跳到主要内容

使用 Sentence Transformers 和 MLflow 进行高级语义搜索

下载此笔记本

开始一个动手实践之旅，探索使用 Sentence Transformers 和 MLflow 进行高级语义搜索。

您将学到什么​

  • 使用 sentence-transformers 实现高级语义搜索。
  • 根据独特的项目需求定制 MLflow 的 PythonModel。
  • 在 MLflow 生态系统内管理和记录模型。
  • 使用 MLflow 为实际应用部署复杂模型。

语义搜索超越了关键词匹配，它利用语言的细微之处和上下文来查找相关结果。这种高级方法反映了人类对语言的理解，考虑了词语在不同场景下的多种含义。

Sentence Transformers 专门用于处理富含上下文的句子嵌入，将搜索查询和文本语料库转换为语义向量。这使得能够识别语义相似的条目，是语义搜索的基石。

MLflow：模型管理和部署的先锋​

MLflow 通过高效的实验日志记录和可定制的模型环境增强了 NLP 项目。它为实验追踪带来了效率，并增加了定制层，这对于独特的 NLP 任务至关重要。

加入本教程，掌握高级语义搜索技术，并了解 MLflow 如何革新您处理 NLP 模型部署和管理的方法。

import warnings

# Disable a few less-than-useful UserWarnings from setuptools and pydantic
warnings.filterwarnings("ignore", category=UserWarning)

使用 MLflow 和 Sentence Transformers 理解语义搜索模型​

深入了解 SemanticSearchModel 的细节，这是使用 MLflow 和 Sentence Transformers 实现语义搜索的自定义模型。

MLflow 和自定义 PyFunc 模型​

MLflow 的自定义 Python 函数 (pyfunc) 模型提供了一种灵活且可部署的解决方案，用于集成复杂逻辑，非常适合我们的 SemanticSearchModel。

模型的核心功能​

  • 上下文加载：对于初始化 Sentence Transformer 模型和准备语料库进行语义比较至关重要。
  • 预测方法 (Predict Method)：语义搜索的核心函数，包括输入验证、查询编码和相似度计算。

预测方法 (Predict Method) 详细分解​

  • 输入验证：确保查询句子的格式正确和提取准确。
  • 查询编码：将查询转换为用于比较的嵌入。
  • 余弦相似度计算：确定每个语料库条目与查询的相关性。
  • 提取顶部结果：根据相似度分数确定最相关的条目。
  • 相关性过滤：基于最小相关性阈值过滤结果，提高实际可用性。
  • 警告机制：如果所有顶部结果都低于相关性阈值，则发出警告，确保始终提供结果。

结论​

这个语义搜索模型展示了 NLP 与 MLflow 的集成，体现了灵活性、用户友好性以及在现代机器学习工作流中的实际应用。

import warnings

import numpy as np
import pandas as pd
from sentence_transformers import SentenceTransformer, util

import mlflow
from mlflow.models.signature import infer_signature
from mlflow.pyfunc import PythonModel


class SemanticSearchModel(PythonModel):
  def load_context(self, context):
      """Load the model context for inference, including the corpus from a file."""
      try:
          # Load the pre-trained sentence transformer model
          self.model = SentenceTransformer.load(context.artifacts["model_path"])

          # Load the corpus from the specified file
          corpus_file = context.artifacts["corpus_file"]
          with open(corpus_file) as file:
              self.corpus = file.read().splitlines()

          # Encode the corpus and convert it to a tensor
          self.corpus_embeddings = self.model.encode(self.corpus, convert_to_tensor=True)

      except Exception as e:
          raise ValueError(f"Error loading model and corpus: {e}")

  def predict(self, context, model_input, params=None):
      """Predict method to perform semantic search over the corpus."""

      if isinstance(model_input, pd.DataFrame):
          if model_input.shape[1] != 1:
              raise ValueError("DataFrame input must have exactly one column.")
          model_input = model_input.iloc[0, 0]
      elif isinstance(model_input, dict):
          model_input = model_input.get("sentence")
          if model_input is None:
              raise ValueError("The input dictionary must have a key named 'sentence'.")
      else:
          raise TypeError(
              f"Unexpected type for model_input: {type(model_input)}. Must be either a Dict or a DataFrame."
          )

