【PaddleNLP】使用预训练模型代码解读（Stack，Pad，Tuple、utils.py，偏函数)_paddlenlp utils

总结

utils.py里面的函数
都是封装了一下，
感觉没必要，

不清楚变量的shape和值就一个一个print输出
多看官方文档和里面的例子

【参考：模型训练、评估与推理-使用文档-PaddlePaddle深度学习平台】

---

项目一

本项目来源于：【参考：『NLP经典项目集』02：使用预训练模型ERNIE优化情感分析 - 飞桨AI Studio】

Stack，Pad，Tuple

• 【参考：使用PaddleNLP进行恶意网页识别（一） - 飞桨AI Studio】
• 【参考：PaddleNLP/data.md at develop · PaddlePaddle/PaddleNLP】

注意，在早前的PaddleNLP版本中，token_type_ids叫做segment_ids

# 单句输入
single_seg_input = tokenizer(text="请输入测试样例")
# 句对输入
multi_seg_input = tokenizer(text="请输入测试样例1", text_pair="请输入测试样例2")

print("单句输入token (str): {}".format(tokenizer.convert_ids_to_tokens(single_seg_input['input_ids'])))
print("单句输入token (int): {}".format(single_seg_input['input_ids']))
# 注意，在早前的PaddleNLP版本中，token_type_ids叫做segment_ids
print("单句输入segment ids : {}".format(single_seg_input['token_type_ids']))

print()
print("句对输入token (str): {}".format(tokenizer.convert_ids_to_tokens(multi_seg_input['input_ids'])))
print("句对输入token (int): {}".format(multi_seg_input['input_ids']))
print("句对输入segment ids : {}".format(multi_seg_input['token_type_ids']))
1
2
3
4
5
6
7
8
9
10
11
12
13
14

Dataset

【参考：dataset — PaddleNLP 文档】

from functools import partial
from paddlenlp.data import Stack, Tuple, Pad
from utils import  convert_example, create_dataloader

# 模型运行批处理大小
batch_size = 32
max_seq_length = 128

#使用partial()来固定convert_example函数的tokenizer, label_list, max_seq_length, is_test等参数值
trans_func = partial(
    convert_example,
    tokenizer=tokenizer,
    max_seq_length=max_seq_length)

# 对获取的批次数据进行处理
batchify_fn = lambda samples, fn=Tuple(
	# 这里要和返回的data 里面的数据相对应，data里面有三个值，下面就写三个
	# axis=0 按行堆叠 pad_val：填充值
    Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input
    Pad(axis=0, pad_val=tokenizer.pad_token_type_id),  # segment
    Stack(dtype="int64")  # label
): [data for data in fn(samples)] # samples每个batch的数据，类型List[List[]]，data类型为 List[]

train_data_loader = create_dataloader(
    train_ds, # MapDataset
    mode='train',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)
dev_data_loader = create_dataloader(
    dev_ds,
    mode='dev',
    batch_size=batch_size,
    batchify_fn=batchify_fn,
    trans_fn=trans_func)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

【参考：“此苹果非彼苹果”看意图识别的那些事儿】
3.调用map()方法批量处理数据
由于我们传入了lazy=False，所以我们使用load_dataset()自定义的数据集是MapDataset对象。
MapDataset是paddle.io.Dataset的功能增强版本。其内置的map()方法适合用来进行批量数据集处理。
map()方法传入的是一个用于数据处理的function。正好可以与tokenizer相配合。
4.Batchify和数据读入
使用paddle.io.BatchSampler和paddlenlp.data中提供的方法把数据组成batch。
然后使用paddle.io.DataLoader接口多线程异步加载数据。
Batchify功能详解：

utils.py

import numpy as np
import paddle
import paddle.nn.functional as F
from paddlenlp.data import Stack, Tuple, Pad


def predict(model, data, tokenizer, label_map, batch_size=1):
    """
    Predicts the data labels.

