From 8934da87ba36fb463b17946441037de533c125f3 Mon Sep 17 00:00:00 2001 From: gitea_eternal <401029566@qq.com> Date: Mon, 27 Apr 2026 21:45:48 +0800 Subject: [PATCH] Upload train.py --- train.py | 265 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 265 insertions(+) create mode 100644 train.py diff --git a/train.py b/train.py new file mode 100644 index 0000000..7fc9402 --- /dev/null +++ b/train.py @@ -0,0 +1,265 @@ +# -*- coding: utf-8 -*- +""" +训练与对比实验模块 + +功能： +1. 单模型训练与评估 +2. 模型对比实验（LR vs MLP） +3. 向量化对比实验（BoW vs TF-IDF） +4. 超参数对比实验 +""" + +import numpy as np +import time +from datetime import datetime +from model_numpy import create_model +from dataset import load_data +from config import * + + +def train_single_model(X_train, y_train, X_test, y_test, + model_type, vectorizer_type, epochs, batch_size, lr, hidden_size): + """ + 训练单个模型并返回结果 + """ + print(f"\n{'='*60}") + print(f"训练配置:") + print(f" 模型: {model_type.upper()}") + print(f" 向量: {vectorizer_type.upper()}") + print(f" 学习率: {lr}") + if model_type == 'mlp': + print(f" 隐藏层大小: {hidden_size}") + print(f" 训练轮数: {epochs}") + print(f"{'='*60}") + + # 重新加载数据（确保向量化方式正确） + X_tr, y_tr, X_te, y_te, _ = load_data( + DATA_DIR, MAX_FEATURES, MAX_SEQ_LEN, vectorizer_type + ) + + # 计算类别权重（解决数据不平衡问题） + pos_count = sum(y_tr) + neg_count = len(y_tr) - pos_count + n_samples = len(y_tr) + n_classes = 2 + + if USE_CLASS_WEIGHT: + # 权重 = n_samples / (n_classes * n_class_i) + # 负面样本少，权重更大 + weight_pos = n_samples / (n_classes * pos_count) + weight_neg = n_samples / (n_classes * neg_count) + class_weight = {0: weight_neg, 1: weight_pos} + print(f" 类别权重: 正面={weight_pos:.2f}, 负面={weight_neg:.2f}") + else: + class_weight = None + print(f" 类别权重: 不使用") + + # 创建模型 + model = create_model( + model_type=model_type, + input_size=X_tr.shape[1], + hidden_size=hidden_size, + num_classes=NUM_CLASSES, + learning_rate=lr, + keep_prob=KEEP_PROB, + class_weight=class_weight + ) + + # 训练 + start_time = time.time() + model.fit(X_tr, y_tr, X_te, y_te, epochs=epochs, batch_size=batch_size, verbose=True) + train_time = time.time() - start_time + + # 最终评估 + train_acc = model.accuracy(X_tr, y_tr) + test_acc = model.accuracy(X_te, y_te) + + print(f"\n最终结果:") + print(f" 训练准确率: {train_acc:.4f}") + print(f" 测试准确率: {test_acc:.4f}") + print(f" 训练时间: {train_time:.2f}秒") + + # 保存模型权重 + # 生成模型文件名（包含配置标识和时间戳） + weight_suffix = "weighted" if USE_CLASS_WEIGHT else "noweight" + timestamp = datetime.now().strftime("%m%d_%H%M%S") + model_path = f"model_{model_type}_{vectorizer_type}_{weight_suffix}_{timestamp}" + model.save(model_path) + print(f" 权重文件: {model_path}_*.npy") + + return { + 'model_type': model_type, + 'vectorizer_type': vectorizer_type, + 'train_acc': train_acc, + 'test_acc': test_acc, + 'train_time': train_time, + 'lr': lr, + 'hidden_size': hidden_size + } + + +def run_comparison_experiments(): + """ + 运行对比实验 + + 实验设计： + 1. 不同向量化方法的效果对比 (BoW vs TF-IDF) + 2. 不同模型的效果对比 (LR vs MLP) + 3. 