      # Encode the query
      query_embedding = self.model.encode(model_input, convert_to_tensor=True)

      # Compute cosine similarity scores
      cos_scores = util.cos_sim(query_embedding, self.corpus_embeddings)[0]

      # Determine the number of top results to return
      top_k = params.get("top_k", 3) if params else 3  # Default to 3 if not specified

      minimum_relevancy = (
          params.get("minimum_relevancy", 0.2) if params else 0.2
      )  # Default to 0.2 if not specified

      # Get the top_k most similar sentences from the corpus
      top_results = np.argsort(cos_scores, axis=0)[-top_k:]

      # Prepare the initial results list
      initial_results = [
          (self.corpus[idx], cos_scores[idx].item()) for idx in reversed(top_results)
      ]

      # Filter the results based on the minimum relevancy threshold
      filtered_results = [result for result in initial_results if result[1] >= minimum_relevancy]

      # If all results are below the threshold, issue a warning and return the top result
      if not filtered_results:
          warnings.warn(
              "All top results are below the minimum relevancy threshold. "
              "Returning the highest match instead.",
              RuntimeWarning,
          )
          return [initial_results[0]]
      else:
          return filtered_results

构建和准备语义搜索语料库​

探索如何构建和准备语义搜索模型的语料库，这是搜索功能的关键组成部分。

模拟真实世界用例​

我们创建了一个简化的合成博客文章语料库，以演示模型的核心功能，复制了典型真实世界场景的缩小版本。

语料库准备的关键步骤​

  • 语料库创建：形成一个列表，代表单个博客文章条目。
  • 写入文件：将语料库保存到文本文件，模拟真实应用中数据提取和预处理的过程。

高效的数据处理以实现可扩展性​

我们的模型将语料库编码成嵌入向量，以便快速比较，展示了一种适合扩展到更大数据集的高效方法。

生产环境考量​

  • 存储嵌入向量：讨论了高效存储和检索嵌入向量的选项，这在大规模应用中至关重要。
  • 可扩展性：强调了可扩展存储系统对于处理海量数据集和复杂查询的重要性。
  • 更新语料库：概述了在动态、不断演变的用例中管理和更新语料库的策略。

实现语义搜索概念​

这种设置虽然简化了，但反映了开发健壮且可扩展的语义搜索系统的基本步骤，将 NLP 技术与高效的数据管理相结合。在真实的生产用例中，处理语料库（创建嵌入向量）将是独立于运行语义搜索的外部过程。下面的语料库示例仅用于演示功能。

