文字分類的遷移學習

Deep Learning for Text with PyTorch

Shubham Jain

Instructor

什麼是遷移學習？

遷移學習

將一個任務的既有知識用於相關任務

省時
共享專長
減少大型資料需求

例如：英文老師改教歷史

遷移學習的運作機制

遷移學習 I

遷移學習的運作機制

遷移學習 II

遷移學習的運作機制

遷移學習 II

遷移學習的運作機制

遷移學習 III

預訓練模型：BERT

Bidirectional Encoder Representations from Transformers

BERT 情緒分析

以語言建模訓練
多層 transformers 結構
以大量文本預先訓練

動手做：實作 BERT

texts = ["I love this!", 
         "This is terrible.", 
         "Amazing experience!", 
         "Not my cup of tea."]
labels = [1, 0, 1, 0]

import torch
from transformers import BertTokenizer, BertForSequenceClassification

tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
model = BertForSequenceClassification.from_pretrained('bert-base-uncased',                 
                                                       num_labels=2)

inputs = tokenizer(texts, padding=True, truncation=True, 
                    return_tensors="pt", max_length=32)
inputs["labels"] = torch.tensor(labels)

微調 BERT

optimizer = torch.optim.AdamW(model.parameters(), lr=0.00001)
model.train()

for epoch in range(1):
    outputs = model(**inputs)

    loss = outputs.loss
    loss.backward()

    optimizer.step()
    optimizer.zero_grad()

    print(f"Epoch: {epoch+1}, Loss: {loss.item()}")

Epoch: 1, Loss: 0.7061821222305298

在新文本上評估

text = "I had an awesome day!"
input_eval = tokenizer(text, return_tensors="pt", truncation=True, 
                       padding=True, max_length=128)

outputs_eval = model(**input_eval)

predictions = torch.nn.functional.softmax(outputs_eval.logits, dim=-1)

predicted_label = 'positive' if torch.argmax(predictions) > 0 else 'negative'
print(f"Text: {text}\nSentiment: {predicted_label}")

Text: I had an awesome day!
Sentiment: positive

一起來練習吧！

Deep Learning for Text with PyTorch