paddle2.0高层API实现自定义数据集文本分类中的情感分析任务

paddle2.0高层API实现自定义数据集文本分类中的情感分析任务

本文包含了：
- 自定义文本分类数据集继承
- 文本分类数据处理
- 循环神经网络RNN, LSTM
- ·seq2vec·
- pretrained预训练模型

『深度学习7日打卡营·day4』

零基础解锁深度学习神器飞桨框架高层API，七天时间助你掌握CV、NLP领域最火模型及应用。

1. 课程地址
   传送门：https://aistudio.baidu.com/aistudio/course/introduce/6771

2. 目标

• 掌握深度学习常用模型基础知识
• 熟练掌握一种国产开源深度学习框架
• 具备独立完成相关深度学习任务的能力
• 能用所学为AI加一份年味

问题定义

情感分析是自然语言处理领域一个老生常谈的任务。句子情感分析目的是为了判别说者的情感倾向，比如在某些话题上给出的的态度明确的观点，或者反映的情绪状态等。情感分析有着广泛应用，比如电商评论分析、舆情分析等。

环境介绍

• PaddlePaddle框架，AI Studio平台已经默认安装最新版2.0。

• PaddleNLP，深度兼容框架2.0，是飞桨框架2.0在NLP领域的最佳实践。

这里使用的是beta版本，马上也会发布rc版哦。AI Studio平台后续会默认安装PaddleNLP，在此之前可使用如下命令安装。

# 下载paddlenlp
!pip install --upgrade paddlenlp==2.0.0b4 -i https://pypi.org/simple

Requirement already up-to-date: paddlenlp==2.0.0b4 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (2.0.0b4)
Requirement already satisfied, skipping upgrade: colorlog in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp==2.0.0b4) (4.1.0)
Requirement already satisfied, skipping upgrade: colorama in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp==2.0.0b4) (0.4.4)
Requirement already satisfied, skipping upgrade: jieba in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp==2.0.0b4) (0.42.1)
Requirement already satisfied, skipping upgrade: h5py in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp==2.0.0b4) (2.9.0)
Requirement already satisfied, skipping upgrade: visualdl in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp==2.0.0b4) (2.1.1)
Requirement already satisfied, skipping upgrade: seqeval in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp==2.0.0b4) (1.2.2)
Requirement already satisfied, skipping upgrade: numpy>=1.7 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from h5py->paddlenlp==2.0.0b4) (1.16.4)
Requirement already satisfied, skipping upgrade: six in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from h5py->paddlenlp==2.0.0b4) (1.15.0)
Requirement already satisfied, skipping upgrade: protobuf>=3.11.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (3.14.0)
Requirement already satisfied, skipping upgrade: Flask-Babel>=1.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (1.0.0)
Requirement already satisfied, skipping upgrade: Pillow>=7.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (7.1.2)
Requirement already satisfied, skipping upgrade: flake8>=3.7.9 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (3.8.2)
Requirement already satisfied, skipping upgrade: pre-commit in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (1.21.0)
Requirement already satisfied, skipping upgrade: shellcheck-py in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (0.7.1.1)
Requirement already satisfied, skipping upgrade: flask>=1.1.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (1.1.1)
Requirement already satisfied, skipping upgrade: requests in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (2.22.0)
Requirement already satisfied, skipping upgrade: bce-python-sdk in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp==2.0.0b4) (0.8.53)
Requirement already satisfied, skipping upgrade: scikit-learn>=0.21.3 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from seqeval->paddlenlp==2.0.0b4) (0.22.1)
Requirement already satisfied, skipping upgrade: pytz in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->paddlenlp==2.0.0b4) (2019.3)
Requirement already satisfied, skipping upgrade: Babel>=2.3 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->paddlenlp==2.0.0b4) (2.8.0)
Requirement already satisfied, skipping upgrade: Jinja2>=2.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Flask-Babel>=1.0.0->visualdl->paddlenlp==2.0.0b4) (2.10.1)
Requirement already satisfied, skipping upgrade: importlib-metadata; python_version < "3.8" in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddlenlp==2.0.0b4) (0.23)
Requirement already satisfied, skipping upgrade: mccabe<0.7.0,>=0.6.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddlenlp==2.0.0b4) (0.6.1)
Requirement already satisfied, skipping upgrade: pycodestyle<2.7.0,>=2.6.0a1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddlenlp==2.0.0b4) (2.6.0)
Requirement already satisfied, skipping upgrade: pyflakes<2.3.0,>=2.2.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flake8>=3.7.9->visualdl->paddlenlp==2.0.0b4) (2.2.0)
Requirement already satisfied, skipping upgrade: toml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (0.10.0)
Requirement already satisfied, skipping upgrade: pyyaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (5.1.2)
Requirement already satisfied, skipping upgrade: nodeenv>=0.11.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (1.3.4)
Requirement already satisfied, skipping upgrade: virtualenv>=15.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (16.7.9)
Requirement already satisfied, skipping upgrade: aspy.yaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (1.3.0)
Requirement already satisfied, skipping upgrade: identify>=1.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (1.4.10)
Requirement already satisfied, skipping upgrade: cfgv>=2.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp==2.0.0b4) (2.0.1)
Requirement already satisfied, skipping upgrade: Werkzeug>=0.15 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddlenlp==2.0.0b4) (0.16.0)
Requirement already satisfied, skipping upgrade: itsdangerous>=0.24 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddlenlp==2.0.0b4) (1.1.0)
Requirement already satisfied, skipping upgrade: click>=5.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from flask>=1.1.1->visualdl->paddlenlp==2.0.0b4) (7.0)
Requirement already satisfied, skipping upgrade: idna<2.9,>=2.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->paddlenlp==2.0.0b4) (2.8)
Requirement already satisfied, skipping upgrade: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->paddlenlp==2.0.0b4) (1.25.6)
Requirement already satisfied, skipping upgrade: certifi>=2017.4.17 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->paddlenlp==2.0.0b4) (2019.9.11)
Requirement already satisfied, skipping upgrade: chardet<3.1.0,>=3.0.2 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from requests->visualdl->paddlenlp==2.0.0b4) (3.0.4)
Requirement already satisfied, skipping upgrade: future>=0.6.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from bce-python-sdk->visualdl->paddlenlp==2.0.0b4) (0.18.0)
Requirement already satisfied, skipping upgrade: pycryptodome>=3.8.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from bce-python-sdk->visualdl->paddlenlp==2.0.0b4) (3.9.9)
Requirement already satisfied, skipping upgrade: joblib>=0.11 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-learn>=0.21.3->seqeval->paddlenlp==2.0.0b4) (0.14.1)
Requirement already satisfied, skipping upgrade: scipy>=0.17.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from scikit-learn>=0.21.3->seqeval->paddlenlp==2.0.0b4) (1.3.0)
Requirement already satisfied, skipping upgrade: MarkupSafe>=0.23 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from Jinja2>=2.5->Flask-Babel>=1.0.0->visualdl->paddlenlp==2.0.0b4) (1.1.1)
Requirement already satisfied, skipping upgrade: zipp>=0.5 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from importlib-metadata; python_version < "3.8"->flake8>=3.7.9->visualdl->paddlenlp==2.0.0b4) (0.6.0)
Requirement already satisfied, skipping upgrade: more-itertools in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from zipp>=0.5->importlib-metadata; python_version < "3.8"->flake8>=3.7.9->visualdl->paddlenlp==2.0.0b4) (7.2.0)