corpus = [
  "Perfecting a Sourdough Bread Recipe: The Joy of Baking. Baking sourdough bread "
  "requires patience, skill, and a good understanding of yeast fermentation. Each "
  "loaf is unique, telling its own story of the baker's journey.",
  "The Mars Rover's Discoveries: Unveiling the Red Planet. NASA's Mars rover has "
  "sent back stunning images and data, revealing the planet's secrets. These "
  "discoveries may hold the key to understanding Mars' history.",
  "The Art of Growing Herbs: Enhancing Your Culinary Skills. Growing your own "
  "herbs can transform your cooking, adding fresh and vibrant flavors. Whether it's "
  "basil, thyme, or rosemary, each herb has its own unique characteristics.",
  "AI in Software Development: Transforming the Tech Landscape. The rapid "
  "advancements in artificial intelligence are reshaping how we approach software "
  "development. From automation to machine learning, the possibilities are endless.",
  "Backpacking Through Europe: A Journey of Discovery. Traveling across Europe by "
  "backpack allows one to immerse in diverse cultures and landscapes. It's an "
  "adventure that combines the thrill of exploration with personal growth.",
  "Shakespeare's Timeless Influence: Reshaping Modern Storytelling. The works of "
  "William Shakespeare continue to inspire and influence contemporary literature. "
  "His mastery of language and deep understanding of human nature are unparalleled.",
  "The Rise of Renewable Energy: A Sustainable Future. Embracing renewable energy "
  "is crucial for achieving a sustainable and environmentally friendly lifestyle. "
  "Solar, wind, and hydro power are leading the way in this green revolution.",
  "The Magic of Jazz: An Exploration of Sound and Harmony. Jazz music, known for "
  "its improvisation and complex harmonies, has a rich and diverse history. It "
  "evokes a range of emotions, often reflecting the soul of the musician.",
  "Yoga for Mind and Body: The Benefits of Regular Practice. Engaging in regular "
  "yoga practice can significantly improve flexibility, strength, and mental "
  "well-being. It's a holistic approach to health, combining physical and spiritual "
  "aspects.",
  "The Egyptian Pyramids: Monuments of Ancient Majesty. The ancient Egyptian "
  "pyramids, monumental tombs for pharaohs, are marvels of architectural "
  "ingenuity. They stand as a testament to the advanced skills of ancient builders.",
  "Vegan Cuisine: A World of Flavor. Exploring vegan cuisine reveals a world of "
  "nutritious and delicious possibilities. From hearty soups to delectable desserts, "
  "plant-based dishes are diverse and satisfying.",
  "Extraterrestrial Life: The Endless Search. The quest to find life beyond Earth "
  "continues to captivate scientists and the public alike. Advances in space "
  "technology are bringing us closer to answering this age-old question.",
  "The Art of Plant Pruning: Promoting Healthy Growth. Regular pruning is essential "
  "for maintaining healthy and vibrant plants. It's not just about cutting back, but "
  "understanding each plant's growth patterns and needs.",
  "Cybersecurity in the Digital Age: Protecting Our Data. With the rise of digital "
  "technology, cybersecurity has become a critical concern. Protecting sensitive "
  "information from cyber threats is an ongoing challenge for individuals and "
  "businesses alike.",
  "The Great Wall of China: A Historical Journey. Visiting the Great Wall offers "
  "more than just breathtaking views; it's a journey through history. This ancient "
  "structure tells stories of empires, invasions, and human resilience.",
  "Mystery Novels: Crafting Suspense and Intrigue. A great mystery novel captivates "
  "the reader with intricate plots and unexpected twists. It's a genre that combines "
  "intellectual challenge with entertainment.",
  "Conserving Endangered Species: A Global Effort. Protecting endangered species "
  "is a critical task that requires international collaboration. From rainforests to "
  "oceans, every effort counts in preserving our planet's biodiversity.",
  "Emotions in Classical Music: A Symphony of Feelings. Classical music is not just "
  "an auditory experience; it's an emotional journey. Each composition tells a story, "
  "conveying feelings from joy to sorrow, tranquility to excitement.",
  "CrossFit: A Test of Strength and Endurance. CrossFit is more than just a fitness "
  "regimen; it's a lifestyle that challenges your physical and mental limits. It "
  "combines various disciplines to create a comprehensive workout.",
  "The Renaissance: An Era of Artistic Genius. The Renaissance marked a period of "
  "extraordinary artistic and scientific achievements. It was a time when creativity "
  "and innovation flourished, reshaping the course of history.",
  "Exploring International Cuisines: A Culinary Adventure. Discovering international "
  "cuisines is an adventure for the palate. Each dish offers a glimpse into the "
  "culture and traditions of its origin.",
  "Astronaut Training: Preparing for the Unknown. Becoming an astronaut involves "
  "rigorous training to prepare for the extreme conditions of space. It's a journey "
  "that tests both physical endurance and mental resilience.",
  "Sustainable Gardening: Nurturing the Environment. Sustainable gardening is not "
  "just about growing plants; it's about cultivating an ecosystem. By embracing "
  "environmentally friendly practices, gardeners can have a positive impact on the "
  "planet.",
  "The Smartphone Revolution: Changing Communication. Smartphones have transformed "
  "how we communicate, offering unprecedented connectivity and convenience. This "
  "technology continues to evolve, shaping our daily interactions.",
  "Experiencing African Safaris: Wildlife and Wilderness. An African safari is an "
  "unforgettable experience that brings you face-to-face with the wonders of "
  "wildlife. It's a journey that connects you with the raw beauty of nature.",
  "Graphic Novels: A Blend of Art and Story. Graphic novels offer a unique medium "
  "where art and narrative intertwine to tell compelling stories. They challenge "
  "traditional forms of storytelling, offering visual and textual richness.",
  "Addressing Ocean Pollution: A Call to Action. The increasing levels of pollution "
  "in our oceans are a pressing environmental concern. Protecting marine life and "
  "ecosystems requires concerted global efforts.",
  "The Origins of Hip Hop: A Cultural Movement. Hip hop music, originating from the "
  "streets of New York, has grown into a powerful cultural movement. Its beats and "
  "lyrics reflect the experiences and voices of a community.",
  "Swimming: A Comprehensive Workout. Swimming offers a full-body workout that is "
  "both challenging and refreshing. It's an exercise that enhances cardiovascular "
  "health, builds muscle, and improves endurance.",
  "The Fall of the Berlin Wall: A Historical Turning Point. The fall of the Berlin "
  "Wall was not just a physical demolition; it was a symbol of political and social "
  "change. This historic event marked the end of an era and the beginning of a new "
  "chapter in world history.",
]

