transformers_doc/zh/tensorflow/preprocessing.ipynb (1,078 lines of code) (raw):
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# 预处理"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"在您可以在数据集上训练模型之前，数据需要被预处理为期望的模型输入格式。无论您的数据是文本、图像还是音频，它们都需要被转换并组合成批量的张量。🤗 Transformers 提供了一组预处理类来帮助准备数据以供模型使用。在本教程中，您将了解以下内容：\n",
"\n",
"* 对于文本，使用[分词器](https://huggingface.co/docs/transformers/main/zh/./main_classes/tokenizer)(`Tokenizer`)将文本转换为一系列标记(`tokens`)，并创建`tokens`的数字表示，将它们组合成张量。\n",
"* 对于语音和音频，使用[特征提取器](https://huggingface.co/docs/transformers/main/zh/./main_classes/feature_extractor)(`Feature extractor`)从音频波形中提取顺序特征并将其转换为张量。\n",
"* 图像输入使用[图像处理器](https://huggingface.co/docs/transformers/main/zh/./main_classes/image)(`ImageProcessor`)将图像转换为张量。\n",
"* 多模态输入，使用[处理器](https://huggingface.co/docs/transformers/main/zh/./main_classes/processors)(`Processor`)结合了`Tokenizer`和`ImageProcessor`或`Processor`。\n",
"\n",
"<Tip>\n",
"\n",
"`AutoProcessor` **始终**有效的自动选择适用于您使用的模型的正确`class`，无论您使用的是`Tokenizer`、`ImageProcessor`、`Feature extractor`还是`Processor`。\n",
"\n",
"</Tip>\n",
"\n",
"在开始之前，请安装🤗 Datasets，以便您可以加载一些数据集来进行实验：\n",
"\n",
"\n",
"```bash\n",
"pip install datasets\n",
"```"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 自然语言处理"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"cellView": "form",
"hide_input": true
},
"outputs": [
{
"data": {
"text/html": [
"<iframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/Yffk5aydLzg?rel=0&controls=0&showinfo=0\" frameborder=\"0\" allowfullscreen></iframe>"
],
"text/plain": [
"<IPython.core.display.HTML object>"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"#@title\n",
"from IPython.display import HTML\n",
"\n",
"HTML('<iframe width=\"560\" height=\"315\" src=\"https://www.youtube.com/embed/Yffk5aydLzg?rel=0&controls=0&showinfo=0\" frameborder=\"0\" allowfullscreen></iframe>')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"处理文本数据的主要工具是[Tokenizer](https://huggingface.co/docs/transformers/main/zh/main_classes/tokenizer)。`Tokenizer`根据一组规则将文本拆分为`tokens`。然后将这些`tokens`转换为数字，然后转换为张量，成为模型的输入。模型所需的任何附加输入都由`Tokenizer`添加。\n",
"\n",
"<Tip>\n",
"\n",
"如果您计划使用预训练模型，重要的是使用与之关联的预训练`Tokenizer`。这确保文本的拆分方式与预训练语料库相同，并在预训练期间使用相同的标记-索引的对应关系（通常称为*词汇表*-`vocab`）。\n",
"\n",
"</Tip>\n",
"\n",
"开始使用`AutoTokenizer.from_pretrained()`方法加载一个预训练`tokenizer`。这将下载模型预训练的`vocab`："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from transformers import AutoTokenizer\n",
"\n",
"tokenizer = AutoTokenizer.from_pretrained(\"google-bert/bert-base-cased\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"然后将您的文本传递给`tokenizer`："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'input_ids': [101, 2079, 2025, 19960, 10362, 1999, 1996, 3821, 1997, 16657, 1010, 2005, 2027, 2024, 11259, 1998, 4248, 2000, 4963, 1012, 102],\n",
" 'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"encoded_input = tokenizer(\"Do not meddle in the affairs of wizards, for they are subtle and quick to anger.\")\n",
"print(encoded_input)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"`tokenizer`返回一个包含三个重要对象的字典：\n",
"\n",
"* [input_ids](https://huggingface.co/docs/transformers/main/zh/glossary#input-ids) 是与句子中每个`token`对应的索引。\n",
"* [attention_mask](https://huggingface.co/docs/transformers/main/zh/glossary#attention-mask) 指示是否应该关注一个`token`。\n",
"* [token_type_ids](https://huggingface.co/docs/transformers/main/zh/glossary#token-type-ids) 在存在多个序列时标识一个`token`属于哪个序列。\n",
"\n",
"通过解码 `input_ids` 来返回您的输入："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'[CLS] Do not meddle in the affairs of wizards, for they are subtle and quick to anger. [SEP]'"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"tokenizer.decode(encoded_input[\"input_ids\"])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"如您所见，`tokenizer`向句子中添加了两个特殊`token` - `CLS` 和 `SEP`（分类器和分隔符）。并非所有模型都需要特殊`token`，但如果需要，`tokenizer`会自动为您添加。\n",
"\n",
"如果有多个句子需要预处理，将它们作为列表传递给`tokenizer`："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102],\n",
" [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],\n",
" [101, 1327, 1164, 5450, 23434, 136, 102]],\n",
" 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0]],\n",
" 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1],\n",