    Args:
        model (obj:`paddle.nn.Layer`): A model to classify texts.
        data (obj:`List(Example)`): The processed data whose each element is a Example (numedtuple) object.
            A Example object contains `text`(word_ids) and `seq_len`(sequence length).
        tokenizer(obj:`PretrainedTokenizer`): This tokenizer inherits from :class:`~paddlenlp.transformers.PretrainedTokenizer` 
            which contains most of the methods. Users should refer to the superclass for more information regarding methods.
        label_map(obj:`dict`): The label id (key) to label str (value) map.
        batch_size(obj:`int`, defaults to 1): The number of batch.

    Returns:
        results(obj:`dict`): All the predictions labels.
    """
    examples = []
    for text in data:
        input_ids, segment_ids = convert_example(
            text,
            tokenizer,
            max_seq_length=128,
            is_test=True)
        examples.append((input_ids, segment_ids))

    batchify_fn = lambda samples, fn=Tuple(
        Pad(axis=0, pad_val=tokenizer.pad_token_id),  # input id
        Pad(axis=0, pad_val=tokenizer.pad_token_id),  # segment id
    ): fn(samples)

    # Seperates data into some batches.
    batches = []
    one_batch = []
    for example in examples:
        one_batch.append(example)
        if len(one_batch) == batch_size:
            batches.append(one_batch)
            one_batch = []
    if one_batch:
        # The last batch whose size is less than the config batch_size setting.
        batches.append(one_batch)

    results = []
    model.eval()
    for batch in batches:
        input_ids, segment_ids = batchify_fn(batch)
        input_ids = paddle.to_tensor(input_ids)
        segment_ids = paddle.to_tensor(segment_ids)
        logits = model(input_ids, segment_ids)
        probs = F.softmax(logits, axis=1)
        idx = paddle.argmax(probs, axis=1).numpy()
        idx = idx.tolist()
        labels = [label_map[i] for i in idx]
        results.extend(labels)
    return results


@paddle.no_grad()
def evaluate(model, criterion, metric, data_loader):
    """
    Given a dataset, it evals model and computes the metric.

    Args:
        model(obj:`paddle.nn.Layer`): A model to classify texts.
        data_loader(obj:`paddle.io.DataLoader`): The dataset loader which generates batches.
        criterion(obj:`paddle.nn.Layer`): It can compute the loss.
        metric(obj:`paddle.metric.Metric`): The evaluation metric.
    """
    model.eval()
    metric.reset() # 重置
    losses = []
    for batch in data_loader:
        input_ids, token_type_ids, labels = batch
        logits = model(input_ids, token_type_ids)
        loss = criterion(logits, labels)
        losses.append(loss.numpy()) 
        correct = metric.compute(logits, labels)
        metric.update(correct)
        accu = metric.accumulate()
    print("eval loss: %.5f, accu: %.5f" % (np.mean(losses), accu))
    model.train() # 继续训练
    metric.reset()


def convert_example(example, tokenizer, max_seq_length=512, is_test=False):
    """
    Builds model inputs from a sequence or a pair of sequence for sequence classification tasks
    by concatenating and adding special tokens. And creates a mask from the two sequences passed 
    to be used in a sequence-pair classification task.
        
    A BERT sequence has the following format:

    - single sequence: ``[CLS] X [SEP]``
    - pair of sequences: ``[CLS] A [SEP] B [SEP]``

    A BERT sequence pair mask has the following format:
    ::
        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
        | first sequence    | second sequence |

    If only one sequence, only returns the first portion of the mask (0's).


    Args:
        example(obj:`list[str]`): List of input data, containing text and label if it have label.
        tokenizer(obj:`PretrainedTokenizer`): This tokenizer inherits from :class:`~paddlenlp.transformers.PretrainedTokenizer` 
            which contains most of the methods. Users should refer to the superclass for more information regarding methods.
        max_seq_len(obj:`int`): The maximum total input sequence length after tokenization. 
            Sequences longer than this will be truncated, sequences shorter will be padded.
        is_test(obj:`False`, defaults to `False`): Whether the example contains label or not.

