{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 使用Loader和Pipe加载并处理数据集\n", "\n", "这一部分是关于如何加载数据集的教程\n", "\n", "## Part I: 数据集容器DataBundle\n", "\n", "而由于对于同一个任务，训练集，验证集和测试集会共用同一个词表以及具有相同的目标值，所以在fastNLP中我们使用了 DataBundle 来承载同一个任务的多个数据集 DataSet 以及它们的词表 Vocabulary 。下面会有例子介绍 DataBundle 的相关使用。\n", "\n", "DataBundle 在fastNLP中主要在各个 Loader 和 Pipe 中被使用。 下面我们先介绍一下 Loader 和 Pipe 。\n", "\n", "## Part II: 加载的各种数据集的Loader\n", "\n", "在fastNLP中，所有的 Loader 都可以通过其文档判断其支持读取的数据格式，以及读取之后返回的 DataSet 的格式, 例如 ChnSentiCorpLoader \n", "\n", "- download() 函数：自动将该数据集下载到缓存地址，默认缓存地址为~/.fastNLP/datasets/。由于版权等原因，不是所有的Loader都实现了该方法。该方法会返回下载后文件所处的缓存地址。\n", "\n", "- _load() 函数：从一个数据文件中读取数据，返回一个 DataSet 。返回的DataSet的格式可从Loader文档判断。\n", "\n", "- load() 函数：从文件或者文件夹中读取数据为 DataSet 并将它们组装成 DataBundle。支持接受的参数类型有以下的几种\n", "\n", " - None, 将尝试读取自动缓存的数据，仅支持提供了自动下载数据的Loader\n", " - 文件夹路径, 默认将尝试在该文件夹下匹配文件名中含有 train , test , dev 的文件，如果有多个文件含有相同的关键字，将无法通过该方式读取\n", " - dict, 例如{'train':\"/path/to/tr.conll\", 'dev':\"/to/validate.conll\", \"test\":\"/to/te.conll\"}。" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "In total 3 datasets:\n", "\ttest has 1944 instances.\n", "\ttrain has 17196 instances.\n", "\tdev has 1858 instances.\n", "\n" ] } ], "source": [ "from fastNLP.io import CWSLoader\n", "\n", "loader = CWSLoader(dataset_name='pku')\n", "data_bundle = loader.load()\n", "print(data_bundle)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "这里表示一共有3个数据集。其中：\n", "\n", " 3个数据集的名称分别为train、dev、test，分别有17223、1831、1944个instance\n", "\n", "也可以取出DataSet，并打印DataSet中的具体内容" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "+----------------------------------------------------------------+\n", "| raw_words |\n", "+----------------------------------------------------------------+\n", "| 迈向 充满 希望 的 新 世纪 —— 一九九八年 新年 讲话 ... |\n", "| 中共中央 总书记 、 国家 主席 江 泽民 |\n", "+----------------------------------------------------------------+\n" ] } ], "source": [ "tr_data = data_bundle.get_dataset('train')\n", "print(tr_data[:2])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Part III: 使用Pipe对数据集进行预处理\n", "\n", "通过 Loader 可以将文本数据读入，但并不能直接被神经网络使用，还需要进行一定的预处理。\n", "\n", "在fastNLP中，我们使用 Pipe 的子类作为数据预处理的类， Loader 和 Pipe 一般具备一一对应的关系，该关系可以从其名称判断， 例如 CWSLoader 与 CWSPipe 是一一对应的。一般情况下Pipe处理包含以下的几个过程，\n", "1. 将raw_words或 raw_chars进行tokenize以切分成不同的词或字; \n", "2. 再建立词或字的 Vocabulary , 并将词或字转换为index; \n", "3. 将target 列建立词表并将target列转为index;\n", "\n", "所有的Pipe都可通过其文档查看该Pipe支持处理的 DataSet 以及返回的 DataBundle 中的Vocabulary的情况; 如 OntoNotesNERPipe\n", "\n", "各种数据集的Pipe当中，都包含了以下的两个函数:\n", "\n", "- process() 函数：对输入的 DataBundle 进行处理, 然后返回处理之后的 DataBundle 。process函数的文档中包含了该Pipe支持处理的DataSet的格式。\n", "- process_from_file() 函数：输入数据集所在文件夹，使用对应的Loader读取数据(所以该函数支持的参数类型是由于其对应的Loader的load函数决定的)，然后调用相对应的process函数对数据进行预处理。相当于是把Load和process放在一个函数中执行。\n", "\n", "接着上面 CWSLoader 的例子，我们展示一下 CWSPipe 的功能：" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "In total 3 datasets:\n", "\ttest has 1944 instances.\n", "\ttrain has 17196 instances.\n", "\tdev has 1858 instances.\n", "In total 2 vocabs:\n", "\tchars has 4777 entries.\n", "\ttarget has 4 entries.\n", "\n" ] } ], "source": [ "from fastNLP.io import CWSPipe\n", "\n", "data_bundle = CWSPipe().process(data_bundle)\n", "print(data_bundle)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "表示一共有3个数据集和2个词表。其中：\n", "\n", "- 3个数据集的名称分别为train、dev、test，分别有17223、1831、1944个instance\n", "- 2个词表分别为chars词表与target词表。其中chars词表为句子文本所构建的词表，一共有4777个不同的字；target词表为目标标签所构建的词表，一共有4种标签。\n", "\n", "相较于之前CWSLoader读取的DataBundle，新增了两个Vocabulary。 