超参数对比 (学习率, 隐藏层大小) + """ + print("\n" + "=" * 70) + print("文本分类对比实验") + print("=" * 70) + + results = [] + + # ===== 实验1: 向量化方法对比 (固定模型=LR) ===== + print("\n\n" + "▶ " * 20) + print("实验1: 向量化方法对比 (模型=LR, 学习率=0.05)") + print("▶ " * 20) + + for vec_type in COMPARE_VECTORS: + result = train_single_model( + None, None, None, None, + model_type='lr', + vectorizer_type=vec_type, + epochs=50, + batch_size=BATCH_SIZE, + lr=0.05, + hidden_size=HIDDEN_SIZE + ) + results.append(result) + + # ===== 实验2: 模型复杂度对比 (固定向量=TF-IDF) ===== + print("\n\n" + "▶ " * 20) + print("实验2: 模型复杂度对比 (向量=TF-IDF)") + print("▶ " * 20) + + for model_type in COMPARE_MODELS: + result = train_single_model( + None, None, None, None, + model_type=model_type, + vectorizer_type='tfidf', + epochs=50, + batch_size=BATCH_SIZE, + lr=LEARNING_RATE, + hidden_size=HIDDEN_SIZE + ) + results.append(result) + + # ===== 实验3: 学习率对比 (MLP) ===== + print("\n\n" + "▶ " * 20) + print("实验3: 学习率对比 (模型=MLP, 向量=TF-IDF)") + print("▶ " * 20) + + for lr in [0.01, 0.1]: + result = train_single_model( + None, None, None, None, + model_type='mlp', + vectorizer_type='tfidf', + epochs=50, + batch_size=BATCH_SIZE, + lr=lr, + hidden_size=64 + ) + results.append(result) + + # ===== 实验4: 隐藏层大小对比 (MLP) ===== + print("\n\n" + "▶ " * 20) + print("实验4: 隐藏层大小对比 (模型=MLP, 向量=TF-IDF)") + print("▶ " * 20) + + for hidden_size in [32, 128]: + result = train_single_model( + None, None, None, None, + model_type='mlp', + vectorizer_type='tfidf', + epochs=50, + batch_size=BATCH_SIZE, + lr=LEARNING_RATE, + hidden_size=hidden_size + ) + results.append(result) + + # ===== 汇总报告 ===== + print("\n\n" + "=" * 70) + print("实验结果汇总") + print("=" * 70) + + print(f"\n{'配置':<40} {'训练准确率':<12} {'测试准确率':<12} {'训练时间(秒)':<10}") + print("-" * 76) + + for r in results: + config = f"{r['model_type'].upper()} + {r['vectorizer_type'].upper()}" + if r['lr'] != LEARNING_RATE: + config += f", lr={r['lr']}" + if r['hidden_size'] != HIDDEN_SIZE: + config += f", h={r['hidden_size']}" + print(f"{config:<40} {r['train_acc']:<12.4f} {r['test_acc']:<12.4f} {r['train_time']:<10.2f}") + + # 分析 + print("\n" + "=" * 70) + print("结果分析") + print("=" * 70) + + # 找出最佳配置 + best = max(results, key=lambda x: x['test_acc']) + print(f"\n最佳测试准确率: {best['test_acc']:.4f}") + print(f"最佳配置: {best['model_type'].upper()} + {best['vectorizer_type'].upper()}") + + # BoW vs TF-IDF 分析 + bow_results = [r for r in results if r['vectorizer_type'] == 'bow' and r['model_type'] == 'lr'] + tfidf_results = [r for r in results if r['vectorizer_type'] == 'tfidf' and r['model_type'] == 'lr'] + + if bow_results and tfidf_results: + bow_acc = bow_results[0]['test_acc'] + tfidf_acc = tfidf_results[0]['test_acc'] + print(f"\n向量化方法影响:") + print(f" BoW测试准确率: {bow_acc:.4f}") + print(f" TF-IDF测试准确率: {tfidf_acc:.4f}") + print(f" 差异: {abs(tfidf_acc - bow_acc):.4f} ({'TF-IDF更优' if tfidf_acc > bow_acc else 'BoW更优'})") + + # LR vs MLP 分析 + lr_results = [r for r in results if r['model_type'] == 'lr' and r['vectorizer_type'] == 'tfidf'] + mlp_results = [r for r in results if r['model_type'] == 'mlp' and r['vectorizer_type'] == 'tfidf'] + + if lr_results and mlp_results: + lr_acc = lr_results[0]['test_acc'] + mlp_acc = mlp_results[0]['test_acc'] + print(f"\n模型复杂度影响:") + print(f" LR测试准确率: {lr_acc:.4f}") + print(f" MLP测试准确率: {mlp_acc:.4f}") + print(f" 差异: {abs(mlp_acc - lr_acc):.4f} ({'MLP更优' if mlp_acc > lr_acc else 'LR更优'})") + + return results + + +def main(): + """主函数""" + print("=" * 70) + print("文本分类实验 - 纯NumPy实现") + print("数据集: ChnSentiCorp (中文酒店评论)") + print("=" * 70) + + if RUN_COMPARISON: + # 运行对比实验 + run_comparison_experiments() + else: + # 运行单个模型 + train_single_model( + None, None, None, None, + model_type=MODEL_TYPE, + vectorizer_type=VECTORIZER_TYPE, + epochs=NUM_EPOCHS, + batch_size=BATCH_SIZE, + lr=LEARNING_RATE, + hidden_size=HIDDEN_SIZE + ) + + print("\n" + "=" * 70) + print("实验完成!") + print("=" * 70) + + +if __name__ == '__main__': + main()