上传文件至 /

config.py

@@ -1,40 +1,19 @@
-# -*- coding: utf-8 -*-
-"""
-配置文件 - 所有超参数集中管理
-
-设计思路：
-将超参数分门别类，学生可以单独修改某一类而不会影响其他
-"""
-
-# ==================== 数据相关 ====================
-DATA_DIR = 'data/ChnSentiCorp'          # 数据集路径
-MAX_FEATURES = 3000                     # 词表最大容量
-MAX_SEQ_LEN = 100                       # 句子最大长度（词数）
-VECTORIZER_TYPE = 'tfidf'               # 'tfidf' 或 'bow'（向量化方式）
-
-# ==================== 模型相关 ====================
-MODEL_TYPE = 'mlp'                       # 'mlp' 或 'lr'（模型类型）
-HIDDEN_SIZE = 64                        # MLP隐藏层大小（LR忽略）
-NUM_CLASSES = 2                         # 类别数（正面/负面二分类）
-KEEP_PROB = 1.0                         # Dropout保留概率（LR忽略，设为1即可）
-
-# ==================== 训练相关 ====================
-LEARNING_RATE = 0.05                    # 学习率
-NUM_EPOCHS = 100                        # 训练轮数
-BATCH_SIZE = 64                         # 批次大小
-
-# ==================== 类别权重（解决数据不平衡问题）====================
-USE_CLASS_WEIGHT = True                 # True=启用类别权重, False=不启用（对比用）
-# 权重计算公式: n_samples / (n_classes * n_class_i)
-# 正面评论多所以权重小，负面评论少所以权重大
-CLASS_WEIGHT_POS = 0.73                # 正面类权重（自动计算）
-CLASS_WEIGHT_NEG = 1.58                # 负面类权重（自动计算）
-
-# ==================== 实验相关 ====================
-RUN_COMPARISON = False                  # True=运行对比实验, False=运行单个模型
-COMPARE_MODELS = ['lr', 'mlp']          # 要对比的模型列表
-COMPARE_VECTORS = ['bow', 'tfidf']      # 要对比的向量化方式
-
-# ==================== 其他 ====================
-RANDOM_SEED = 42                        # 随机种子（保证可复现）
-VERBOSE = True                          # 打印详细日志
+# ģ<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+MODEL_TYPE = 'mlp'  # 'lr' <20>߼<EFBFBD><DFBC>ع<EFBFBD> / 'mlp' <20><><EFBFBD><EFBFBD>֪<EFBFBD><D6AA>
+VECTORIZER_TYPE = 'tfidf'  # 'bow' / 'tfidf'
+
+# ѵ<><D1B5><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+NUM_EPOCHS = 100
+LEARNING_RATE = 0.05
+BATCH_SIZE = 32
+HIDDEN_SIZE = 64
+
+# <20>ı<EFBFBD><C4B1><EFBFBD><EFBFBD><EFBFBD>
+MAX_FEATURES = 3000  # <20>ʱ<EFBFBD><CAB1><EFBFBD>С
+MAX_SEQ_LEN = 100    # <20><><EFBFBD><EFBFBD>ı<EFBFBD><C4B1><EFBFBD><EFBFBD><EFBFBD>
+
+# <EFBFBD><EFBFBD><EFBFBD>ݲ<EFBFBD>ƽ<EFBFBD>⴦<EFBFBD><EFBFBD>
+USE_CLASS_WEIGHT = True  # <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>Ȩ<EFBFBD><C8A8>
+
+# ʵ<><CAB5>Ա<EFBFBD>
+RUN_COMPARISON = True

dataset.py

@@ -1,286 +1,88 @@
-# -*- coding: utf-8 -*-
-"""
-数据加载与向量化模块
-
-支持两种向量化方法：
-1. BoW (Bag of Words) - 词频向量
-2. TF-IDF - 词频-逆文档频率向量
-
-TF-IDF 的优势：
-- 降低常见词（如"的"、"是"）的权重
-- 提升罕见词的信息量
-- 通常效果优于简单BoW
-"""
-
-import os
-import re
-import csv
-import math
-import jieba
-import numpy as np
-from collections import Counter
-
-try:
-    import urllib.request
-    import ssl
-    DOWNLOAD_AVAILABLE = True
-except ImportError:
-    DOWNLOAD_AVAILABLE = False
-
-
-DATASET_URL = "https://raw.githubusercontent.com/SophonPlus/ChineseNlpCorpus/master/datasets/ChnSentiCorp_htl_all/ChnSentiCorp_htl_all.csv"
-
-
-def download_dataset(data_dir):
-    """下载数据集（如果不存在）"""
-    csv_path = os.path.join(data_dir, 'ChnSentiCorp_htl_all.csv')
-    
-    if os.path.exists(csv_path):
-        print(f"数据已存在: {csv_path}")
-        return True
-    
-    if not DOWNLOAD_AVAILABLE:
-        return False
-    
-    print("正在下载数据集...")
