From b2d4ce660d7a59cea41f4369779a26366cd23063 Mon Sep 17 00:00:00 2001 From: QCQCQC <1220204124@zust.edu.cn> Date: Wed, 23 Apr 2025 15:18:30 +0800 Subject: [PATCH] bug fixed --- perf.py | 85 ++++++++++++++++++++++++++++++++++++++----------- vector_tests.py | 13 ++++---- 2 files changed, 73 insertions(+), 25 deletions(-) diff --git a/perf.py b/perf.py index 47ad18d..15da726 100644 --- a/perf.py +++ b/perf.py @@ -5,12 +5,32 @@ import faiss import numpy as np import random import string +import psutil +import os +# import pynvml + +# 初始化 GPU 监控 +# pynvml.nvmlInit() +# gpu_handle = pynvml.nvmlDeviceGetHandleByIndex(0) # 默认用第一个GPU + +def get_memory_info(): + process = psutil.Process(os.getpid()) + mem_info = process.memory_info() + return mem_info.rss / 1024 / 1024 # 单位 MB + +def get_gpu_memory_info(): + # mem = pynvml.nvmlDeviceGetMemoryInfo(gpu_handle) + # used = mem.used / 1024 / 1024 # MB + # total = mem.total / 1024 / 1024 + # return used, total + return 0, 0 + # 1. 初始化模型 model = SentenceTransformer('all-MiniLM-L6-v2') # 2. 模拟一个更大的错误数据库 -NUM_ERRORS = 1000 +NUM_ERRORS = 100 error_db_large = [] def generate_random_traceback(num_lines=10): @@ -61,20 +81,32 @@ for entry in error_db_large: "original_error": entry["error"], "solution": entry["solution"] }) + cleaning_time = time.time() - start_time print(f"清洗 {NUM_ERRORS} 条错误信息耗时: {cleaning_time:.4f} 秒") +print(f"内存使用: {get_memory_info():.2f} MB") +gpu_used, gpu_total = get_gpu_memory_info() +print(f"GPU 显存使用: {gpu_used:.2f} / {gpu_total:.2f} MB") # 5. 向量化清洗后的错误信息并构建索引 start_time = time.time() embeddings_large = np.array([model.encode(e['cleaned_error']) for e in cleaned_error_db_large]) embedding_time = time.time() - start_time print(f"向量化 {NUM_ERRORS} 条错误信息耗时: {embedding_time:.4f} 秒") +print(f"内存使用: {get_memory_info():.2f} MB") +gpu_used, gpu_total = get_gpu_memory_info() +print(f"GPU 显存使用: {gpu_used:.2f} / {gpu_total:.2f} MB") + start_time = time.time() index_large = faiss.IndexFlatL2(embeddings_large.shape[1]) index_large.add(embeddings_large) index_building_time = time.time() - start_time print(f"构建包含 {NUM_ERRORS} 条向量的 FAISS 索引耗时: {index_building_time:.4f} 秒") +print(f"内存使用: {get_memory_info():.2f} MB") +gpu_used, gpu_total = get_gpu_memory_info() +print(f"GPU 显存使用: {gpu_used:.2f} / {gpu_total:.2f} MB") + # 6. 用户报错（测试查询） user_error_performance = generate_random_traceback(num_lines=12) @@ -82,23 +114,38 @@ cleaned_user_error_performance = clean_traceback(user_error_performance) query_vec_performance = model.encode(cleaned_user_error_performance) query_vec_performance = np.array([query_vec_performance]) -# 7. 在索引中搜索最相似的错误 -start_time = time.time() -D_performance, I_performance = index_large.search(query_vec_performance, k=1) -search_time = time.time() - start_time -print(f"在包含 {NUM_ERRORS} 条向量的 FAISS 索引中搜索耗时: {search_time:.6f} 秒") -# 8. 输出结果 -matched_index_performance = I_performance[0][0] -distance_performance = D_performance[0][0] -cosine_similarity_performance = 1 - (distance_performance**2) / 2 -matched_entry_performance = cleaned_error_db_large[matched_index_performance] +NUM_QUERIES = 100 # 查询次数 +k = 10 # 每次查询返回前 k 个最相似的错误 -print("\n--- 性能测试结果 ---") -print("\n--- 清洗后的用户错误信息 ---") -print(cleaned_user_error_performance) -print("\n--- 最相似的已知错误信息 ---") -print(matched_entry_performance['cleaned_error']) -print("\n--- 建议的解决方案 ---") -print(matched_entry_performance['solution']) -print(f"\n--- 匹配度 (Cosine Similarity): {cosine_similarity_performance:.4f}") \ No newline at end of file +query_times = [] +similarities = [] + +for i in range(NUM_QUERIES): + user_error = generate_random_traceback(num_lines=random.randint(8, 14)) + cleaned_user_error = clean_traceback(user_error) + query_vec = model.encode(cleaned_user_error) + query_vec = np.array([query_vec]) + + start_query_time = time.time() + D, I = index_large.search(query_vec, k=k) + elapsed = time.time() - start_query_time + query_times.append(elapsed) + + matched_index = I[0][0] + distance = D[0][0] + similarity = 1 - (distance ** 2) / 2 + similarities.append(similarity) + +# 输出统计结果 +print(f"\n--- 多次查询性能测试 ---") +print(f"总查询次数: {NUM_QUERIES}") +print(f"平均查询耗时: {np.mean(query_times) * 1000:.4f} ms") +print(f"最小查询耗时: {np.min(query_times) * 1000:.4f} ms") +print(f"最大查询耗时: {np.max(query_times) * 1000:.4f} ms") +print(f"查询耗时标准差: {np.std(query_times) * 