查看安装的版本

import paddle
import paddlenlp

print(paddle.__version__, paddlenlp.__version__)

paddle.set_device('gpu')

2.0.0 2.0.0b4





CUDAPlace(0)

PaddleNLP和Paddle框架是什么关系？

• Paddle框架是基础底座，提供深度学习任务全流程API。PaddleNLP基于Paddle框架开发，适用于NLP任务。

PaddleNLP中数据处理、数据集、组网单元等API未来会沉淀到框架paddle.text中。

• 代码中继承
  class TSVDataset(paddle.io.Dataset)

使用飞桨完成深度学习任务的通用流程

• 数据集和数据处理
  paddle.io.Dataset
paddle.io.DataLoader
paddlenlp.data

• 组网和网络配置

  paddle.nn.Embedding
paddlenlp.seq2vec
paddle.nn.Linear
paddle.tanh

  paddle.nn.CrossEntropyLoss
paddle.metric.Accuracy
paddle.optimizer

  model.prepare

• 网络训练和评估

  model.fit
model.evaluate

• 预测

  model.predict

注意：建议在GPU下运行。

import numpy as np
from functools import partial

import paddle.nn as nn
import paddle.nn.functional as F
import paddlenlp as ppnlp
from paddlenlp.data import Pad, Stack, Tuple
from paddlenlp.datasets import MapDatasetWrapper

from utils import load_vocab, convert_example

数据集和数据处理

自定义数据集

映射式(map-style)数据集需要继承paddle.io.Dataset

• __getitem__: 根据给定索引获取数据集中指定样本，在 paddle.io.DataLoader 中需要使用此函数通过下标获取样本。

• __len__: 返回数据集样本个数， paddle.io.BatchSampler 中需要样本个数生成下标序列。

class SelfDefinedDataset(paddle.io.Dataset):
    def __init__(self, data):
        super(SelfDefinedDataset, self).__init__()
        self.data = data

    def __getitem__(self, idx):
        return self.data[idx]

    def __len__(self):
        return len(self.data)
        
    def get_labels(self):
        return ["0", "1"]

def txt_to_list(file_name):
    res_list = []
    for line in open(file_name):
        res_list.append(line.strip().split('\t'))
    return res_list

trainlst = txt_to_list('train.txt')
devlst = txt_to_list('dev.txt')
testlst = txt_to_list('test.txt')