-    ssl_context = ssl.create_default_context()
-    ssl_context.check_hostname = False
-    ssl_context.verify_mode = ssl.CERT_NONE
-    
-    try:
-        request = urllib.request.Request(DATASET_URL, headers={'User-Agent': 'Mozilla/5.0'})
-        response = urllib.request.urlopen(request, timeout=120, context=ssl_context)
-        os.makedirs(data_dir, exist_ok=True)
-        with open(csv_path, 'wb') as f:
-            f.write(response.read())
-        print(f"下载完成: {csv_path}")
-        return True
-    except Exception as e:
-        print(f"下载失败: {e}")
-        return False
-
-
-def load_raw_data(data_dir):
-    """加载原始数据"""
-    csv_path = os.path.join(data_dir, 'ChnSentiCorp_htl_all.csv')
-    texts, labels = [], []
-    
-    with open(csv_path, 'r', encoding='utf-8') as f:
-        reader = csv.reader(f)
-        for row in reader:
-            if len(row) < 2:
-                continue
-            try:
-                label = int(row[0])
-                review = row[1].strip()
-                if review:
-                    texts.append(review)
-                    labels.append(label)
-            except (ValueError, IndexError):
-                continue
-    
-    return texts, np.array(labels)
-
-
-def tokenize(text):
-    """中文分词"""
-    text = re.sub(r'[^\u4e00-\u9fa5a-zA-Z]', ' ', text)
-    words = jieba.lcut(text)
-    return [w for w in words if len(w) > 1]
-
-
-# ==================== 向量化器 ====================
-
-class BaseVectorizer:
-    """向量化器基类"""
-    def fit(self, texts): pass
-    def transform(self, texts): pass
-    def fit_transform(self, texts): pass
-
-
-class BoWVectorizer(BaseVectorizer):
-    """
-    词袋模型 (Bag of Words)
-    
-    原理：统计每个词在文本中出现的次数
-    向量维度 = 词表大小
-    每个维度 = 该词在本文本中出现的次数
-    """
-    
-    def __init__(self, max_features, max_seq_len):
-        self.max_features = max_features
-        self.max_seq_len = max_seq_len
-        self.vocab = {}
-        self.doc_freq = {}  # 文档频率
-        self.num_docs = 0
-    
-    def fit(self, texts):
-        """构建词表（基于词频）"""
-        counter = Counter()
-        doc_counter = Counter()  # 统计包含该词的文档数
-        
-        for text in texts:
-            words = tokenize(text)
-            unique_words = set(words)
-            counter.update(words)
-            for w in unique_words:
-                doc_counter[w] += 1
-        
-        self.num_docs = len(texts)
-        
-        # 取最高频的词
-        most_common = counter.most_common(self.max_features)
-        self.vocab = {word: idx for idx, (word, _) in enumerate(most_common)}
-        
-        # 记录文档频率（用于TF-IDF）
-        self.doc_freq = {w: doc_counter[w] for w in self.vocab}
-        
-        print(f"  BoW词表大小: {len(self.vocab)}")
-        return self
-    
-    def transform(self, texts):
-        """将文本转换为词频向量"""
-        vectors = []
-        for text in texts:
-            words = tokenize(text)
-            freq = [0] * self.max_seq_len
-            for i, word in enumerate(words[:self.max_seq_len]):
-                if word in self.vocab:
-                    freq[i] = 1  # 二值（出现=1，不出现=0）
-            vectors.append(freq)