" [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],\n",
" [1, 1, 1, 1, 1, 1, 1]]}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"batch_sentences = [\n",
" \"But what about second breakfast?\",\n",
" \"Don't think he knows about second breakfast, Pip.\",\n",
" \"What about elevensies?\",\n",
"]\n",
"encoded_inputs = tokenizer(batch_sentences)\n",
"print(encoded_inputs)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 填充"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"句子的长度并不总是相同，这可能会成为一个问题，因为模型输入的张量需要具有统一的形状。填充是一种策略，通过在较短的句子中添加一个特殊的`padding token`，以确保张量是矩形的。\n",
"\n",
"将 `padding` 参数设置为 `True`，以使批次中较短的序列填充到与最长序列相匹配的长度："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],\n",
" [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],\n",
" [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
" 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
" 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],\n",
" [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],\n",
" [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"batch_sentences = [\n",
" \"But what about second breakfast?\",\n",
" \"Don't think he knows about second breakfast, Pip.\",\n",
" \"What about elevensies?\",\n",
"]\n",
"encoded_input = tokenizer(batch_sentences, padding=True)\n",
"print(encoded_input)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"第一句和第三句因为较短，通过`0`进行填充，。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 截断"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"另一方面，有时候一个序列可能对模型来说太长了。在这种情况下，您需要将序列截断为更短的长度。\n",
"\n",
"将 `truncation` 参数设置为 `True`，以将序列截断为模型接受的最大长度："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'input_ids': [[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],\n",
" [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],\n",
" [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
" 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
" 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],\n",
" [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],\n",
" [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]]}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"batch_sentences = [\n",
" \"But what about second breakfast?\",\n",
" \"Don't think he knows about second breakfast, Pip.\",\n",
" \"What about elevensies?\",\n",
"]\n",
"encoded_input = tokenizer(batch_sentences, padding=True, truncation=True)\n",
"print(encoded_input)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<Tip>\n",
"\n",
"查看[填充和截断](https://huggingface.co/docs/transformers/main/zh/./pad_truncation)概念指南，了解更多有关填充和截断参数的信息。\n",
"\n",
"</Tip>"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 构建张量"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"最后，`tokenizer`可以返回实际输入到模型的张量。\n",
"\n",
"将 `return_tensors` 参数设置为 `pt`（对于PyTorch）或 `tf`（对于TensorFlow）："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'input_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=\n",
"array([[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],\n",
" [101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],\n",
" [101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],\n",
" dtype=int32)>,\n",
" 'token_type_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=\n",
"array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],\n",
" [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>,\n",
" 'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=\n",
"array([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],\n",
" [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],\n",
" [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"batch_sentences = [\n",
" \"But what about second breakfast?\",\n",
" \"Don't think he knows about second breakfast, Pip.\",\n",
" \"What about elevensies?\",\n",
"]\n",
"encoded_input = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors=\"tf\")\n",
"print(encoded_input)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 音频"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"对于音频任务，您需要[feature extractor](https://huggingface.co/docs/transformers/main/zh/main_classes/feature_extractor)来准备您的数据集以供模型使用。`feature extractor`旨在从原始音频数据中提取特征，并将它们转换为张量。\n",
"\n",