# 通过get_datasets()函数，将list数据转换为dataset。
# get_datasets()可接收[list]参数，或[str]参数，根据自定义数据集的写法自由选择。
# train_ds, dev_ds, test_ds = ppnlp.datasets.ChnSentiCorp.get_datasets(['train', 'dev', 'test'])
train_ds, dev_ds, test_ds = SelfDefinedDataset.get_datasets([trainlst, devlst, testlst])

训练数据查看

label_list = train_ds.get_labels()
print(label_list)

for i in range(10):
    print(train_ds[i])

['0', '1']
['赢在心理，输在出品！杨枝太酸，三文鱼熟了，酥皮焗杏汁杂果可以换个名（九唔搭八）', '0']
['服务一般，客人多，服务员少，但食品很不错', '1']
['東坡肉竟然有好多毛，問佢地點解，佢地仲話係咁架\ue107\ue107\ue107\ue107\ue107\ue107\ue107冇天理，第一次食東坡肉有毛，波羅包就幾好食', '0']
['父亲节去的，人很多，口味还可以上菜快！但是结账的时候，算错了没有打折，我也忘记拿清单了。说好打8折的，收银员没有打，人太多一时自己也没有想起。不知道收银员忘记，还是故意那钱露入自己钱包。。', '0']
['吃野味，吃个新鲜，你当然一定要来广州吃鹿肉啦*价格便宜，量好足，', '1']
['味道几好服务都五错推荐鹅肝乳鸽飞鱼', '1']
['作为老字号，水准保持算是不错，龟岗分店可能是位置问题，人不算多，基本不用等位，自从抢了券，去过好几次了，每次都可以打85以上的评分，算是可以了～粉丝煲每次必点，哈哈，鱼也不错，还会来帮衬的，楼下还可以免费停车！', '1']
['边到正宗啊？味味都咸死人啦，粤菜讲求鲜甜，五知点解感多人话好吃。', '0']
['环境卫生差，出品垃圾，冇下次，不知所为', '0']
['和苑真是精致粤菜第一家，服务菜品都一流', '1']

数据处理

为了将原始数据处理成模型可以读入的格式，本项目将对数据作以下处理：

• 首先使用jieba切词，之后将jieba切完后的单词映射词表中单词id。

• 使用paddle.io.DataLoader接口多线程异步加载数据。

其中用到了PaddleNLP中关于数据处理的API。PaddleNLP提供了许多关于NLP任务中构建有效的数据pipeline的常用API

API	简介
paddlenlp.data.Stack	堆叠N个具有相同shape的输入数据来构建一个batch，它的输入必须具有相同的shape，输出便是这些输入的堆叠组成的batch数据。
paddlenlp.data.Pad	堆叠N个输入数据来构建一个batch，每个输入数据将会被padding到N个输入数据中最大的长度
paddlenlp.data.Tuple	将多个组batch的函数包装在一起

更多数据处理操作详见： https://github.com/PaddlePaddle/PaddleNLP/blob/develop/docs/data.md

# 下载词汇表文件word_dict.txt，用于构造词-id映射关系。
# !wget https://paddlenlp.bj.bcebos.com/data/senta_word_dict.txt

# 加载词表
vocab = load_vocab('./senta_word_dict.txt')

# 打印填补单词及对应向量
for k, v in vocab.items():
    print(k, v)
    break

[PAD] 0

构造dataloder

下面的create_data_loader函数用于创建运行和预测时所需要的DataLoader对象。

• paddle.io.DataLoader返回一个迭代器，该迭代器根据batch_sampler指定的顺序迭代返回dataset数据。异步加载数据。

• batch_sampler：DataLoader通过 batch_sampler 产生的mini-batch索引列表来 dataset 中索引样本并组成mini-batch

• collate_fn：指定如何将样本列表组合为mini-batch数据。传给它参数需要是一个callable对象，需要实现对组建的batch的处理逻辑，并返回每个batch的数据。在这里传入的是prepare_input函数，对产生的数据进行pad操作，并返回实际长度等。

# Reads data and generates mini-batches.
def create_dataloader(dataset,
                      trans_function=None,
                      mode='train',
                      batch_size=1,
                      pad_token_id=0,
                      batchify_fn=None):
    if trans_function:
        dataset = dataset.apply(trans_function, lazy=True)

    # return_list 数据是否以list形式返回
    # collate_fn  指定如何将样本列表组合为mini-batch数据。传给它参数需要是一个callable对象，需要实现对组建的batch的处理逻辑，并返回每个batch的数据。在这里传入的是`prepare_input`函数，对产生的数据进行pad操作，并返回实际长度等。
    dataloader = paddle.io.DataLoader(
        dataset,
        return_list=True,
        batch_size=batch_size,
        collate_fn=batchify_fn)
        
    return dataloader

# python中的偏函数partial，把一个函数的某些参数固定住（也就是设置默认值），返回一个新的函数，调用这个新函数会更简单。
trans_function = partial(
    convert_example,
    vocab=vocab,
    unk_token_id=vocab.get('[UNK]', 1),
    is_test=False)