# Write the corpus to a file
corpus_file = "/tmp/search_corpus.txt"
with open(corpus_file, "w") as file:
  for sentence in corpus:
      file.write(sentence + "
")

MLflow 中的模型准备和配置​

探索准备和配置 Sentence Transformer 模型以与 MLflow 集成的步骤，这对于部署就绪至关重要。

加载和保存 Sentence Transformer 模型​

  • 模型初始化：加载 "all-MiniLM-L6-v2" 模型，该模型以其在性能和速度方面的平衡而闻名，适用于语义搜索任务。
  • 模型存储：将模型保存到目录，这对于稍后通过 MLflow 部署至关重要。选择 /tmp/search_model 是为了教程的便利性，以便您的当前工作目录不会被模型文件填满。您可以将其更改为任意位置。

准备模型 Artifacts 和 Signature​

  • Artifacts 字典：创建一个包含模型和语料库文件路径的字典，引导 MLflow 找到初始化自定义模型对象所需的组件。
  • 输入示例和测试输出：定义样本输入和输出，以说明模型期望的数据格式。
  • 模型 Signature：使用 infer_signature 自动生成 Signature，包括输入、输出和操作参数。

模型 Signature 的重要性​

Signature 确保训练和部署之间的数据一致性，增强模型可用性并减少潜在错误。指定 Signature 可确保在推理时进行类型验证，防止因无效类型转换导致不正确或令人困惑的推理结果而产生意外行为。

结论​

这一全面的准备过程保证模型已为部署做好准备，所有依赖项和操作要求都已明确定义。

# Load a pre-trained sentence transformer model
model = SentenceTransformer("all-MiniLM-L6-v2")

# Create an input example DataFrame
input_example = ["Something I want to find matches for."]

# Save the model in the /tmp directory
model_directory = "/tmp/search_model"
model.save(model_directory)

artifacts = {"model_path": model_directory, "corpus_file": corpus_file}

# Generate test output for signature
test_output = ["match 1", "match 2", "match 3"]

# Define the signature associated with the model
signature = infer_signature(
  input_example, test_output, params={"top_k": 3, "minimum_relevancy": 0.2}
)

# Visualize the signature
signature
inputs: 
[string]
outputs: 
[string]
params: 
['top_k': long (default: 3), 'minimum_relevancy': double (default: 0.2)]

创建实验​

我们创建一个新的 MLflow 实验，以便我们要记录模型的运行不会记录到默认实验中，而是拥有自己的上下文相关条目。

# If you are running this tutorial in local mode, leave the next line commented out.
# Otherwise, uncomment the following line and set your tracking uri to your local or remote tracking server.