"加载[MInDS-14](https://huggingface.co/datasets/PolyAI/minds14)数据集（有关如何加载数据集的更多详细信息，请参阅🤗 [Datasets教程](https://huggingface.co/docs/datasets/load_hub)）以了解如何在音频数据集中使用`feature extractor`："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from datasets import load_dataset, Audio\n",
"\n",
"dataset = load_dataset(\"PolyAI/minds14\", name=\"en-US\", split=\"train\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"访问 `audio` 列的第一个元素以查看输入。调用 `audio` 列会自动加载和重新采样音频文件："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'array': array([ 0. , 0.00024414, -0.00024414, ..., -0.00024414,\n",
" 0. , 0. ], dtype=float32),\n",
" 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',\n",
" 'sampling_rate': 8000}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"dataset[0][\"audio\"]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"这会返回三个对象：\n",
"\n",
"* `array` 是加载的语音信号 - 并在必要时重新采为`1D array`。\n",
"* `path` 指向音频文件的位置。\n",
"* `sampling_rate` 是每秒测量的语音信号数据点数量。\n",
"\n",
"对于本教程，您将使用[Wav2Vec2](https://huggingface.co/facebook/wav2vec2-base)模型。查看模型卡片，您将了解到Wav2Vec2是在16kHz采样的语音音频数据上预训练的。重要的是，您的音频数据的采样率要与用于预训练模型的数据集的采样率匹配。如果您的数据的采样率不同，那么您需要对数据进行重新采样。\n",
"\n",
"1. 使用🤗 Datasets的`cast_column`方法将采样率提升到16kHz："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"dataset = dataset.cast_column(\"audio\", Audio(sampling_rate=16_000))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"2. 再次调用 `audio` 列以重新采样音频文件："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'array': array([ 2.3443763e-05, 2.1729663e-04, 2.2145823e-04, ...,\n",
" 3.8356509e-05, -7.3497440e-06, -2.1754686e-05], dtype=float32),\n",
" 'path': '/root/.cache/huggingface/datasets/downloads/extracted/f14948e0e84be638dd7943ac36518a4cf3324e8b7aa331c5ab11541518e9368c/en-US~JOINT_ACCOUNT/602ba55abb1e6d0fbce92065.wav',\n",
" 'sampling_rate': 16000}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"dataset[0][\"audio\"]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"接下来，加载一个`feature extractor`以对输入进行标准化和填充。当填充文本数据时，会为较短的序列添加 `0`。相同的理念适用于音频数据。`feature extractor`添加 `0` - 被解释为静音 - 到`array` 。\n",
"\n",
"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(\"facebook/wav2vec2-base\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"将音频 `array` 传递给`feature extractor`。我们还建议在`feature extractor`中添加 `sampling_rate` 参数，以更好地调试可能发生的静音错误："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'input_values': [array([ 3.8106556e-04, 2.7506407e-03, 2.8015103e-03, ...,\n",
" 5.6335266e-04, 4.6588284e-06, -1.7142107e-04], dtype=float32)]}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"audio_input = [dataset[0][\"audio\"][\"array\"]]\n",
"feature_extractor(audio_input, sampling_rate=16000)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"就像`tokenizer`一样，您可以应用填充或截断来处理批次中的可变序列。请查看这两个音频样本的序列长度："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"(173398,)"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"dataset[0][\"audio\"][\"array\"].shape"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"(106496,)"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"dataset[1][\"audio\"][\"array\"].shape"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"创建一个函数来预处理数据集，以使音频样本具有相同的长度。通过指定最大样本长度，`feature extractor`将填充或截断序列以使其匹配："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def preprocess_function(examples):\n",
" audio_arrays = [x[\"array\"] for x in examples[\"audio\"]]\n",
" inputs = feature_extractor(\n",
" audio_arrays,\n",
" sampling_rate=16000,\n",
" padding=True,\n",
" max_length=100000,\n",
" truncation=True,\n",
" )\n",
" return inputs"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"将`preprocess_function`应用于数据集中的前几个示例："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"processed_dataset = preprocess_function(dataset[:5])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"现在样本长度是相同的，并且与指定的最大长度匹配。您现在可以将经过处理的数据集传递给模型了！"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"(100000,)"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"processed_dataset[\"input_values\"][0].shape"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"(100000,)"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"processed_dataset[\"input_values\"][1].shape"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 计算机视觉"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"对于计算机视觉任务，您需要一个[ image processor](https://huggingface.co/docs/transformers/main/zh/main_classes/image_processor)来准备数据集以供模型使用。图像预处理包括多个步骤将图像转换为模型期望输入的格式。这些步骤包括但不限于调整大小、标准化、颜色通道校正以及将图像转换为张量。\n",