# 将读入的数据batch化处理，便于模型batch化运算。
# batch中的每个句子将会padding到这个batch中的文本最大长度batch_max_seq_len。
# 当文本长度大于batch_max_seq时，将会截断到batch_max_seq_len；当文本长度小于batch_max_seq时，将会padding补齐到batch_max_seq_len.
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=vocab['[PAD]']),  # input_ids
    Stack(dtype="int64"),  # seq len
    Stack(dtype="int64")  # label
): [data for data in fn(samples)]


train_loader = create_dataloader(
    train_ds,
    trans_function=trans_function,
    batch_size=128,
    mode='train',
    batchify_fn=batchify_fn)

dev_loader = create_dataloader(
    dev_ds,
    trans_function=trans_function,
    batch_size=128,
    mode='validation',
    batchify_fn=batchify_fn)
    
test_loader = create_dataloader(
    test_ds,
    trans_function=trans_function,
    batch_size=128,
    mode='test',
    batchify_fn=batchify_fn)

模型搭建

使用LSTMencoder搭建一个BiLSTM模型用于进行句子建模，得到句子的向量表示。

然后接一个线性变换层，完成二分类任务。

• paddle.nn.Embedding组建word-embedding层
• ppnlp.seq2vec.LSTMEncoder组建句子建模层
• paddle.nn.Linear构造二分类器

图1：seq2vec示意图

• 除LSTM外，seq2vec还提供了许多语义表征方法，详细可参考：seq2vec介绍

• LSTMEncoder参数：

• input_size: int，必选。输入特征Tensor的最后一维维度。
• hidden_size: int，必选。lstm运算的hidden size。
• num_layers:int，可选，lstm层数，默认为1。
• direction: str，可选，lstm运算方向，可选forward， bidirectional。默认forward。
• dropout: float，可选，dropout概率值。如果设置非0，则将对每一层lstm输出做dropout操作。默认为0.0。
• pooling_type: str， 可选，默认为None。可选sum，max，mean。如pooling_type=None， 则将最后一层lstm的最后一个step hidden输出作为文本语义表征; 如pooling_type!=None， 则将最后一层lstm的所有step的hidden输出做指定pooling操作，其结果作为文本语义表征。

更多seq2vec信息参考：https://github.com/PaddlePaddle/models/blob/develop/PaddleNLP/paddlenlp/seq2vec/encoder.py

class LSTMModel(nn.Layer):
    def __init__(self,
                 vocab_size,
                 num_classes,
                 emb_dim=128,
                 padding_idx=0,
                 lstm_hidden_size=198,
                 direction='forward',
                 lstm_layers=1,
                 dropout_rate=0,
                 pooling_type=None,
                 fc_hidden_size=96):
        super().__init__()

        # 首先将输入word id 查表后映射成 word embedding
        self.embedder = nn.Embedding(
            num_embeddings=vocab_size,
            embedding_dim=emb_dim,
            padding_idx=padding_idx)

        # 将word embedding经过LSTMEncoder变换到文本语义表征空间中
        self.lstm_encoder = ppnlp.seq2vec.LSTMEncoder(
            emb_dim,
            lstm_hidden_size,
            num_layers=lstm_layers,
            direction=direction,
            dropout=dropout_rate,
            pooling_type=pooling_type)

        # LSTMEncoder.get_output_dim()方法可以获取经过encoder之后的文本表示hidden_size
        self.fc = nn.Linear(self.lstm_encoder.get_output_dim(), fc_hidden_size)

        # 最后的分类器
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text, seq_len):
        # text shape: (batch_size, num_tokens)
        # print('input :', text.shape)
        
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)
        # print('after word-embeding:', embedded_text.shape)

        # Shape: (batch_size, num_tokens, num_directions*lstm_hidden_size)
        # num_directions = 2 if direction is 'bidirectional' else 1
        text_repr = self.lstm_encoder(embedded_text, sequence_length=seq_len)
        # print('after lstm:', text_repr.shape)


        # Shape: (batch_size, fc_hidden_size)
        fc_out = paddle.tanh(self.fc(text_repr))
        # print('after Linear classifier:', fc_out.shape)

        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc_out)
        # print('output:', logits.shape)
        
        # probs 分类概率值
        probs = F.softmax(logits, axis=-1)
        # print('output probability:', probs.shape)
        return probs

model= LSTMModel(
        len(vocab),
        len(label_list),
        direction='bidirectional',
        padding_idx=vocab['[PAD]'])
        
model = paddle.Model(model)

模型配置和训练

模型配置

optimizer = paddle.optimizer.Adam(
        parameters=model.parameters(), learning_rate=5e-5)

loss = paddle.nn.CrossEntropyLoss()

metric = paddle.metric.Accuracy()

model.prepare(optimizer, loss, metric)

# 设置visualdl路径
log_dir = './visualdl'
callbacks = paddle.callbacks.VisualDL(log_dir=log_dir)

模型训练

训练过程中会输出loss、acc等信息。

这里一共训练了10个epoch，在训练集上准确率约97%。

model.fit(train_loader, 
            dev_loader, 
            epochs=10, 
            save_dir='./checkpoints', 
            save_freq=5, 
            callbacks=callbacks)