-        return np.array(vectors, dtype=np.float32)
-    
-    def fit_transform(self, texts):
-        self.fit(texts)
-        return self.transform(texts)
-
-
-class TFIDFVectorizer(BaseVectorizer):
-    """
-    TF-IDF 向量器
-    
-    原理：
-    - TF(词频) = 词在本文本中出现的次数
-    - IDF(逆文档频率) = log(总文档数 / 包含该词的文档数)
-    - TF-IDF = TF × IDF
-    
-    优势：
-    - 降低常见无意义词的权重（如"的"、"是"）
-    - 提升罕见但有信息量的词
-    """
-    
-    def __init__(self, max_features, max_seq_len):
-        self.max_features = max_features
-        self.max_seq_len = max_seq_len
-        self.vocab = {}
-        self.idf = {}  # 存储每个词的IDF值
-        self.num_docs = 0
-    
-    def fit(self, texts):
-        """构建词表并计算IDF"""
-        counter = Counter()
-        doc_counter = Counter()
-        
-        for text in texts:
-            words = tokenize(text)
-            unique_words = set(words)
-            counter.update(words)
-            for w in unique_words:
-                doc_counter[w] += 1
-        
-        self.num_docs = len(texts)
-        
-        # 计算每个词的IDF
-        # IDF = log(总文档数 / 包含该词的文档数)
-        idf_values = {}
-        for word, df in doc_counter.items():
-            idf_values[word] = math.log(self.num_docs / (df + 1)) + 1  # 加1防零
-        
-        # 取IDF值最高的词（信息量最大的词）
-        sorted_words = sorted(idf_values.items(), key=lambda x: x[1], reverse=True)
-        self.vocab = {word: idx for idx, (word, _) in enumerate(sorted_words[:self.max_features])}
-        
-        # 保存IDF值
-        self.idf = {word: idf_values[word] for word in self.vocab}
-        
-        print(f"  TF-IDF词表大小: {len(self.vocab)}")
-        print(f"  平均IDF: {np.mean(list(self.idf.values())):.3f}")
-        return self
-    
-    def transform(self, texts):
-        """将文本转换为TF-IDF向量"""
-        vectors = []
-        for text in texts:
-            words = tokenize(text)
-            
-            # 计算TF
-            tf = Counter(words)
-            tf_sum = len(words) if words else 1
-            
-            # 生成向量
-            vec = [0.0] * self.max_seq_len
-            for i, word in enumerate(words[:self.max_seq_len]):
-                if word in self.vocab:
-                    # TF × IDF
-                    vec[i] = (tf[word] / tf_sum) * self.idf.get(word, 0)
-            vectors.append(vec)
-        
-        return np.array(vectors, dtype=np.float32)
-    
-    def fit_transform(self, texts):
-        self.fit(texts)
-        return self.transform(texts)
-
-
-def load_data(data_dir, max_features, max_seq_len, vectorizer_type='tfidf'):
-    """
-    加载并向量化数据
-    
-    参数：
-    - vectorizer_type: 'tfidf' 或 'bow'
-    """
-    if not download_dataset(data_dir):
-        raise RuntimeError("数据加载失败，请检查网络或手动下载数据集")
-    
-    print("正在加载数据...")
-    texts, labels = load_raw_data(data_dir)
-    print(f"总评论数: {len(texts)}, 正面: {sum(labels)}, 负面: {len(labels) - sum(labels)}")
-    
-    # 选择向量化器
-    if vectorizer_type == 'tfidf':
-        vectorizer = TFIDFVectorizer(max_features, max_seq_len)
-        vec_name = "TF-IDF"
-    else:
-        vectorizer = BoWVectorizer(max_features, max_seq_len)
-        vec_name = "BoW"
-    
-    print(f"正在使用{vec_name}向量化...")