"\n",
"<Tip>\n",
"\n",
"图像预处理通常遵循某种形式的图像增强。图像预处理和图像增强都会改变图像数据，但它们有不同的目的：\n",
"\n",
"* 图像增强可以帮助防止过拟合并增加模型的鲁棒性。您可以在数据增强方面充分发挥创造性 - 调整亮度和颜色、裁剪、旋转、调整大小、缩放等。但要注意不要改变图像的含义。\n",
"* 图像预处理确保图像与模型预期的输入格式匹配。在微调计算机视觉模型时，必须对图像进行与模型训练时相同的预处理。\n",
"\n",
"您可以使用任何您喜欢的图像增强库。对于图像预处理，请使用与模型相关联的`ImageProcessor`。\n",
"\n",
"</Tip>\n",
"\n",
"加载[food101](https://huggingface.co/datasets/food101)数据集（有关如何加载数据集的更多详细信息，请参阅🤗 [Datasets教程](https://huggingface.co/docs/datasets/load_hub)）以了解如何在计算机视觉数据集中使用图像处理器：\n",
"\n",
"<Tip>\n",
"\n",
"因为数据集相当大，请使用🤗 Datasets的`split`参数加载训练集中的少量样本！\n",
"\n",
"</Tip>"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from datasets import load_dataset\n",
"\n",
"dataset = load_dataset(\"food101\", split=\"train[:100]\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"接下来，使用🤗 Datasets的[`Image`](https://huggingface.co/docs/datasets/package_reference/main_classes?highlight=image#datasets.Image)功能查看图像："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"dataset[0][\"image\"]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<div class=\"flex justify-center\">\n",
" <img src=\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/vision-preprocess-tutorial.png\"/>\n",
"</div>\n",
"\n",
"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": [
"首先，让我们进行图像增强。您可以使用任何您喜欢的库，但在本教程中，我们将使用torchvision的[`transforms`](https://pytorch.org/vision/stable/transforms.html)模块。如果您有兴趣使用其他数据增强库，请参阅[Albumentations](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification_albumentations.ipynb)或[Kornia notebooks](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification_kornia.ipynb)中的示例。\n",
"\n",
"1. 在这里，我们使用[`Compose`](https://pytorch.org/vision/master/generated/torchvision.transforms.Compose.html)将[`RandomResizedCrop`](https://pytorch.org/vision/main/generated/torchvision.transforms.RandomResizedCrop.html)和 [`ColorJitter`](https://pytorch.org/vision/main/generated/torchvision.transforms.ColorJitter.html)变换连接在一起。请注意，对于调整大小，我们可以从`image_processor`中获取图像尺寸要求。对于一些模型，精确的高度和宽度需要被定义，对于其他模型只需定义`shortest_edge`。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from torchvision.transforms import RandomResizedCrop, ColorJitter, Compose\n",
"\n",
"size = (\n",
" image_processor.size[\"shortest_edge\"]\n",
" if \"shortest_edge\" in image_processor.size\n",
" else (image_processor.size[\"height\"], image_processor.size[\"width\"])\n",
")\n",
"\n",
"_transforms = Compose([RandomResizedCrop(size), ColorJitter(brightness=0.5, hue=0.5)])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"2. 模型接受 [`pixel_values`](https://huggingface.co/docs/transformers/main/zh/model_doc/visionencoderdecoder#transformers.VisionEncoderDecoderModel.forward.pixel_values) 作为输入。`ImageProcessor` 可以进行图像的标准化，并生成适当的张量。创建一个函数，将图像增强和图像预处理步骤组合起来处理批量图像，并生成 `pixel_values`："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def transforms(examples):\n",
" images = [_transforms(img.convert(\"RGB\")) for img in examples[\"image\"]]\n",
" examples[\"pixel_values\"] = image_processor(images, do_resize=False, return_tensors=\"pt\")[\"pixel_values\"]\n",
" return examples"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<Tip>\n",
"\n",
"在上面的示例中，我们设置`do_resize=False`，因为我们已经在图像增强转换中调整了图像的大小，并利用了适当的`image_processor`的`size`属性。如果您在图像增强期间不调整图像的大小，请将此参数排除在外。默认情况下`ImageProcessor`将处理调整大小。\n",
"\n",
"如果希望将图像标准化步骤为图像增强的一部分，请使用`image_processor.image_mean`和`image_processor.image_std`。\n",
"\n",
"</Tip>\n",
"\n",
"3. 然后使用🤗 Datasets的[`set_transform`](https://huggingface.co/docs/datasets/process#format-transform)在运行时应用这些变换："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"dataset.set_transform(transforms)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"4. 现在，当您访问图像时，您将注意到`image processor`已添加了 `pixel_values`。您现在可以将经过处理的数据集传递给模型了！"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"dataset[0].keys()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"这是在应用变换后的图像样子。图像已被随机裁剪，并其颜色属性发生了变化。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import numpy as np\n",