The loss value printed in the log is the current step, and the metric is the average value of previous step.
Epoch 1/10


Building prefix dict from the default dictionary ...
Dumping model to file cache /tmp/jieba.cache
Loading model cost 0.867 seconds.
Prefix dict has been built successfully.
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:77: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  return (isinstance(seq, collections.Sequence) and


step  10/125 - loss: 0.6945 - acc: 0.4734 - 234ms/step
step  20/125 - loss: 0.6929 - acc: 0.4914 - 163ms/step
step  30/125 - loss: 0.6919 - acc: 0.5068 - 138ms/step
step  40/125 - loss: 0.6904 - acc: 0.5109 - 125ms/step
step  50/125 - loss: 0.6878 - acc: 0.5145 - 119ms/step
step  60/125 - loss: 0.6949 - acc: 0.5137 - 115ms/step
step  70/125 - loss: 0.6923 - acc: 0.5143 - 113ms/step
step  80/125 - loss: 0.6877 - acc: 0.5125 - 111ms/step
step  90/125 - loss: 0.6898 - acc: 0.5122 - 109ms/step
step 100/125 - loss: 0.6846 - acc: 0.5141 - 107ms/step
step 110/125 - loss: 0.6800 - acc: 0.5156 - 105ms/step
step 120/125 - loss: 0.6790 - acc: 0.5281 - 104ms/step
step 125/125 - loss: 0.6796 - acc: 0.5379 - 102ms/step
save checkpoint at /home/aistudio/checkpoints/0
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6775 - acc: 0.7937 - 96ms/step
step 20/84 - loss: 0.6774 - acc: 0.7941 - 84ms/step
step 30/84 - loss: 0.6776 - acc: 0.7964 - 78ms/step
step 40/84 - loss: 0.6765 - acc: 0.7986 - 74ms/step
step 50/84 - loss: 0.6798 - acc: 0.7972 - 71ms/step
step 60/84 - loss: 0.6748 - acc: 0.7991 - 69ms/step
step 70/84 - loss: 0.6782 - acc: 0.8012 - 68ms/step
step 80/84 - loss: 0.6776 - acc: 0.8011 - 66ms/step
step 84/84 - loss: 0.6750 - acc: 0.8011 - 63ms/step
Eval samples: 10646
Epoch 2/10
step  10/125 - loss: 0.6812 - acc: 0.7531 - 125ms/step
step  20/125 - loss: 0.6665 - acc: 0.7902 - 110ms/step
step  30/125 - loss: 0.6578 - acc: 0.7987 - 108ms/step
step  40/125 - loss: 0.6452 - acc: 0.7977 - 104ms/step
step  50/125 - loss: 0.6238 - acc: 0.8003 - 103ms/step
step  60/125 - loss: 0.5803 - acc: 0.8124 - 102ms/step
step  70/125 - loss: 0.4889 - acc: 0.8177 - 101ms/step
step  80/125 - loss: 0.4504 - acc: 0.8218 - 100ms/step
step  90/125 - loss: 0.4354 - acc: 0.8266 - 99ms/step
step 100/125 - loss: 0.3977 - acc: 0.8316 - 98ms/step
step 110/125 - loss: 0.4341 - acc: 0.8364 - 97ms/step
step 120/125 - loss: 0.4397 - acc: 0.8417 - 97ms/step
step 125/125 - loss: 0.4236 - acc: 0.8430 - 95ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.4360 - acc: 0.8906 - 87ms/step
step 20/84 - loss: 0.4167 - acc: 0.8898 - 71ms/step
step 30/84 - loss: 0.4203 - acc: 0.8971 - 67ms/step
step 40/84 - loss: 0.3835 - acc: 0.8986 - 64ms/step
step 50/84 - loss: 0.3996 - acc: 0.8978 - 64ms/step
step 60/84 - loss: 0.4477 - acc: 0.8962 - 62ms/step
step 70/84 - loss: 0.4174 - acc: 0.8952 - 63ms/step
step 80/84 - loss: 0.4231 - acc: 0.8960 - 65ms/step
step 84/84 - loss: 0.4522 - acc: 0.8966 - 62ms/step
Eval samples: 10646
Epoch 3/10
step  10/125 - loss: 0.4684 - acc: 0.8922 - 107ms/step
step  20/125 - loss: 0.4446 - acc: 0.8938 - 96ms/step
step  30/125 - loss: 0.4317 - acc: 0.9008 - 100ms/step
step  40/125 - loss: 0.4128 - acc: 0.9084 - 101ms/step
step  50/125 - loss: 0.4111 - acc: 0.9125 - 98ms/step
step  60/125 - loss: 0.3678 - acc: 0.9182 - 96ms/step
step  70/125 - loss: 0.3552 - acc: 0.9212 - 95ms/step