-    X = vectorizer.fit_transform(texts)
-    y = labels
-    
-    # 打乱并划分
-    np.random.seed(42)
-    indices = np.random.permutation(len(X))
-    X = X[indices]
-    y = y[indices]
-    
-    split_idx = int(len(X) * 0.8)
-    X_train, X_test = X[:split_idx], X[split_idx:]
-    y_train, y_test = y[:split_idx], y[split_idx:]
-    
-    print(f"训练集: {len(X_train)}条, 测试集: {len(X_test)}条")
-    
-    return X_train, y_train, X_test, y_test, vectorizer
-
-
-if __name__ == '__main__':
-    # 测试
-    print("=" * 60)
-    print("测试 TF-IDF 向量化")
-    print("=" * 60)
-    X_train, y_train, X_test, y_test, vec = load_data(
-        'data/ChnSentiCorp', max_features=3000, max_seq_len=100, 
-        vectorizer_type='tfidf'
-    )
-    print(f"\nX_train shape: {X_train.shape}")
-    print(f"X_train sample (前5个特征): {X_train[0][:5]}")
+import numpy as np
+import jieba
+from collections import Counter
+import os
+import requests
+
+# <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ݼ<EFBFBD>
+def download_data():
+    url = "https://github.com/SophonPlus/ChineseNlpCorpus/raw/master/datasets/ChnSentiCorp_htl_all/ChnSentiCorp_htl_all.csv"
+    path = "ChnSentiCorp_htl_all.csv"
+    if not os.path.exists(path):
+        print("<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>ݼ<EFBFBD>...")
+        r = requests.get(url)
+        with open(path, "wb") as f:
+            f.write(r.content)
+    return path
+
+# <20>ִʣ<D6B4><CAA3><EFBFBD><EFBFBD>˵<EFBFBD><CBB5>֣<EFBFBD>
+def tokenize(text):
+    words = jieba.lcut(str(text).strip())
+    return [w for w in words if len(w) > 1]
+
+# <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+def load_data():
+    path = download_data()
+    texts, labels = [], []
+    with open(path, "r", encoding="utf-8") as f:
+        next(f)
+        for line in f:
+            parts = line.strip().split(",", 1)
+            if len(parts) != 2:
+                continue
+            label, text = parts
+            texts.append(text)
+            labels.append(int(label))
+    return texts, labels
+
+# BoW <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+class BoWVectorizer:
+    def __init__(self, max_features=3000):
+        self.max_features = max_features
+        self.vocab = {}
+
+    def fit(self, texts):
+        counter = Counter()
+        for t in texts:
+            counter.update(tokenize(t))
+        words = [w for w, _ in counter.most_common(self.max_features)]
+        self.vocab = {w:i for i, w in enumerate(words)}
+
+    def transform(self, text):
+        words = tokenize(text)
+        vec = np.zeros(self.max_features)
+        for w in words:
+            if w in self.vocab:
+                vec[self.vocab[w]] += 1
+        return vec
+
+# TF-IDF <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+class TFIDFVectorizer:
+    def __init__(self, max_features=3000):
+        self.max_features = max_features
+        self.vocab = {}
+        self.idf = {}
+
+    def fit(self, texts):
+        counter = Counter()
+        doc_freq = Counter()
+        for t in texts:
+            ws = set(tokenize(t))
+            counter.update(tokenize(t))
+            for w in ws:
+                doc_freq[w] += 1
