用于文本生成的注意力机制

使用 PyTorch 的文本深度学习

Shubham Jain

Instructor

文本处理中的歧义

"The monkey ate that banana because it was too hungry"
词语"it"指代什么？

人类与机器

注意力机制

为词分配重要性
让机器理解与人类理解一致

注意力图

¹ Xie, Huiqiang & Qin, Zhijin & Li, Geoffrey & Juang, Biing-Hwang. (2020). Deep Learning Enabled Semantic Communication Systems

自注意力与多头注意力

自注意力：为句内词语分配权重
- "The cat, which was on the roof, was scared"
- 将"was scared"关联到"The cat"
多头注意力：像多束聚光灯，捕捉不同侧面
- 理解"was scared"可关联到
- "The cat""the roof"或"was on"

注意力机制——设定词表与数据

data = ["the cat sat on the mat", ...]

vocab = set(' '.join(data).split())

word_to_ix = {word: i for i, word in enumerate(vocab)}
ix_to_word = {i: word for word, i in word_to_ix.items()}

pairs = [sentence.split() for sentence in data]
input_data = [[word_to_ix[word] for word in sentence[:-1]] for sentence in pairs]
target_data = [word_to_ix[sentence[-1]] for sentence in pairs]
inputs = [torch.tensor(seq, dtype=torch.long) for seq in input_data]
targets = torch.tensor(target_data, dtype=torch.long)

模型定义

embedding_dim = 10
hidden_dim = 16


class RNNWithAttentionModel(nn.Module):
    def __init__(self):
        super(RNNWithAttentionModel, self).__init__()

        self.embeddings = nn.Embedding(vocab_size, embedding_dim)
        self.rnn = nn.RNN(embedding_dim, hidden_dim, batch_first=True)

        self.attention = nn.Linear(hidden_dim, 1)

        self.fc = nn.Linear(hidden_dim, vocab_size)

带注意力的前向传播

def forward(self, x):
    x = self.embeddings(x)
    out, _ = self.rnn(x)

    attn_weights = torch.nn.functional.softmax(self.attention(out).squeeze(2), 
                                               dim=1)

    context = torch.sum(attn_weights.unsqueeze(2) * out, dim=1)
    out = self.fc(context)
    return out

def pad_sequences(batch): 
    max_len = max([len(seq) for seq in batch])
    return torch.stack([torch.cat([seq, torch.zeros(max_len-len(seq)).long()]) 
                        for seq in batch])

训练准备

criterion = nn.CrossEntropyLoss()

attention_model = RNNWithAttentionModel()
optimizer = torch.optim.Adam(attention_model.parameters(), lr=0.01)

for epoch in range(300):
    attention_model.train()
    optimizer.zero_grad()

    padded_inputs = pad_sequences(inputs)
    outputs = attention_model(padded_inputs)

    loss = criterion(outputs, targets)
    loss.backward()
    optimizer.step()

模型评估

for input_seq, target in zip(input_data, target_data):
    input_test = torch.tensor(input_seq, dtype=torch.long).unsqueeze(0)

    attention_model.eval()
    attention_output = attention_model(input_test)

    attention_prediction = ix_to_word[torch.argmax(attention_output).item()]

    print(f"\nInput: {' '.join([ix_to_word[ix] for ix in input_seq])}")
    print(f"Target: {ix_to_word[target]}")
    print(f"RNN with Attention prediction: {attention_prediction}")

Input: the cat sat on the
Target: mat
RNN with Attention prediction: mat

Passons à la pratique !

使用 PyTorch 的文本深度学习