1000:.4f} ms") +print(f"平均匹配相似度: {np.mean(similarities):.4f}") +print(f"相似度标准差: {np.std(similarities):.4f}") +print(f"内存使用: {get_memory_info():.2f} MB") +gpu_used, gpu_total = get_gpu_memory_info() +print(f"GPU 显存使用: {gpu_used:.2f} / {gpu_total:.2f} MB") diff --git a/vector_tests.py b/vector_tests.py index 2c46215..7e1ac0e 100644 --- a/vector_tests.py +++ b/vector_tests.py @@ -90,7 +90,7 @@ for entry, agg_vec in zip(error_db, aggregate_vectors): layered_error_db.append({ "layers": layers, - "vectors": embeddings, # 保留原始向量（可选） + "vectors": embeddings, # 保留原始向量 "index": index, # 添加 FAISS 索引 "solution": entry["solution"], "agg_vector": agg_vec # 添加聚合向量 @@ -105,9 +105,9 @@ logging.info(f"创建数据库耗时：{end_db_time - start_db_time:.2f} 秒") # 完全不同 # user_error = """"Traceback (most recent call last):\n File \"D:\\Develop\\Projects\\bash-hook\\matches\\perf_2.py\", line 94, in \n D, I = index.search(query_vec, k=1)\n File \"C:\\Users\\qcqcqc\\.conda\\envs\\sentence_transformers\\lib\\site-packages\\faiss\\class_wrappers.py\", line 329, in replacement_search\n assert d == self.d\nAssertionError\n""" # 底层不同 -user_error = """Traceback (most recent call last):\n File \"image_error.py\", line 55, in \n loss = criterion(outputs, labels)\n File \"module.py\", line 1553, in _wrapped_call_impl\n return self._call_impl(*args, **kwargs)\n File \"module.py\", line 1570, in _call_impl\n return forward_call(*args, **kwargs)\n File \"loss.py\", line 1200, in forward\n return F.mse_loss(input, target, reduction=self.reduction)\n File \"functional.py\", line 2301, in mse_loss\n return mean_squared_error_loss(input, target, ...)\nValueError: The size of input tensor must match the size of target tensor.\n""" +# user_error = """Traceback (most recent call last):\n File \"image_error.py\", line 55, in \n loss = criterion(outputs, labels)\n File \"module.py\", line 1553, in _wrapped_call_impl\n return self._call_impl(*args, **kwargs)\n File \"module.py\", line 1570, in _call_impl\n return forward_call(*args, **kwargs)\n File \"loss.py\", line 1200, in forward\n return F.mse_loss(input, target, reduction=self.reduction)\n File \"functional.py\", line 2301, in mse_loss\n return mean_squared_error_loss(input, target, ...)\nValueError: The size of input tensor must match the size of target tensor.\n""" # 顶层不同 -# user_error = """Traceback (most recent call last):\n File \"train_model.py\", line 72, in \n loss = loss_function(predictions, ground_truth)\n File \"core.py\", line 442, in _call_function\n return self._execute(*args, **kwargs)\n File \"core.py\", line 450, in _execute\n return forward_execution(*args, **kwargs)\n File \"loss_functions.py\", line 205, in forward\n return F.cross_entropy(input, target, weight=self.weight,\n File \"functional.py\", line 3104, in cross_entropy\n return cross_entropy_loss(input, target, ...)\nIndexError: Target 15 is out of bounds.\n""" +user_error = """Traceback (most recent call last):\n File \"train_model.py\", line 72, in \n loss = loss_function(predictions, ground_truth)\n File \"core.py\", line 442, in _call_function\n return self._execute(*args, **kwargs)\n File \"core.py\", line 450, in _execute\n return forward_execution(*args, **kwargs)\n File \"loss_functions.py\", line 205, in forward\n return F.cross_entropy(input, target, weight=self.weight,\n File \"functional.py\", line 3104, in cross_entropy\n return cross_entropy_loss(input, target, ...)\nIndexError: Target 15 is out of bounds.\n""" # 无错误信息 # user_error = "success." @@ -134,11 +134,12 @@ def compute_layered_similarity_avg(user_vecs, db_index, layer_weights=None): def compute_layered_similarity_sco(user_vecs, db_index, layer_weights=None): - score = 0.0 weighted_score = 0.0 if layer_weights is None: - layer_weights = np.linspace(0.1, 1, len(user_vecs)) # 默认从浅层到深层递增权重 + # 使用对数函数生成非线性权重 + layer_weights = np.logspace(0, 1, len(user_vecs)) # 从浅层到深层递增权重 + layer_weights /= np.sum(layer_weights) # 归一化权重，使其和为1 for i, u_vec in enumerate(user_vecs): u_vec = u_vec.astype('float32').reshape(1, -1) @@ -147,7 +148,7 @@ def compute_layered_similarity_sco(user_vecs, db_index, layer_weights=None): weighted_score += max_sim * layer_weights[i] - return weighted_score / np.sum(layer_weights) if len(user_vecs) > 0 else 0.0 + return weighted_score if len(user_vecs) > 0 else 0.0 # 计时：查询