docs/tutorials/3_autoclass.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 写在最前\n",
    "\n",
    "请记住，在本章的内容中，“架构”指的是模型的结构，而 \"checkpoint\" 是给定架构的权重（也就是预训练模型的权重，也可指代预训练模型本身）。\n",
    "\n",
    "例如，“BERT” 是一种架构，而 “google-bert/bert-base-uncased” 是一个 checkpoint。“模型”则是一个通用术语，可以指代架构或checkpoint。\n",
    "\n",
    "在 `🤗 Transformers` 的示例代码中，你通过替换 \"model\" 参数的值来选择不同的模型。\n",
    "\n",
    "在这个教程中，将会学习如何：\n",
    "\n",
    "1. 加载预训练的分词器（tokenizer）\n",
    "2. 加载预训练的图像处理器(image processor)\n",
    "3. 加载预训练的特征提取器(feature extractor)\n",
    "4. 加载预训练的处理器(processor)\n",
    "5. 加载预训练的模型。\n",
    "\n",
    "# 使用AutoClass加载预训练实例\n",
    "\n",
    "由于存在许多不同的 Transformer 架构，因此为 checkpoint 创建一个可用的架构可能会具有挑战性。\n",
    "\n",
    "**通过 `AutoClass` 可以自动推断并从给定的 checkpoint 加载正确的架构**, 这也是体现 🤗 Transformers 易于使用、简单且灵活的核心部分。\n",
    "\n",
    "**如果你的代码适用于一个 checkpoint，它将适用于另一个 checkpoint**，只要它们是为了相似的任务进行训练的，即使它们的架构有所不同。\n",
    "\n",
    "`from_pretrained()` 方法允许你快速加载任何架构的预训练模型，不需要花费时间和精力从头开始训练模型。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## AutoTokenizer\n",
    "\n",
    "几乎所有的 NLP 任务都以 `tokenizer` 开始。Tokenizer 能够将你的输入转换为模型可以理解和处理的格式(即模型正确的输入格式)。\n",
    "\n",
    "使用 `AutoTokenizer.from_pretrained()` 加载 tokenizer："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoTokenizer\n",
    "\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"google-bert/bert-base-uncased\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "然后按照如下方式对输入进行分词："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "sequence = \"In a hole in the ground there lived a hobbit.\"\n",
    "print(tokenizer(sequence))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## AutoImageProcessor\n",
    "\n",
    "对于视觉任务，`ImageProcessor` 能够将图像处理成正确的输入格式。\n",
    "\n",
    "这里使用`AutoImageProcessor.from_pretrained()`加载 Image Processor："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoImageProcessor\n",
    "\n",
    "image_processor = AutoImageProcessor.from_pretrained(\"google/vit-base-patch16-224\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## AutoFeatureExtractor\n",
    "\n",
    "对于音频任务, `Feature Extractor` 将音频信号处理成正确的输入格式。\n",
    "\n",
    "这里使用`AutoFeatureExtractor.from_pretrained()`加载 Feature Extractor："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoFeatureExtractor\n",
    "\n",
    "feature_extractor = AutoFeatureExtractor.from_pretrained(\n",
    "    \"ehcalabres/wav2vec2-lg-xlsr-en-speech-emotion-recognition\"\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## AutoProcessor\n",
    "\n",
    "对于多模态任务只需要一个 `Processor` 就能够将两种类型的预处理工具结合起来。\n",
    "\n",
    "例如，LayoutLMV2 模型需要一个 image processor 来处理图像和一个 tokenizer 来处理文本；Processor 会将两者结合起来。\n",
    "\n",
    "使用`AutoProcessor.from_pretrained()`加载 Processor："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoProcessor\n",
    "\n",
    "processor = AutoProcessor.from_pretrained(\"microsoft/layoutlmv2-base-uncased\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## AutoModel\n",
    "\n",
    "`AutoModelFor*` 类能够加载给定任务类型的预训练模型（[支持的任务列表](https://huggingface.co/docs/transformers/main/en/model_doc/auto)）。\n",
    "\n",
    "例如，使用 `AutoModelForSequenceClassification.from_pretrained()` 加载用于序列分类的模型："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# PyTorch\n",
    "from transformers import AutoModelForSequenceClassification\n",
    "\n",
    "model = AutoModelForSequenceClassification.from_pretrained(\"distilbert/distilbert-base-uncased\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "可以使用相同的 checkpoint 来为不同的任务加载模型架构："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# PyTorch\n",
    "from transformers import AutoModelForTokenClassification\n",
    "\n",
    "model = AutoModelForTokenClassification.from_pretrained(\"distilbert/distilbert-base-uncased\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "对于 PyTorch 模型，`from_pretrained()`方法在加载预训练模型时，通常使用`torch.load()`，它内部使用的是不安全的 `pickle`， pickle 加载不可信的数据可能存在安全风险（如代码执行漏洞）。\n",
    "\n",
    "永远不要加载那些来自不可信来源或可能被篡改的模型。建议对于托管在 Hugging Face Hub 上的公共模型，因为每次提交都会进行恶意软件扫描，在一定程度上缓解了安全风险。\n",
    "\n",
    "建议使用[使用 GPG 进行签名提交验证](https://huggingface.co/docs/hub/security-gpg#signing-commits-with-gpg)等最佳实践来进一步确保安全性，可参阅[Hub 文档](https://huggingface.co/docs/hub/security)以了解最佳实践。\n",
    "\n",
    "**TensorFlow 和 Flax 在加载 checkpoints 不使用 pickle ，不存在由 pickle 引起的安全问题。** \n",
    "\n",
    "在 PyTorch 中可以通过 `from_pretrained()` 方法的 `from_tf` 和 `from_flax` 参数来加载 TensorFlow 和 Flax 的 checkpoints，进而绕过 torch.load() 使用 pickle 的安全风险。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 实际操作示例\n",
    "\n",
    "通过加载 TensorFlow和 Flax 的 checkpoints，可以避免使用 torch.load()及其pickle的安全风险。\n",
    "\n",
    "假设你想要加载一个在 Hugging Face Hub 上托管的 TensorFlow 模型到 PyTorch 中，可以这样做："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoModel\n",
    "\n",
    "# 加载TensorFlow模型，并将 TensorFlow 模型转换为 PyTorch 模型\n",
    "pt_model = AutoModel.from_pretrained(\"bert-base-uncased\", from_tf=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "类似地，如果你要加载一个 Flax 模型到 PyTorch 中："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoModel\n",
    "\n",
    "# 加载 Flax 模型，并将 Flax 模型转换为 PyTorch 模型\n",
    "pt_model = AutoModel.from_pretrained(\"bert-base-uncased\", from_flax=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "一般来说，我们不建议使用通用的`AutoModel`来加载预训练模型，而是使用`AutoTokenizer`类和`AutoModelFor*`类来加载预训练的模型实例，因为这样可以确保加载的是正确的架构。\n",
    "\n",
    "在下一个[教程](../docs/4_preprocess_data.ipynb)中，学习如何使用新加载的 tokenizer, image processor, feature extractor 和 processor对数据集进行预处理以进行微调。"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}