# mlflow.set_tracking_uri("http://127.0.0.1:8080")

mlflow.set_experiment("Semantic Similarity")
<Experiment: artifact_location='file:///Users/benjamin.wilson/repos/mlflow-fork/mlflow/docs/source/llms/sentence-transformers/tutorials/semantic-search/mlruns/405641275158666585', creation_time=1701278766302, experiment_id='405641275158666585', last_update_time=1701278766302, lifecycle_stage='active', name='Semantic Similarity', tags={}>

使用 MLflow 记录模型​

了解在 MLflow 中记录模型的过程，这是在 MLflow 框架内管理和部署模型的关键步骤。

启动 MLflow Run​

  • 上下文管理：使用 with mlflow.start_run() 启动 MLflow run，这对于追踪和管理模型相关操作至关重要。

记录模型​

  • 模型记录：利用 mlflow.pyfunc.log_model 记录自定义的 SemanticSearchModel，包括模型名称、实例、输入示例、签名、artifacts 和依赖项等关键参数。

模型记录的结果​

  • 模型注册：确保模型及其所有必需组件在 MLflow 中注册，以便部署。
  • 可复现性和可追溯性：促进模型部署的一致性，并追踪版本和相关数据。

结论​

完成这一关键步骤将模型从开发阶段转换为部署就绪状态，并封装在 MLflow 生态系统内。

with mlflow.start_run() as run:
  model_info = mlflow.pyfunc.log_model(
      "semantic_search",
      python_model=SemanticSearchModel(),
      input_example=input_example,
      signature=signature,
      artifacts=artifacts,
      pip_requirements=["sentence_transformers", "numpy"],
  )
Downloading artifacts:   0%|          | 0/11 [00:00<?, ?it/s]
2023/11/30 15:57:53 INFO mlflow.store.artifact.artifact_repo: The progress bar can be disabled by setting the environment variable MLFLOW_ENABLE_ARTIFACTS_PROGRESS_BAR to false
Downloading artifacts:   0%|          | 0/1 [00:00<?, ?it/s]

模型推理和预测演示​

观察我们的语义搜索模型的实际应用，演示其响应用户查询并提供相关预测的能力。

加载模型进行推理​

  • 模型加载：利用 mlflow.pyfunc.load_model 加载模型，准备处理语义搜索查询。

进行预测​

  • 运行查询：将示例查询传递给已加载的模型，演示其语义搜索能力。

理解预测输出​

  • 输出格式：分析预测输出，通过相关性分数展示模型的语义理解能力。
  • 示例结果：演示模型的结果，包括与查询相关的各种条目的相关性分数。

结论​

此演示强调了模型在语义搜索中的有效性，突显了其在推荐和知识检索应用中的潜力。

# Load our model as a PyFuncModel.
# Note that unlike the example shown in the Introductory Tutorial, there is no 'native' flavor for PyFunc models.
# This model cannot be loaded with `mlflow.sentence_transformers.load_model()` because it is not in the native model format.
loaded_dynamic = mlflow.pyfunc.load_model(model_info.model_uri)

# Make sure that it generates a reasonable output
loaded_dynamic.predict(["I'd like some ideas for a meal to cook."])
[('Exploring International Cuisines: A Culinary Adventure. Discovering international cuisines is an adventure for the palate. Each dish offers a glimpse into the culture and traditions of its origin.',
0.43857115507125854),
('Vegan Cuisine: A World of Flavor. Exploring vegan cuisine reveals a world of nutritious and delicious possibilities. From hearty soups to delectable desserts, plant-based dishes are diverse and satisfying.',
0.34688490629196167),
("The Art of Growing Herbs: Enhancing Your Culinary Skills. Growing your own herbs can transform your cooking, adding fresh and vibrant flavors. Whether it's basil, thyme, or rosemary, each herb has its own unique characteristics.",
0.22686949372291565)]