    Returns:
        input_ids(obj:`list[int]`): The list of token ids.
        token_type_ids(obj: `list[int]`): List of sequence pair mask.
        label(obj:`numpy.array`, data type of int64, optional): The input label if not is_test.
    """
    encoded_inputs = tokenizer(text=example["text"], max_seq_len=max_seq_length)
    input_ids = encoded_inputs["input_ids"]
    token_type_ids = encoded_inputs["token_type_ids"]

    if not is_test:
        label = np.array([example["label"]], dtype="int64")
        return input_ids, token_type_ids, label
    else:
        return input_ids, token_type_ids


def create_dataloader(dataset,
                      mode='train',
                      batch_size=1,
                      batchify_fn=None,
                      trans_fn=None):
    if trans_fn:
        # 相当于给convert_example函数传入example=dataset（dataset:paddlenlp.datasets.dataset.MapDataset, dataset.data:List[str]）
        # 【参考：[dataset — PaddleNLP 文档](https://paddlenlp.readthedocs.io/zh/latest/source/paddlenlp.datasets.dataset.html)】
        # MapDataset下有map函数，默认给fn传入一个single sample
        dataset = dataset.map(trans_fn) 
        

    shuffle = True if mode == 'train' else False #如果不是训练集，则不打乱顺序
    if mode == 'train':
        # DistributedBatchSampler 分布式批采样器加载数据
        batch_sampler = paddle.io.DistributedBatchSampler(
            dataset, batch_size=batch_size, shuffle=shuffle)
    else:
        # 生成一个取样器
        batch_sampler = paddle.io.BatchSampler(
            dataset, batch_size=batch_size, shuffle=shuffle)

    return paddle.io.DataLoader(
        dataset=dataset,
        batch_sampler=batch_sampler, # 批次采样
        collate_fn=batchify_fn, # 对抽取的批次数据进行处理
        return_list=True) # 返回List


1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162

项目二

【参考：『NLP经典项目集』01：seq2vec是什么? 瞧瞧怎么用它做情感分析_副本 - 人工智能学习与实训社区】

使用高阶API训练

# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import defaultdict
import re

from paddlenlp import Taskflow
import numpy as np
import paddle

word_segmenter = Taskflow("word_segmentation") # 分词器


def create_dataloader(dataset,
                      trans_fn=None,
                      mode='train',
                      batch_size=1,
                      batchify_fn=None):
    """
    Creats dataloader.

    Args:
        dataset(obj:`paddle.io.Dataset`): Dataset instance.
        trans_fn(obj:`callable`, optional, defaults to `None`): function to convert a data sample to input ids, etc.
        mode(obj:`str`, optional, defaults to obj:`train`): If mode is 'train', it will shuffle the dataset randomly.
        batch_size(obj:`int`, optional, defaults to 1): The sample number of a mini-batch.
        batchify_fn(obj:`callable`, optional, defaults to `None`): function to generate mini-batch data by merging
            the sample list, None for only stack each fields of sample in axis
            0(same as :attr::`np.stack(..., axis=0)`).

    Returns:
        dataloader(obj:`paddle.io.DataLoader`): The dataloader which generates batches.
    """
    # 相当于给convert_example函数传入example=dataset（dataset:paddlenlp.datasets.dataset.MapDataset, dataset.data:List[str]）
    # 【参考：[dataset — PaddleNLP 文档](https://paddlenlp.readthedocs.io/zh/latest/source/paddlenlp.datasets.dataset.html)】
    # MapDataset下有map函数，默认给fn传入一个single sample
    if trans_fn:
        dataset = dataset.map(trans_fn)

    shuffle = True if mode == 'train' else False # 如果不是训练集，则不打乱顺序
    if mode == "train":
        # DistributedBatchSampler 分布式批采样器加载数据
        sampler = paddle.io.DistributedBatchSampler(
            dataset=dataset, batch_size=batch_size, shuffle=shuffle)
    else:
        # 生成一个取样器
        sampler = paddle.io.BatchSampler(
            dataset=dataset, batch_size=batch_size, shuffle=shuffle)

    dataloader = paddle.io.DataLoader(
        dataset,
        batch_sampler=sampler, # 批次采样
        collate_fn=batchify_fn) # 对抽取的批次数据进行处理

    return dataloader

# 数据转换 str->int
def convert_example(example, tokenizer, is_test=False):
    """
    Builds model inputs from a sequence for sequence classification tasks. 
    It use `jieba.cut` to tokenize text.

