# ======================================================================================
# 1.·基础 RNN 模型构建,
# 题目:使用PyTorch 构建一个简单的RNN模型，用于序列分类任务。
# 要求:
# ·实现基本的 RNN 单元
# ·处理序列数据(如时间序列或文本)
# ·添加全连接层进行分类.
# ·实现训练过程
# 2.序列数据预处理.
# 题目:实现序列数据的预处理流程，包括:
# ·序列填充(padding)
# ·序列截断(truncation)
# ·数据标准化。
# ·创建数据加载器.
# 3. 使用简单RNN和LSTM（长短时记忆网络）模型，在一个“情感分类”任务上（如IMDb电影评论）进行训练。
# 题目：
# 1. 记录两个模型在测试集上的准确率。
# 2. 从测试集中找出一条非常长的、情感复杂的评论（例如，前半段好评，后半段差评）。
# 3. 报告两个模型对这条特定评论的预测结果。
# 4. 尝试解释为什么LSTM在处理这种长距离依赖的评论时，通常表现得比简单RNN更好。
# ======================================================================================
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
import numpy as np
from collections import Counter
import re
import torch.nn.functional as F
import matplotlib.pyplot as plt
import random
# 设置随机种子以保证结果可重现
torch.manual_seed(42)
np.random.seed(42)
random.seed(42)
# ============================================================================
# 1. 基础RNN模型构建
# ============================================================================
class SimpleRNN(nn.Module):
def __init__(self, vocab_size, embedding_dim, hidden_dim, output_dim, n_layers=1, dropout=0.2):
s
uper(SimpleRNN, self).__init__()
self.embedding = nn.Embedding(vocab_size, embedding_dim)
self.rnn = nn.RNN(embedding_dim, hidden_dim, n_layers,
batch_first=True, dropout=dropout)
self.fc = nn.Linear(hidden_dim, output_dim)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
embedded = self.dropout(self.embedding(x))
output, hidden = self.rnn(embedded)
last_output = output[:, -1, :]
return self.fc(last_output)
class LSTMClassifier(nn.Module):
def __init__(self, vocab_size, embedding_dim, hidden_dim, output_dim, n_layers=1, dropout=0.2):
s
uper(LSTMClassifier, self).__init__()
self.embedding = nn.Embedding(vocab_size, embedding_dim)
self.lstm = nn.LSTM(embedding_dim, hidden_dim, n_layers,
batch_first=True, dropout=dropout)
self.fc = nn.Linear(hidden_dim, output_dim)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
embedded = self.dropout(self.embedding(x))
output, (hidden, cell) = self.lstm(embedded)
last_output = output[:, -1, :]
return self.fc(last_output)
# ============================================================================
# 2. 序列数据预处理
# ============================================================================
class TextProcessor:
def __init__(self, max_vocab_size=10000, max_length=200):
self.max_vocab_size = max_vocab_size
self.max_length = max_length
self.vocab = {}
self.word2idx = {}
self.idx2word = {}
def preprocess_text(self, text):
text = text.lower()
text = re.sub(r'[^a-zA-Z\s]', '', text)
tokens = text.split()
return tokens
def build_vocab(self, texts):
counter = Counter()
for text in texts:
tokens = self.preprocess_text(text)
counter.
update(tokens)
common_words = counter.most_common(self.max_vocab_size - 2)
self.word2idx = {'<pad>': 0, '<unk>': 1}
self.idx2word = {0: '<pad>', 1: '<unk>'}
for idx, (word, _) in enumerate(common_words, start=2):
self.word2idx[word] = idx
self.idx2word[idx] = word
self.vocab_size = len(self.word2idx)
def text_to_sequence(self, text):
tokens = self.preprocess_text(text)
sequence = [self.word2idx.get(token, 1) for token in tokens]
return sequence
def pad_sequence(self, sequence):
if len(sequence) > self.max_length:
return sequence[:self.max_length]
else:
return sequence + [0] * (self.max_length - len(sequence))
def process_dataset(self, texts, labels):
sequences = []
for text in texts:
seq = self.text_to_sequence(text)
seq = self.pad_sequence(seq)
sequences.append(seq)
return torch.tensor(sequences), torch.tensor(labels)
def create_data_loader(texts, labels, batch_size=32, shuffle=True):
processor = TextProcessor()
processor.build_vocab(texts)
sequences, labels_tensor = processor.process_dataset(texts, labels)
dataset = TensorDataset(sequences, labels_tensor)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=shuffle)
return dataloader, processor
# ============================================================================
# 3. 训练和评估函数 (不使用sklearn)
# ============================================================================
def calculate_accuracy(predictions, targets):
"""手动计算准确率，替代sklearn的accuracy_score"""
correct = (predictions == targets).sum().item()
total = len(targets)
return correct / total
def train_model(model, train_loader, val_loader, num_epochs=10, learning_rate=0.001):
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
train_losses = []
val_accuracies = []
for epoch in range(num_epochs):
model.train()
total_loss = 0
for batch_idx, (data, target) in enumerate(train_loader):
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
total_loss += loss.item()
avg_loss = total_loss / len(train_loader)
train_losses.append(avg_loss)
# 验证
model.eval()
val_predictions = []
val_targets = []
with torch.no_grad():
for data, target in val_loader:
output = model(data)
pred = output.argmax(dim=1)
val_predictions.append(pred)
val_targets.append(target)