我们可以打印一下处理之后的DataSet" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "+---------------------+---------------------+---------------------+---------+\n", "| raw_words | chars | target | seq_len |\n", "+---------------------+---------------------+---------------------+---------+\n", "| 迈向 充满 希望... | [1224, 178, 674,... | [0, 1, 0, 1, 0, ... | 29 |\n", "| 中共中央 总书记... | [11, 212, 11, 33... | [0, 3, 3, 1, 0, ... | 15 |\n", "+---------------------+---------------------+---------------------+---------+\n" ] } ], "source": [ "tr_data = data_bundle.get_dataset('train')\n", "print(tr_data[:2])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "可以看到有两列为int的field: chars和target。这两列的名称同时也是DataBundle中的Vocabulary的名称。可以通过下列的代码获取并查看Vocabulary的 信息" ] }, { "cell_type": "code", "execution_count": 5, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Vocabulary(['B', 'E', 'S', 'M']...)\n" ] } ], "source": [ "vocab = data_bundle.get_vocab('target')\n", "print(vocab)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Part IV: fastNLP封装好的Loader和Pipe\n", "\n", "fastNLP封装了多种任务/数据集的 Loader 和 Pipe 并提供自动下载功能，具体参见文档 [数据集](https://docs.qq.com/sheet/DVnpkTnF6VW9UeXdh?c=A1A0A0)\n", "\n", "## Part V: 不同格式类型的基础Loader\n", "\n", "除了上面提到的针对具体任务的Loader，我们还提供了CSV格式和JSON格式的Loader\n", "\n", "**CSVLoader** 读取CSV类型的数据集文件。例子如下：\n", "\n", "```python\n", "from fastNLP.io.loader import CSVLoader\n", "data_set_loader = CSVLoader(\n", " headers=('raw_words', 'target'), sep='\\t'\n", ")\n", "```\n", "\n", "表示将CSV文件中每一行的第一项将填入'raw_words' field，第二项填入'target' field。其中项之间由'\\t'分割开来\n", "\n", "```python\n", "data_set = data_set_loader._load('path/to/your/file')\n", "```\n", "\n", "文件内容样例如下\n", "\n", "```csv\n", "But it does not leave you with much . 1\n", "You could hate it for the same reason . 1\n", "The performances are an absolute joy . 4\n", "```\n", "\n", "读取之后的DataSet具有以下的field\n", "\n", "| raw_words | target |\n", "| --------------------------------------- | ------ |\n", "| But it does not leave you with much . | 1 |\n", "| You could hate it for the same reason . | 1 |\n", "| The performances are an absolute joy . | 4 |\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "**JsonLoader** 读取Json类型的数据集文件，数据必须按行存储，每行是一个包含各类属性的Json对象。例子如下\n", "\n", "```python\n", "from fastNLP.io.loader import JsonLoader\n", "loader = JsonLoader(\n", " fields={'sentence1': 'raw_words1', 'sentence2': 'raw_words2', 'gold_label': 'target'}\n", ")\n", "```\n", "\n", "表示将Json对象中'sentence1'、'sentence2'和'gold_label'对应的值赋给'raw_words1'、'raw_words2'、'target'这三个fields\n", "\n", "```python\n", "data_set = loader._load('path/to/your/file')\n", "```\n", "\n", "数据集内容样例如下\n", "```\n", "{\"annotator_labels\": [\"neutral\"], \"captionID\": \"3416050480.jpg#4\", \"gold_label\": \"neutral\", ... }\n", "{\"annotator_labels\": [\"contradiction\"], \"captionID\": \"3416050480.jpg#4\", \"gold_label\": \"contradiction\", ... }\n", "{\"annotator_labels\": [\"entailment\"], \"captionID\": \"3416050480.jpg#4\", \"gold_label\": \"entailment\", ... }\n", "```\n", "\n", "读取之后的DataSet具有以下的field\n", "\n", "| raw_words0 | raw_words1 | target |\n", "| ------------------------------------------------------ | ------------------------------------------------- | ------------- |\n", "| A person on a horse jumps over a broken down airplane. | A person is training his horse for a competition. | neutral |\n", "| A person on a horse jumps over a broken down airplane. | A person is at a diner, ordering an omelette. | contradiction |\n", "| A person on a horse jumps over a broken down airplane. | A person is outdoors, on a horse. | entailment |" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python Now", "language": "python", "name": "now" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.8.0" } }, "nbformat": 4, "nbformat_minor": 2 }