PyTorch 実例 - テキスト感情分析プロジェクト

import torch
import torch.nn as nn
import torch.optim as optim
from torchtext.data import Field, TabularDataset, BucketIterator
import spacy
import numpy as np

# フィールド処理を定義
TEXT = Field(tokenize='spacy',
            tokenizer_language='en_core_web_sm',
            include_lengths=True)
LABEL = Field(sequential=False, use_vocab=False)

# データセットを読み込み
train_data, test_data = TabularDataset.splits(
    path='./data',
    train='train.csv',
    test='test.csv',
    format='csv',
    fields=[('text', TEXT), ('label', LABEL)]
)

# ボキャブラリを構築
TEXT.build_vocab(train_data,
                max_size=25000,
                vectors="glove.6B.100d")

class SentimentLSTM(nn.Module):
    def __init__(self, vocab_size, embedding_dim, hidden_dim, output_dim, n_layers):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        self.lstm = nn.LSTM(embedding_dim,
                           hidden_dim,
                           num_layers=n_layers,
                           bidirectional=True)
        self.fc = nn.Linear(hidden_dim * 2, output_dim)
        self.dropout = nn.Dropout(0.5)
       
    def forward(self, text, text_lengths):
        embedded = self.dropout(self.embedding(text))
        packed_embedded = nn.utils.rnn.pack_padded_sequence(
            embedded, text_lengths.to('cpu'))
        packed_output, (hidden, cell) = self.lstm(packed_embedded)
        hidden = self.dropout(torch.cat((hidden[-2,:,:], hidden[-1,:,:]), dim=1))
        return self.fc(hidden)

# モデルパラメータ
INPUT_DIM = len(TEXT.vocab)
EMBEDDING_DIM = 100
HIDDEN_DIM = 256
OUTPUT_DIM = 1
N_LAYERS = 2

# モデルの初期化
model = SentimentLSTM(INPUT_DIM, EMBEDDING_DIM, HIDDEN_DIM, OUTPUT_DIM, N_LAYERS)

# オプティマイザと損失関数
optimizer = optim.Adam(model.parameters())
criterion = nn.BCEWithLogitsLoss()

def train(model, iterator, optimizer, criterion):
    epoch_loss = 0
    epoch_acc = 0
   
    model.train()
   
    for batch in iterator:
        text, text_lengths = batch.text
        predictions = model(text, text_lengths).squeeze(1)
        loss = criterion(predictions, batch.label)
       
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
       
        epoch_loss += loss.item()
        epoch_acc += accuracy(predictions, batch.label)
       
    return epoch_loss / len(iterator), epoch_acc / len(iterator)

def evaluate(model, iterator, criterion):
    epoch_loss = 0
    epoch_acc = 0
   
    model.eval()
   
    with torch.no_grad():
        for batch in iterator:
            text, text_lengths = batch.text
            predictions = model(text, text_lengths).squeeze(1)
            loss = criterion(predictions, batch.label)
            epoch_loss += loss.item()
            epoch_acc += accuracy(predictions, batch.label)
           
    return epoch_loss / len(iterator), epoch_acc / len(iterator)

def accuracy(preds, y):
    rounded_preds = torch.round(torch.sigmoid(preds))
    correct = (rounded_preds == y).float()
    acc = correct.sum() / len(correct)
    return acc

def predict_sentiment(model, sentence):
    tokenized = [tok.text for tok in nlp.tokenizer(sentence)]
    indexed = [TEXT.vocab.stoi[t] for t in tokenized]
    length = [len(indexed)]
    tensor = torch.LongTensor(indexed).to(device)
    tensor = tensor.unsqueeze(1)
    length_tensor = torch.LongTensor(length)
    prediction = torch.sigmoid(model(tensor, length_tensor))
    return prediction.item()

positive_review = "This movie was fantastic! I really enjoyed it."
negative_review = "The film was terrible and boring."

print(f"Positive review score: {predict_sentiment(model, positive_review):.4f}")
print(f"Negative review score: {predict_sentiment(model, negative_review):.4f}")