"import matplotlib.pyplot as plt\n",
"\n",
"img = dataset[0][\"pixel_values\"]\n",
"plt.imshow(img.permute(1, 2, 0))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<div class=\"flex justify-center\">\n",
" <img src=\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/preprocessed_image.png\"/>\n",
"</div>\n",
"\n",
"<Tip>\n",
"\n",
"对于诸如目标检测、语义分割、实例分割和全景分割等任务，`ImageProcessor`提供了训练后处理方法。这些方法将模型的原始输出转换为有意义的预测，如边界框或分割地图。\n",
"\n",
"</Tip>"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 填充"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"在某些情况下，例如，在微调[DETR](https://huggingface.co/docs/transformers/main/zh/./model_doc/detr)时，模型在训练时应用了尺度增强。这可能导致批处理中的图像大小不同。您可以使用`DetrImageProcessor.pad()`来指定自定义的`collate_fn`将图像批处理在一起。"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def collate_fn(batch):\n",
" pixel_values = [item[\"pixel_values\"] for item in batch]\n",
" encoding = image_processor.pad(pixel_values, return_tensors=\"pt\")\n",
" labels = [item[\"labels\"] for item in batch]\n",
" batch = {}\n",
" batch[\"pixel_values\"] = encoding[\"pixel_values\"]\n",
" batch[\"pixel_mask\"] = encoding[\"pixel_mask\"]\n",
" batch[\"labels\"] = labels\n",
" return batch"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 多模态"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"对于涉及多模态输入的任务，您需要[processor](https://huggingface.co/docs/transformers/main/zh/main_classes/processors)来为模型准备数据集。`processor`将两个处理对象-例如`tokenizer`和`feature extractor`-组合在一起。\n",
"\n",
"加载[LJ Speech](https://huggingface.co/datasets/lj_speech)数据集（有关如何加载数据集的更多详细信息，请参阅🤗 [Datasets 教程](https://huggingface.co/docs/datasets/load_hub)）以了解如何使用`processor`进行自动语音识别（ASR）："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from datasets import load_dataset\n",
"\n",
"lj_speech = load_dataset(\"lj_speech\", split=\"train\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"对于ASR（自动语音识别），主要关注`audio`和`text`，因此可以删除其他列："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"lj_speech = lj_speech.map(remove_columns=[\"file\", \"id\", \"normalized_text\"])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"现在查看`audio`和`text`列："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'array': array([-7.3242188e-04, -7.6293945e-04, -6.4086914e-04, ...,\n",
" 7.3242188e-04, 2.1362305e-04, 6.1035156e-05], dtype=float32),\n",
" 'path': '/root/.cache/huggingface/datasets/downloads/extracted/917ece08c95cf0c4115e45294e3cd0dee724a1165b7fc11798369308a465bd26/LJSpeech-1.1/wavs/LJ001-0001.wav',\n",
" 'sampling_rate': 22050}"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"lj_speech[0][\"audio\"]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'Printing, in the only sense with which we are at present concerned, differs from most if not from all the arts and crafts represented in the Exhibition'"
]
},
"execution_count": null,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"lj_speech[0][\"text\"]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"请记住，您应始终[重新采样](https://huggingface.co/docs/transformers/main/zh/preprocessing#audio)音频数据集的采样率，以匹配用于预训练模型数据集的采样率！"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"lj_speech = lj_speech.cast_column(\"audio\", Audio(sampling_rate=16_000))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"使用`AutoProcessor.from_pretrained()`加载一个`processor`："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from transformers import AutoProcessor\n",
"\n",
"processor = AutoProcessor.from_pretrained(\"facebook/wav2vec2-base-960h\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"1. 创建一个函数，用于将包含在 `array` 中的音频数据处理为 `input_values`，并将 `text` 标记为 `labels`。这些将是输入模型的数据："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def prepare_dataset(example):\n",
" audio = example[\"audio\"]\n",
"\n",
" example.update(processor(audio=audio[\"array\"], text=example[\"text\"], sampling_rate=16000))\n",
"\n",
" return example"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"2. 将 `prepare_dataset` 函数应用于一个示例："
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"prepare_dataset(lj_speech[0])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"`processor`现在已经添加了 `input_values` 和 `labels`，并且采样率也正确降低为为16kHz。现在可以将处理后的数据集传递给模型！"
]
}
],
"metadata": {},
"nbformat": 4,
"nbformat_minor": 4
}