# 合并所有批次的预测结果
val_predictions = torch.cat(val_predictions)
val_targets = torch.cat(val_targets)
val_accuracy = calculate_accuracy(val_predictions, val_targets)
val_accuracies.append(val_accuracy)
print(f'Epoch {epoch + 1}/{num_epochs}, Loss: {avg_loss:.4f}, Val Accuracy: {val_accuracy:.4f}')
return train_losses, val_accuracies
def evaluate_model(model, test_loader):
model.eval()
predictions = []
targets = []
with torch.no_grad():
for data, target in test_loader:
output = model(data)
pred = output.argmax(dim=1)
predictions.append(pred)
targets.append(target)
# 合并所有批次的预测结果
predictions = torch.cat(predictions)
targets = torch.cat(targets)
accuracy = calculate_accuracy(predictions, targets)
return accuracy, predictions.numpy(), targets.numpy()
def predict_single_text(model, processor, text):
model.eval()
sequence = processor.text_to_sequence(text)
padded_sequence = processor.pad_sequence(sequence)
input_tensor = torch.tensor([padded_sequence])
with torch.no_grad():
output = model(input_tensor)
probabilities = F.softmax(output, dim=1)
prediction = output.argmax(dim=1).item()
confidence = probabilities[0][prediction].item()
sentiment = "Positive" if prediction == 1 else "Negative"
return sentiment, confidence
# ============================================================================
# 4. 数据生成和主程序
# ============================================================================
def create_sample_imdb_data(num_samples=1000):
positive_texts = [
"This movie was absolutely fantastic! The acting was s
uperb and the plot was engaging from start to finish.",
"I loved
every minute of this film. The cinematography was beautiful and the characters were well-developed.",
"An outstanding performance by all actors. The story was heartwarming and inspiring.",
"One of the
best movies I've seen this year. Highly recommended for all movie lovers.",
"The director did an amazing job with this film. The visuals were stunning and the music was perfect.",
"A masterpiece of modern cinema that will be remembered for years to come.",
"Brilliant storytelling combined with exceptional acting makes this a must-see movie.",
"I was completely captivated from beginning to end. What an incredible film!",
"The character development was phenomenal and the plot twists were unexpected yet satisfying.",
"This film exceeded all my expectations. Truly a work of art in
every aspect."
]
negative_texts = [
"This was a terrible movie. Poor acting and a boring plot made it unwatchable.",
"I was very disappointed with this film. The story made no sense and the characters were flat.",
"Waste of time and money. The movie was poorly directed and the script was awful.",
"One of the worst films I've
ever seen. I can't believe I sat through the whole thing.",
"The acting was wooden and the dialogue was cringe-worthy. Avoid this movie at all costs.",
"A complete disaster from start to finish. I want my two hours back.",
"The plot was predictable and the characters were one-dimensional and uninteresting.",
"Poorly executed with terrible special effects and unconvincing performances.",
"I struggled to stay awake during this boring and poorly written film.",
"An embarrassing attempt at filmmaking that fails on
every level."
]
texts = []
labels = []
for _ in range(num_samples // 2):
base_text = np.random.choice(positive_texts)
variations = ["Really ", "Absolutely ", "Truly ", "Honestly ", "Without a doubt "]
variation = np.random.choice(variations)
text = variation + base_text.lower()
texts.append(text)
labels.append(1)
base_text = np.random.choice(negative_texts)
variation = np.random.choice(variations)
text = variation + base_text.lower()
texts.append(text)
labels.append(0)
return texts, labels
def create_complex_review():
return """
I must admit that I was initially blown away by this movie. The opening scenes were absolutely breathtaking,
with stunning cinematography that captured the beauty of
every frame. The lead actor delivered a powerful performance
in the first half, bringing genuine emotion and depth to their character. The storyline started strong,
with an intriguing premise that promised an unforgettable cinematic experience. The musical score was
perfectly matched to the tone of the film, enhancing
every emotional beat. How
ever, as the movie progressed
into its second half, I found myself growing increasingly disappointed. The plot began to unravel,
introducing unnecessary subplots that added nothing to the main narrative. The character development
that seemed so promising early on was completely abandoned, leaving the protagonists feeling hollow and
underdeveloped. By the final act, the film had devolved into a mess of clichés and predictable twists
that undermined
everything that had been built
up earlier. The ending felt rushed and unsatisfying,
as if the writers had simply run out of
ideas. What started as a potential masterpiece ended
up being
just another forgettable Hollywood production that squandered its early promise.