step  80/125 - loss: 0.3769 - acc: 0.9218 - 94ms/step
step  90/125 - loss: 0.3651 - acc: 0.9234 - 94ms/step
step 100/125 - loss: 0.3755 - acc: 0.9231 - 93ms/step
step 110/125 - loss: 0.3678 - acc: 0.9234 - 93ms/step
step 120/125 - loss: 0.3909 - acc: 0.9249 - 92ms/step
step 125/125 - loss: 0.3978 - acc: 0.9244 - 90ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3998 - acc: 0.9320 - 64ms/step
step 20/84 - loss: 0.3875 - acc: 0.9309 - 65ms/step
step 30/84 - loss: 0.3662 - acc: 0.9305 - 67ms/step
step 40/84 - loss: 0.3615 - acc: 0.9322 - 67ms/step
step 50/84 - loss: 0.3896 - acc: 0.9309 - 66ms/step
step 60/84 - loss: 0.3854 - acc: 0.9326 - 65ms/step
step 70/84 - loss: 0.3862 - acc: 0.9317 - 64ms/step
step 80/84 - loss: 0.3754 - acc: 0.9324 - 62ms/step
step 84/84 - loss: 0.4394 - acc: 0.9332 - 60ms/step
Eval samples: 10646
Epoch 4/10
step  10/125 - loss: 0.4256 - acc: 0.9219 - 129ms/step
step  20/125 - loss: 0.4016 - acc: 0.9305 - 108ms/step
step  30/125 - loss: 0.3773 - acc: 0.9315 - 101ms/step
step  40/125 - loss: 0.3954 - acc: 0.9346 - 96ms/step
step  50/125 - loss: 0.3782 - acc: 0.9353 - 95ms/step
step  60/125 - loss: 0.3464 - acc: 0.9398 - 93ms/step
step  70/125 - loss: 0.3456 - acc: 0.9427 - 93ms/step
step  80/125 - loss: 0.3636 - acc: 0.9429 - 93ms/step
step  90/125 - loss: 0.3477 - acc: 0.9435 - 93ms/step
step 100/125 - loss: 0.3602 - acc: 0.9432 - 92ms/step
step 110/125 - loss: 0.3622 - acc: 0.9431 - 92ms/step
step 120/125 - loss: 0.3756 - acc: 0.9439 - 92ms/step
step 125/125 - loss: 0.3703 - acc: 0.9433 - 90ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3898 - acc: 0.9391 - 84ms/step
step 20/84 - loss: 0.3813 - acc: 0.9410 - 73ms/step
step 30/84 - loss: 0.3603 - acc: 0.9414 - 73ms/step
step 40/84 - loss: 0.3644 - acc: 0.9422 - 71ms/step
step 50/84 - loss: 0.3744 - acc: 0.9417 - 70ms/step
step 60/84 - loss: 0.3567 - acc: 0.9437 - 70ms/step
step 70/84 - loss: 0.3745 - acc: 0.9420 - 70ms/step
step 80/84 - loss: 0.3677 - acc: 0.9426 - 68ms/step
step 84/84 - loss: 0.4366 - acc: 0.9432 - 66ms/step
Eval samples: 10646
Epoch 5/10
step  10/125 - loss: 0.3941 - acc: 0.9328 - 114ms/step
step  20/125 - loss: 0.3838 - acc: 0.9387 - 107ms/step
step  30/125 - loss: 0.3766 - acc: 0.9414 - 102ms/step
step  40/125 - loss: 0.3818 - acc: 0.9439 - 98ms/step
step  50/125 - loss: 0.3641 - acc: 0.9450 - 97ms/step
step  60/125 - loss: 0.3353 - acc: 0.9488 - 96ms/step
step  70/125 - loss: 0.3363 - acc: 0.9510 - 95ms/step
step  80/125 - loss: 0.3508 - acc: 0.9511 - 95ms/step
step  90/125 - loss: 0.3450 - acc: 0.9513 - 95ms/step
step 100/125 - loss: 0.3450 - acc: 0.9514 - 95ms/step
step 110/125 - loss: 0.3547 - acc: 0.9513 - 94ms/step
step 120/125 - loss: 0.3697 - acc: 0.9520 - 93ms/step
step 125/125 - loss: 0.3807 - acc: 0.9512 - 92ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3845 - acc: 0.9414 - 86ms/step
step 20/84 - loss: 0.3750 - acc: 0.9465 - 71ms/step
step 30/84 - loss: 0.3583 - acc: 0.9458 - 67ms/step
step 40/84 - loss: 0.3670 - acc: 0.9463 - 65ms/step
step 50/84 - loss: 0.3688 - acc: 0.9453 - 64ms/step
step 60/84 - loss: 0.3614 - acc: 0.9467 - 63ms/step
step 70/84 - loss: 0.3717 - acc: 0.9452 - 64ms/step
step 80/84 - loss: 0.3554 - acc: 0.9458 - 65ms/step
step 84/84 - loss: 0.4361 - acc: 0.9465 - 64ms/step
Eval samples: 10646
Epoch 6/10