    Args:
        example(obj:`list[str]`): List of input data, containing text and label if it have label.
        tokenizer(obj: paddlenlp.data.JiebaTokenizer): It use jieba to cut the chinese string.
        is_test(obj:`False`, defaults to `False`): Whether the example contains label or not.

    Returns:
        input_ids(obj:`list[int]`): The list of token ids.
        valid_length(obj:`int`): The input sequence valid length.
        label(obj:`numpy.array`, data type of int64, optional): The input label if not is_test.
    """

    input_ids = tokenizer.encode(example["text"])
    valid_length = np.array(len(input_ids), dtype='int64')
    input_ids = np.array(input_ids, dtype='int64')

    if not is_test:
        label = np.array(example["label"], dtype="int64")
        return input_ids, valid_length, label
    else:
        return input_ids, valid_length


def preprocess_prediction_data(data, tokenizer):
    """
    It process the prediction data as the format used as training.

    Args:
        data (obj:`List[str]`): The prediction data whose each element is  a tokenized text.
        tokenizer(obj: paddlenlp.data.JiebaTokenizer): It use jieba to cut the chinese string.

    Returns:
        examples (obj:`List(Example)`): The processed data whose each element is a Example (numedtuple) object.
            A Example object contains `text`(word_ids) and `seq_len`(sequence length).

    """
    examples = []
    for text in data:
        ids = tokenizer.encode(text)
        examples.append([ids, len(ids)])
    return examples


# 构建词汇表
def build_vocab(texts,
                stopwords=[],
                num_words=None,
                min_freq=10,
                unk_token="[UNK]",
                pad_token="[PAD]"):
    """
    According to the texts, it is to build vocabulary.

    Args:
        texts (obj:`List[str]`): The raw corpus data.
        num_words (obj:`int`): the maximum size of vocabulary.
        stopwords (obj:`List[str]`): The list where each element is a word that will be
            filtered from the texts.
        min_freq (obj:`int`): the minimum word frequency of words to be kept.
        unk_token (obj:`str`): Special token for unknow token.
        pad_token (obj:`str`): Special token for padding token.

    Returns:
        word_index (obj:`Dict`): The vocabulary from the corpus data.

    """
    word_counts = defaultdict(int) # defaultdict默认找不到键时返回默认值，int对应0
    for text in texts:
        if not text: # text为None时
            continue
        for word in word_segmenter(text): #  word_segmenter(text) 返回一个分词后的列表
            if word in stopwords: # 去停用词
                continue
            word_counts[word] += 1

    wcounts = []
    for word, count in word_counts.items():
        if count < min_freq: # 词的频率小于min_freq
            continue
        wcounts.append((word, count))
    wcounts.sort(key=lambda x: x[1], reverse=True) # 按照词的频率递减排序（高频排前面）
    # -2 for the pad_token and unk_token which will be added to vocab.
    if num_words is not None and len(wcounts) > (num_words - 2):
        wcounts = wcounts[:(num_words - 2)] # 取频率最大的前num_words - 2个
    # add the special pad_token and unk_token to the vocabulary 
    sorted_voc = [pad_token, unk_token]
    sorted_voc.extend(wc[0] for wc in wcounts) # 获取词典中所有的词
    # list(range(len(sorted_voc)))) 根据sorted_voc的长度生成[1,...len]的序列，序列的值即为词的编号
    # zip返回一个tuple，然后依据这个tuple生成字典
    word_index = dict(zip(sorted_voc, list(range(len(sorted_voc)))))
    return word_index

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163