"""
def main():
print("=" * 80)
print("IMDb情感分类任务: RNN vs LSTM 比较")
print("=" * 80)
# 创建模拟数据
print("\n1. 创建模拟IMDb数据集...")
texts, labels = create_sample_imdb_data(2000)
print(f"数据集大小: {len(texts)} 条评论")
print(f"正面评论: {sum(labels)}, 负面评论: {len(labels) - sum(labels)}")
# 分割数据集
split_idx = int(0.8 * len(texts))
train_texts, train_labels = texts[:split_idx], labels[:split_idx]
test_texts, test_labels = texts[split_idx:], labels[split_idx:]
# 创建数据加载器
print("\n2. 创建数据加载器...")
train_loader, processor = create_data_loader(train_texts, train_labels, batch_size=32)
test_loader, _ = create_data_loader(test_texts, test_labels, batch_size=32, shuffle=False)
# 初始化模型
vocab_size = processor.vocab_size
embedding_dim = 100
hidden_dim = 128
output_dim = 2
print(f"\n3. 初始化模型 (词汇表大小: {vocab_size})...")
rnn_model = SimpleRNN(vocab_size, embedding_dim, hidden_dim, output_dim)
lstm_model = LSTMClassifier(vocab_size, embedding_dim, hidden_dim, output_dim)
print(f"RNN模型参数数量: {sum(p.numel() for p in rnn_model.parameters()):,}")
print(f"LSTM模型参数数量: {sum(p.numel() for p in lstm_model.parameters()):,}")
# 训练RNN模型
print("\n4. 训练RNN模型...")
rnn_train_loss, rnn_val_acc = train_model(rnn_model, train_loader, test_loader, num_epochs=10)
# 训练LSTM模型
print("\n5. 训练LSTM模型...")
lstm_train_loss, lstm_val_acc = train_model(lstm_model, train_loader, test_loader, num_epochs=10)
# 评估模型
print("\n6. 在测试集上评估模型...")
rnn_accuracy, rnn_preds, rnn_targets = evaluate_model(rnn_model, test_loader)
lstm_accuracy, lstm_preds, lstm_targets = evaluate_model(lstm_model, test_loader)
print(f"RNN测试集准确率: {rnn_accuracy:.4f}")
print(f"LSTM测试集准确率: {lstm_accuracy:.4f}")
# 测试复杂评论
complex_review = create_complex_review()
print(f"\n7. 测试复杂评论 (前半段好评，后半段差评):")
print("=" * 60)
print(complex_review)
print("=" * 60)
rnn_sentiment, rnn_confidence = predict_single_text(rnn_model, processor, complex_review)
lstm_sentiment, lstm_confidence = predict_single_text(lstm_model, processor, complex_review)
print(f"\n预测结果:")
print(f"RNN预测: {rnn_sentiment} (置信度: {rnn_confidence:.4f})")
print(f"LSTM预测: {lstm_sentiment} (置信度: {lstm_confidence:.4f})")
# 解释LSTM优势
print("\n" + "=" * 80)
print("LSTM在处理长距离依赖时表现更好的原因:")
print("=" * 80)
reasons = [
"1. 门控机制: LSTM有输入门、遗忘门、输出门，可以精确控制信息的流动",
"2. 长期记忆: LSTM的细胞状态可以保持长期信息，避免梯度消失问题",
"3. 选择性记忆: LSTM可以选择记住重要信息，忘记不重要信息",
"4. 上下文理解: 对于情感复杂的评论，LSTM能更好地平衡前后文信息",
"5. 梯度流: 更好的梯度传播机制，缓解了简单RNN的梯度消失问题",
"6. 序列建模: 能够更好地建模长序列中的依赖关系",
"7. 对于这种前半段好评、后半段差评的复杂文本，LSTM能综合考虑整个序列",
"8. 简单RNN容易受最近信息的影响，而LSTM能考虑整个序列的上下文"
]
for reason in reasons:
print(reason)
# 绘制训练曲线
print("\n8. 生成训练曲线图...")
plt.figure(figsize=(12, 5))
plt.subplot(1, 2, 1)
plt.plot(rnn_train_loss, 'b-', label='RNN Loss', linewidth=2)
plt.plot(lstm_train_loss, 'r-', label='LSTM Loss', linewidth=2)
plt.title('
Training Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.grid(True, alpha=0.3)
plt.subplot(1, 2, 2)
plt.plot(rnn_val_acc, 'b-', label='RNN Accuracy', linewidth=2)
plt.plot(lstm_val_acc, 'r-', label='LSTM Accuracy', linewidth=2)
plt.title('Validation Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig('
training_curves.png', dpi=300, bbox_inches='tight')
plt.show()
print("\n训练完成！图表已保存为 '
training_curves.png'")
if __name__ == "__main__":
main()
将数据集换成IMDb数据集
团队设立了一个C++培训研讨会，每个人轮流主持不同的主题。资深成员Jackson将整合大家的想法并分享他对每个主题的看法。Jackson拥有10年的经验，在C++和软件设计方面非常专业。
扫一扫
135
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