+
+        words = [w for w, _ in counter.most_common(self.max_features)]
+        self.vocab = {w:i for i, w in enumerate(words)}
+        N = len(texts)
+        for w in self.vocab:
+            self.idf[w] = np.log(N / (doc_freq.get(w, 0) + 1))
+
+    def transform(self, text):
+        words = tokenize(text)
+        vec = np.zeros(self.max_features)
+        tf = Counter(words)
+        for w, cnt in tf.items():
+            if w in self.vocab:
+                vec[self.vocab[w]] = cnt * self.idf[w]
+        return vec

main.py

@@ -1,34 +1,60 @@
-# -*- coding: utf-8 -*-
-"""
-主程序入口
-
-使用方式：
-
-1. 运行单个模型（默认）：
-   python main.py
-   
-   修改 config.py 中的 MODEL_TYPE 和 VECTORIZER_TYPE 来切换配置
-
-2. 运行对比实验：
-   修改 config.py 中 RUN_COMPARISON = True
-   
-   这会依次运行：
-   - 实验1: BoW vs TF-IDF (固定LR模型)
-   - 实验2: LR vs MLP (固定TF-IDF)
-   - 实验3: 不同学习率对比
-   - 实验4: 不同隐藏层大小对比
-   
-   最后输出汇总报告
-"""
-
-from train import main
-
-if __name__ == '__main__':
-    print("\n" + "=" * 70)
-    print("文本分类实验 - 纯NumPy实现")
-    print("数据集: ChnSentiCorp (中文酒店评论)")
-    print("模型: Logistic Regression / MLP")
-    print("向量化: BoW / TF-IDF")
-    print("=" * 70 + "\n")
-    
-    main()
+import numpy as np
+from dataset import load_data, BoWVectorizer, TFIDFVectorizer
+from train import train
+import config as cfg
+import pickle
+import time
+
+# <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+texts, labels = load_data()
+labels = np.array(labels)
+
+# <20><><EFBFBD><EFBFBD>ѵ<EFBFBD><D1B5><EFBFBD><EFBFBD>/<2F><><EFBFBD>Լ<EFBFBD>
+np.random.seed(42)
+indices = np.random.permutation(len(texts))
+split = int(0.8 * len(texts))
+train_idx, test_idx = indices[:split], indices[split:]
+train_texts = [texts[i] for i in train_idx]
+test_texts = [texts[i] for i in test_idx]
+y_train, y_test = labels[train_idx], labels[test_idx]
+
+# <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
+if cfg.VECTORIZER_TYPE == "bow":
+    vec = BoWVectorizer(cfg.MAX_FEATURES)
+else:
+    vec = TFIDFVectorizer(cfg.MAX_FEATURES)
+
+vec.fit(train_texts)
+X_train = np.array([vec.transform(t) for t in train_texts])
+X_test = np.array([vec.transform(t) for t in test_texts])
+
+# ѵ<><D1B5>
+print("="*50)
+print(f"ѵ<EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>:\n  ģ<><C4A3>: {cfg.MODEL_TYPE}\n  <20><><EFBFBD><EFBFBD>: {cfg.VECTORIZER_TYPE}\n  ѧϰ<D1A7><CFB0>: {cfg.LEARNING_RATE}")
+print("="*50)
+
+model, t = train(
+    X_train, y_train, X_test, y_test,
+    model_type=cfg.MODEL_TYPE,
+    lr=cfg.LEARNING_RATE,
+    epochs=cfg.NUM_EPOCHS,
+    use_weight=cfg.USE_CLASS_WEIGHT
+)
+
+# <20><><EFBFBD><EFBFBD>
+ts = time.strftime("%m%d_%H%M%S")
+name = f"model_{cfg.MODEL_TYPE}_{cfg.VECTORIZER_TYPE}_{'weighted' if cfg.USE_CLASS_WEIGHT else 'raw'}_{ts}"
+
+if cfg.MODEL_TYPE == "lr":
+    np.save(f"{name}_W.npy", model.W)
+    np.save(f"{name}_b.npy", model.b)
+else:
+    np.save(f"{name}_W1.npy", model.W1)
+    np.save(f"{name}_b1.npy", model.b1)
+    np.save(f"{name}_W2.npy", model.W2)
+    np.save(f"{name}_b2.npy", model.b2)
+
+with open(f"{name}_vec.pkl", "wb") as f:
+    pickle.dump(vec, f)
+
+print(f"\nģ<EFBFBD><EFBFBD><EFBFBD>ѱ<EFBFBD><EFBFBD><EFBFBD>: {name}_*.npy/*.pkl")