step  10/125 - loss: 0.3749 - acc: 0.9477 - 75ms/step
step  20/125 - loss: 0.3694 - acc: 0.9504 - 75ms/step
step  30/125 - loss: 0.3521 - acc: 0.9539 - 73ms/step
step  40/125 - loss: 0.3791 - acc: 0.9541 - 78ms/step
step  50/125 - loss: 0.3515 - acc: 0.9544 - 81ms/step
step  60/125 - loss: 0.3352 - acc: 0.9574 - 81ms/step
step  70/125 - loss: 0.3314 - acc: 0.9590 - 81ms/step
step  80/125 - loss: 0.3496 - acc: 0.9584 - 82ms/step
step  90/125 - loss: 0.3433 - acc: 0.9582 - 82ms/step
step 100/125 - loss: 0.3400 - acc: 0.9580 - 83ms/step
step 110/125 - loss: 0.3451 - acc: 0.9580 - 83ms/step
step 120/125 - loss: 0.3599 - acc: 0.9589 - 83ms/step
step 125/125 - loss: 0.3598 - acc: 0.9585 - 81ms/step
save checkpoint at /home/aistudio/checkpoints/5
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3754 - acc: 0.9484 - 85ms/step
step 20/84 - loss: 0.3714 - acc: 0.9535 - 69ms/step
step 30/84 - loss: 0.3558 - acc: 0.9523 - 65ms/step
step 40/84 - loss: 0.3603 - acc: 0.9533 - 62ms/step
step 50/84 - loss: 0.3719 - acc: 0.9514 - 61ms/step
step 60/84 - loss: 0.3442 - acc: 0.9525 - 60ms/step
step 70/84 - loss: 0.3654 - acc: 0.9513 - 59ms/step
step 80/84 - loss: 0.3602 - acc: 0.9514 - 58ms/step
step 84/84 - loss: 0.4414 - acc: 0.9520 - 56ms/step
Eval samples: 10646
Epoch 7/10
step  10/125 - loss: 0.3673 - acc: 0.9523 - 104ms/step
step  20/125 - loss: 0.3638 - acc: 0.9566 - 93ms/step
step  30/125 - loss: 0.3488 - acc: 0.9599 - 90ms/step
step  40/125 - loss: 0.3790 - acc: 0.9600 - 88ms/step
step  50/125 - loss: 0.3557 - acc: 0.9583 - 88ms/step
step  60/125 - loss: 0.3309 - acc: 0.9604 - 88ms/step
step  70/125 - loss: 0.3366 - acc: 0.9621 - 88ms/step
step  80/125 - loss: 0.3372 - acc: 0.9618 - 87ms/step
step  90/125 - loss: 0.3326 - acc: 0.9615 - 87ms/step
step 100/125 - loss: 0.3365 - acc: 0.9612 - 88ms/step
step 110/125 - loss: 0.3404 - acc: 0.9615 - 88ms/step
step 120/125 - loss: 0.3582 - acc: 0.9626 - 87ms/step
step 125/125 - loss: 0.3549 - acc: 0.9621 - 86ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3800 - acc: 0.9531 - 85ms/step
step 20/84 - loss: 0.3618 - acc: 0.9559 - 69ms/step
step 30/84 - loss: 0.3520 - acc: 0.9555 - 66ms/step
step 40/84 - loss: 0.3568 - acc: 0.9566 - 63ms/step
step 50/84 - loss: 0.3752 - acc: 0.9552 - 62ms/step
step 60/84 - loss: 0.3430 - acc: 0.9559 - 62ms/step
step 70/84 - loss: 0.3786 - acc: 0.9550 - 61ms/step
step 80/84 - loss: 0.3554 - acc: 0.9557 - 59ms/step
step 84/84 - loss: 0.3533 - acc: 0.9563 - 57ms/step
Eval samples: 10646
Epoch 8/10
step  10/125 - loss: 0.3558 - acc: 0.9617 - 109ms/step
step  20/125 - loss: 0.3595 - acc: 0.9641 - 97ms/step
step  30/125 - loss: 0.3484 - acc: 0.9654 - 93ms/step
step  40/125 - loss: 0.3728 - acc: 0.9639 - 90ms/step
step  50/125 - loss: 0.3405 - acc: 0.9639 - 89ms/step
step  60/125 - loss: 0.3275 - acc: 0.9660 - 88ms/step
step  70/125 - loss: 0.3262 - acc: 0.9673 - 87ms/step
step  80/125 - loss: 0.3359 - acc: 0.9668 - 87ms/step
step  90/125 - loss: 0.3285 - acc: 0.9667 - 87ms/step
step 100/125 - loss: 0.3344 - acc: 0.9663 - 87ms/step
step 110/125 - loss: 0.3351 - acc: 0.9666 - 87ms/step
step 120/125 - loss: 0.3564 - acc: 0.9676 - 87ms/step
step 125/125 - loss: 0.3524 - acc: 0.9672 - 86ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3692 - acc: 0.9602 - 73ms/step
step 20/84 - loss: 0.3669 - acc: 0.9582 - 72ms/step
