用於文字處理的 Transformers

Deep Learning for Text with PyTorch

Shubham Jain

Instructor

為什麼用 transformers 做文字處理？

Transformers 標誌

速度快
跨距理解詞與詞關係
近似人類的回應

Transformer 的組成元件

Encoder：處理輸入資料

Decoder：重建輸出

前饋式神經網路：精煉理解

位置編碼：確保順序有效

多頭注意力：擷取多重訊息或情緒

準備資料：訓練／測試切分

sentences = ["I love this product", "This is terrible", 
             "Could be better", "This is the best"]
labels = [1, 0, 0, 1]

train_sentences = sentences[:3]
train_labels = labels[:3]
test_sentences = sentences[3:]
test_labels = labels[3:]

建立 transformer 模型

class TransformerEncoder(nn.Module):

    def __init__(self, embed_size, heads, num_layers, dropout):
        super(TransformerEncoder, self).__init__()

        self.encoder = nn.TransformerEncoder(
         nn.TransformerEncoderLayer(d_model=embed_size, nhead=heads),
         num_layers=num_layers)

        self.fc = nn.Linear(embed_size, 2)

    def forward(self, x):

        x = self.encoder(x)

        x = x.mean(dim=1)

        return self.fc(x)

model = TransformerEncoder(embed_size=512, heads=8, num_layers=3, dropout=0.5)

optimizer = optim.Adam(model.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()

訓練 transformers

for epoch in range(5):

 for sentence, label in zip(train_sentences, train_labels):
        tokens = sentence.split()

        data = torch.stack([token_embeddings[token] for token in tokens], dim=1)

        output = model(data)

        loss = criterion(output, torch.tensor([label]))

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        print(f"Epoch {epoch}, Loss: {loss.item()}")

Epoch 0, Loss: 13.788233757019043
Epoch 1, Loss: 3.9480819702148438
Epoch 2, Loss: 2.4790847301483154
Epoch 3, Loss: 1.3020926713943481
Epoch 4, Loss: 0.4660853147506714

使用 transformers 進行預測

def predict(sentence):
    model.eval()

    with torch.no_grad():

        tokens = sentence.split()
        data = torch.stack([token_embeddings.get(token, torch.rand((1, 512))) 
                            for token in tokens], dim=1)

        output = model(data)

        predicted = torch.argmax(output, dim=1)

        return "Positive" if predicted.item() == 1 else "Negative"

對新文字進行預測

sample_sentence = "This product can be better"
print(f"'{sample_sentence}' is {predict(sample_sentence)}")

'This product can be better' is Negative

一起來練習吧！

Deep Learning for Text with PyTorch