使用可定制参数和警告机制进行高级查询处理​

探索模型的高级功能，包括可定制的搜索参数和独特的警告机制，以提供最佳用户体验。

执行带警告的定制预测​

  • 带有挑战性参数的定制查询：测试模型在高相关性阈值查询下区分高度相关内容的能力。
  • 触发警告：当搜索条件过于严格时，提醒用户的机制，增强用户反馈。

理解模型的响应​

  • 在挑战性场景下的结果：分析模型对严格搜索条件的响应，包括未达到相关性阈值的情况。

影响和最佳实践​

  • 平衡相关性与覆盖率：讨论设置适当相关性阈值的重要性，以确保精度和结果覆盖率之间的平衡。
  • 利用用户反馈改进语料库：利用警告作为反馈来优化语料库并增强搜索系统。

结论​

这一高级功能集展示了模型的适应性以及微调搜索参数对于动态响应式搜索体验的重要性。

# Verify that the fallback logic works correctly by returning the 'best, closest' result, even though the parameters submitted should return no results.
# We are also validating that the warning is issued, alerting us to the fact that this behavior is occurring.
loaded_dynamic.predict(
  ["Latest stories on computing"], params={"top_k": 10, "minimum_relevancy": 0.4}
)
/var/folders/cd/n8n0rm2x53l_s0xv_j_xklb00000gp/T/ipykernel_55915/1325605132.py:71: RuntimeWarning: All top results are below the minimum relevancy threshold. Returning the highest match instead.
warnings.warn(
[('AI in Software Development: Transforming the Tech Landscape. The rapid advancements in artificial intelligence are reshaping how we approach software development. From automation to machine learning, the possibilities are endless.',
0.2533860206604004)]

结论：使用 MLflow 的 PythonModel 构建自定义逻辑​

在结束本教程之际，让我们回顾一下主要学习内容以及 MLflow 的 PythonModel 在为实际应用构建自定义逻辑方面的强大能力，尤其是在集成像 sentence-transformers 这样的高级库时。

主要收获​

  1. PythonModel 的灵活性:

    • MLflow 中的 PythonModel 在定义自定义逻辑方面提供了无与伦比的灵活性。在本教程中，我们利用这一点构建了一个根据我们的特定需求量身定制的语义搜索模型。
    • 当处理超出标准模型实现的复杂用例时，这种灵活性是无价的。

  2. 与 Sentence Transformers 的集成:

    • 我们在 MLflow 模型中无缝集成了 sentence-transformers 库。这展示了如何将高级 NLP 功能嵌入到自定义模型中，以处理像语义搜索这样的复杂任务。
    • 使用 Transformer 模型生成嵌入向量展示了如何在实际场景中应用前沿的 NLP 技术。

  3. 定制和用户体验:

    • 我们的模型不仅执行了语义搜索的核心任务，还允许定制搜索参数（top_k 和 minimum_relevancy）。这种程度的定制对于使模型的输出与不同的用户需求保持一致至关重要。
    • 包含警告机制通过提供有价值的反馈进一步丰富了模型，增强了用户体验。

  4. 实际应用和可扩展性:

    • 虽然我们的教程侧重于受控数据集，但其原则和方法适用于更大、更真实的world数据集。关于使用向量数据库和 Redis 或 Elasticsearch 等内存数据库来实现可扩展性的讨论，突显了模型如何适应大规模应用。

赋能实际应用​

  • MLflow 的 PythonModel 与 sentence-transformers 等高级库的结合，简化了复杂实际应用的创建过程。
  • 封装复杂逻辑、管理依赖项和确保模型可移植性的能力使 MLflow 成为现代数据科学家工具包中不可或缺的工具。

未来展望​

  • 最后，请记住旅程并未在此结束。本教程中探讨的概念和技术为 NLP 及更广泛领域的进一步探索和创新奠定了基础。
  • 我们鼓励您运用这些知识，使用自己的数据集进行实验，并继续突破 MLflow 和先进 NLP 技术所能实现的界限。

感谢您加入我们这个使用 Sentence Transformers 和 MLflow 进行语义搜索的启迪之旅！