Loading Text Dataset

Ascend GPU CPU Data Preparation

  

Overview

The mindspore.dataset module provided by MindSpore enables users to customize their data fetching strategy from disk. At the same time, data processing and tokenization operators are applied to the data. Pipelined data processing produces a continuous flow of data to the training network, improving overall performance.

In addition, MindSpore supports data loading in distributed scenarios. Users can define the number of shards while loading. For more details, see Loading the Dataset in Data Parallel Mode.

This tutorial briefly demonstrates how to load and process text data using MindSpore.

Preparations

1. Prepare the following text data.

Welcome to Beijing!

北京欢迎您！

我喜欢English!

2. Create the tokenizer.txt file, copy the text data to the file, and save the file under ./datasets directory. The directory structure is as follow.

import os

if not os.path.exists('./datasets'):
    os.mkdir('./datasets')
file_handle = open('./datasets/tokenizer.txt', mode='w')
file_handle.write('Welcome to Beijing \n北京欢迎您！ \n我喜欢English! \n')
file_handle.close()

The dataset structure is:

./datasets
└── tokenizer.txt

0 directories, 1 file

3. Import the mindspore.dataset and mindspore.dataset.text modules.

import mindspore.dataset as ds
import mindspore.dataset.text as text

Loading Dataset

MindSpore supports loading common datasets in the field of text processing that come in a variety of on-disk formats. Users can also implement custom dataset class to load customized data. For detailed loading methods of various datasets, please refer to the Loading Dataset chapter in the Programming Guide.

The following tutorial demonstrates loading datasets using the TextFileDataset in the mindspore.dataset module.

1. Configure the dataset directory as follows and create a dataset object.

DATA_FILE = "./datasets/tokenizer.txt"
dataset = ds.TextFileDataset(DATA_FILE, shuffle=False)

2. Create an iterator then obtain data through the iterator.

for data in dataset.create_dict_iterator(output_numpy=True):
    print(text.to_str(data['text']))

Welcome to Beijing
北京欢迎您！
我喜欢English!

Processing Data

For the data processing operators currently supported by MindSpore and their detailed usage methods, please refer to the Processing Data chapter in the Programming Guide

The following tutorial demonstrates how to construct a pipeline and perform operations such as shuffle and RegexReplace on the text dataset.

1. Shuffle the dataset.

    ds.config.set_seed(58)
    dataset = dataset.shuffle(buffer_size=3)

    for data in dataset.create_dict_iterator(output_numpy=True):
        print(text.to_str(data['text']))

我喜欢English!
Welcome to Beijing
北京欢迎您！

2. Perform RegexReplace on the dataset.

    replace_op1 = text.RegexReplace("Beijing", "Shanghai")
    replace_op2 = text.RegexReplace("北京", "上海")
    dataset = dataset.map(operations=replace_op1)
    dataset = dataset.map(operations=replace_op2)

    for data in dataset.create_dict_iterator(output_numpy=True):
        print(text.to_str(data['text']))

我喜欢English!
Welcome to Shanghai
上海欢迎您！

Tokenization

For the data tokenization operators currently supported by MindSpore and their detailed usage methods, please refer to the Tokenizer chapter in the Programming Guide.

The following tutorial demonstrates how to use the WhitespaceTokenizer to tokenize words with space.

1. Create a tokenizer.

    tokenizer = text.WhitespaceTokenizer()

2. Apply the tokenizer.

    dataset = dataset.map(operations=tokenizer)

3. Create an iterator and obtain data through the iterator.

    for data in dataset.create_dict_iterator(output_numpy=True):
        print(text.to_str(data['text']).tolist())

['我喜欢English!']
['Welcome', 'to', 'Shanghai']
['上海欢迎您！']