PaddleNLP Data API

from paddlenlp.data import Vocab
# 从文件构建
vocab1 = Vocab.load_vocabulary(vocab_file_path)
# 从字典构建
# dic = {'unk':0, 'pad':1, 'bos':2, 'eos':3, ...}
vocab2 = Vocab.from_dict(dic)
# 从json构建，一般是已构建好的Vocab对象先保存为json_str或json文件后再进行恢复
# json_str方式
json_str = vocab1.to_json()
vocab3 = Vocab.from_json(json_str)
# json文件方式
vocab1.to_json(json_file_path)
vocab4 = Vocab.from_json(json_file_path)

from paddlenlp.data import Vocab, JiebaTokenizer
# 词表文件路径，运行示例程序可先下载词表文件
# wget https://bj.bcebos.com/paddlenlp/data/senta_word_dict.txt
vocab_file_path = './senta_word_dict.txt'
# 构建词表
vocab = Vocab.load_vocabulary(
    vocab_file_path,
    unk_token='[UNK]',
    pad_token='[PAD]')
tokenizer = JiebaTokenizer(vocab)
tokens = tokenizer.cut('我爱你中国') # ['我爱你', '中国']
ids = tokenizer.encode('我爱你中国') # [1170578, 575565]

from paddlenlp.data import SamplerHelper
from paddle.io import Dataset

class MyDataset(Dataset):
    def __init__(self):
        super(MyDataset, self).__init__()
        self.data = [
            [[1, 2, 3, 4], [1]],
            [[5, 6, 7], [0]],
            [[8, 9], [1]],
        ]

    def __getitem__(self, index):
        data = self.data[index][0]
        label = self.data[index][1]
        return data, label

    def __len__(self):
        return len(self.data)

dataset = MyDataset()
# SamplerHelper返回的是数据索引的可迭代对象，产生的迭代的索引为：[0, 1, 2]
sampler = SamplerHelper(dataset)
# `shuffle()`的作用是随机打乱索引顺序，产生的迭代的索引为：[0, 2, 1]
sampler = sampler.shuffle()
# sort()的作用是按照指定key为排序方式并在buffer_size大小个样本中排序
# 示例中以样本第一个字段的长度进行升序排序，产生的迭代的索引为：[2, 0, 1]
key = (lambda x, data_source: len(data_source[x][0]))
sampler = sampler.sort(key=key, buffer_size=2)
# batch()的作用是按照batch_size组建mini-batch，产生的迭代的索引为：[[2, 0], [1]]
sampler = sampler.batch(batch_size=2)
# shard()的作用是为多卡训练切分数据集，当前卡产生的迭代的索引为：[[2, 0]]
sampler = sampler.shard(num_replicas=2)

from paddlenlp.data import Stack
a = [1, 2, 3, 4]
b = [3, 4, 5, 6]
c = [5, 6, 7, 8]
result = Stack()([a, b, c])
"""
[[1, 2, 3, 4],
 [3, 4, 5, 6],
 [5, 6, 7, 8]]
"""

from paddlenlp.data import Pad
a = [1, 2, 3, 4]
b = [5, 6, 7]
c = [8, 9]
result = Pad(pad_val=0)([a, b, c])
"""
[[1, 2, 3, 4],
 [5, 6, 7, 0],
 [8, 9, 0, 0]]
"""

from paddlenlp.data import Stack, Pad, Tuple
data = [
        [[1, 2, 3, 4], [1]],
        [[5, 6, 7], [0]],
        [[8, 9], [1]],
       ]
batchify_fn = Tuple(Pad(pad_val=0), Stack())
ids, label = batchify_fn(data)
"""
ids:
[[1, 2, 3, 4],
 [5, 6, 7, 0],
 [8, 9, 0, 0]]
label: [[1], [0], [1]]
"""

from paddlenlp.data import Stack, Pad, Dict
data = [
        {'labels':[1], 'token_ids':[1, 2, 3, 4]},
        {'labels':[0], 'token_ids':[5, 6, 7]},
        {'labels':[1], 'token_ids':[8, 9]},
       ]
batchify_fn = Dict({'token_ids':Pad(pad_val=0), 'labels':Stack()})
ids, label = batchify_fn(data)
"""
ids:
[[1, 2, 3, 4],
 [5, 6, 7, 0],
 [8, 9, 0, 0]]
label: [[1], [0], [1]]
"""

from paddlenlp.data import Vocab, JiebaTokenizer, Stack, Pad, Tuple, SamplerHelper
from paddlenlp.datasets import load_dataset
from paddle.io import DataLoader

# 词表文件路径，运行示例程序可先下载词表文件
# wget https://bj.bcebos.com/paddlenlp/data/senta_word_dict.txt
vocab_file_path = './senta_word_dict.txt'
# 构建词表
vocab = Vocab.load_vocabulary(
    vocab_file_path,
    unk_token='[UNK]',
    pad_token='[PAD]')
# 初始化分词器
tokenizer = JiebaTokenizer(vocab)

def convert_example(example):
    text, label = example['text'], example['label']
    ids = tokenizer.encode(text)
    label = [label]
    return ids, label

dataset = load_dataset('chnsenticorp', splits='train')
dataset = dataset.map(convert_example, lazy=True)

pad_id = vocab.token_to_idx[vocab.pad_token]
batchify_fn = Tuple(
    Pad(axis=0, pad_val=pad_id),  # ids
    Stack(dtype='int64')  # label
)

batch_sampler = SamplerHelper(dataset).shuffle().batch(batch_size=16)
data_loader = DataLoader(
    dataset,
    batch_sampler=batch_sampler,
    collate_fn=batchify_fn,
    return_list=True)

# 测试数据集
for batch in data_loader:
    ids, label = batch
    print(ids.shape, label.shape)
    print(ids)
    print(label)
    break