()
decoder.eval()
# Choose a random sentences
i = random.randint(0,len(en_inputs)-1)
h = encoder.init_hidden()
inp = torch.tensor(en_inputs[i]).unsqueeze(0).to(device)
encoder_outputs, h = encoder(inp,h)

decoder_input = torch.tensor([en_w2i['_SOS']],device=device)
decoder_hidden = h
output = []
attentions = []
while True:
  decoder_output, decoder_hidden, attn_weights = decoder(decoder_input, decoder_hidden, encoder_outputs)
  _, top_index = decoder_output.topk(1)
  decoder_input = torch.tensor([top_index.item()],device=device)
  # If the decoder output is the End Of Sentence token, stop decoding process
  if top_index.item() == de_w2i["_EOS"]:
    break
  output.append(top_index.item())
  attentions.append(attn_weights.squeeze().cpu().detach().numpy())
print("English: "+ " ".join([en_i2w[x] for x in en_inputs[i]]))
print("Predicted: " + " ".join([de_i2w[x] for x in output]))
print("Actual: " + " ".join([de_i2w[x] for x in de_inputs[i]]))

# Plotting the heatmap for the Attention weights
fig = plt.figure(figsize=(12,9))
ax = fig.add_subplot(111)
cax = ax.matshow(np.array(attentions))
fig.colorbar(cax)
ax.set_xticklabels(['']+[en_i2w[x] for x in en_inputs[i]])
ax.set_yticklabels(['']+[de_i2w[x] for x in output])
ax.xaxis.set_major_locator(ticker.MultipleLocator(1))
ax.yaxis.set_major_locator(ticker.MultipleLocator(1))
plt.show()
 
[Out]: English: she kissed him . _EOS
       Predicted: sie küsste ihn .
       Actual: sie küsste ihn . _EOS
 
 
Plot showing the weights assigned to each input word for each output 
From the example above, we can see that for each output word from the decoder, the weights assigned to the input words are different and we can see the relationship between the inputs and outputs that the model is able to draw. You can try this on a few more examples to test the results of the translator. 
In our training, we have clearly overfitted our model to the training sentences. If we were to test the trained model on sentences it has never seen before, it is unlikely to produce decent results. Nevertheless, this process acts as a sanity check to ensure that our model works and is able to function end-to-end and learn. 
The challenge of training an effective model can be attributed largely to the lack of training data and training time. Due to the complex nature of the different languages involved and a large number of vocabulary and grammatical permutations, an effective model will require tons of data and training time before any results can be seen on evaluation data. 
Conclusion 
The Attention mechanism has revolutionised the way we create NLP models and is currently a standard fixture in most state-of-the-art NLP models. This is because it enables the model to “remember” all the words in the input and focus on specific words when formulating a response. 
We covered the early implementations of Attention in seq2seq models with RNNs in this article. However, the more recent adaptations of Attention has seen models move beyond RNNs to Self-Attention and the realm of Transformer models. Google’s BERT, OpenAI’s GPT and the more recent XLNet are the more popular NLP models today and are largely based on self-attention and the Transformer architecture. 
I’ll be covering the workings of these models and how you can implement and fine-tune them for your own downstream tasks in my next article. Stay tuned! 
转自：https://blog.floydhub.com/attention-mechanism/