step 30/84 - loss: 0.3471 - acc: 0.9586 - 70ms/step
step 40/84 - loss: 0.3467 - acc: 0.9586 - 69ms/step
step 50/84 - loss: 0.3713 - acc: 0.9573 - 69ms/step
step 60/84 - loss: 0.3442 - acc: 0.9578 - 69ms/step
step 70/84 - loss: 0.3561 - acc: 0.9576 - 69ms/step
step 80/84 - loss: 0.3410 - acc: 0.9579 - 69ms/step
step 84/84 - loss: 0.4010 - acc: 0.9585 - 66ms/step
Eval samples: 10646
Epoch 9/10
step  10/125 - loss: 0.3517 - acc: 0.9602 - 106ms/step
step  20/125 - loss: 0.3577 - acc: 0.9648 - 95ms/step
step  30/125 - loss: 0.3434 - acc: 0.9669 - 92ms/step
step  40/125 - loss: 0.3667 - acc: 0.9660 - 89ms/step
step  50/125 - loss: 0.3391 - acc: 0.9661 - 89ms/step
step  60/125 - loss: 0.3251 - acc: 0.9680 - 88ms/step
step  70/125 - loss: 0.3235 - acc: 0.9695 - 87ms/step
step  80/125 - loss: 0.3325 - acc: 0.9692 - 87ms/step
step  90/125 - loss: 0.3263 - acc: 0.9694 - 90ms/step
step 100/125 - loss: 0.3323 - acc: 0.9692 - 92ms/step
step 110/125 - loss: 0.3316 - acc: 0.9694 - 93ms/step
step 120/125 - loss: 0.3547 - acc: 0.9702 - 93ms/step
step 125/125 - loss: 0.3506 - acc: 0.9699 - 91ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3649 - acc: 0.9609 - 84ms/step
step 20/84 - loss: 0.3670 - acc: 0.9586 - 69ms/step
step 30/84 - loss: 0.3463 - acc: 0.9586 - 64ms/step
step 40/84 - loss: 0.3450 - acc: 0.9594 - 62ms/step
step 50/84 - loss: 0.3687 - acc: 0.9583 - 61ms/step
step 60/84 - loss: 0.3484 - acc: 0.9587 - 60ms/step
step 70/84 - loss: 0.3511 - acc: 0.9587 - 59ms/step
step 80/84 - loss: 0.3392 - acc: 0.9592 - 58ms/step
step 84/84 - loss: 0.4006 - acc: 0.9597 - 56ms/step
Eval samples: 10646
Epoch 10/10
step  10/125 - loss: 0.3449 - acc: 0.9625 - 89ms/step
step  20/125 - loss: 0.3561 - acc: 0.9676 - 105ms/step
step  30/125 - loss: 0.3506 - acc: 0.9693 - 108ms/step
step  40/125 - loss: 0.3665 - acc: 0.9676 - 109ms/step
step  50/125 - loss: 0.3380 - acc: 0.9675 - 107ms/step
step  60/125 - loss: 0.3235 - acc: 0.9693 - 106ms/step
step  70/125 - loss: 0.3228 - acc: 0.9708 - 105ms/step
step  80/125 - loss: 0.3274 - acc: 0.9706 - 106ms/step
step  90/125 - loss: 0.3256 - acc: 0.9705 - 105ms/step
step 100/125 - loss: 0.3307 - acc: 0.9702 - 103ms/step
step 110/125 - loss: 0.3350 - acc: 0.9700 - 102ms/step
step 120/125 - loss: 0.3551 - acc: 0.9709 - 100ms/step
step 125/125 - loss: 0.3524 - acc: 0.9706 - 98ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3675 - acc: 0.9602 - 89ms/step
step 20/84 - loss: 0.3638 - acc: 0.9602 - 71ms/step
step 30/84 - loss: 0.3522 - acc: 0.9599 - 66ms/step
step 40/84 - loss: 0.3541 - acc: 0.9600 - 63ms/step
step 50/84 - loss: 0.3643 - acc: 0.9587 - 62ms/step
step 60/84 - loss: 0.3276 - acc: 0.9600 - 60ms/step
step 70/84 - loss: 0.3688 - acc: 0.9596 - 60ms/step
step 80/84 - loss: 0.3467 - acc: 0.9597 - 58ms/step
step 84/84 - loss: 0.3462 - acc: 0.9603 - 56ms/step
Eval samples: 10646
save checkpoint at /home/aistudio/checkpoints/final

启动VisualDL查看训练过程可视化结果

启动步骤：

• 1、切换到本界面左侧「可视化」
• 2、日志文件路径选择 ‘visualdl’
• 3、点击「启动VisualDL」后点击「打开VisualDL」，即可查看可视化结果：

Accuracy和Loss的实时变化趋势如下：

评估

最终得到的评估准确率为96%

results = model.evaluate(dev_loader)
print("Finally test acc: %.5f" % results['acc'])

Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.3675 - acc: 0.9602 - 85ms/step
step 20/84 - loss: 0.3638 - acc: 0.9602 - 70ms/step
step 30/84 - loss: 0.3522 - acc: 0.9599 - 66ms/step
step 40/84 - loss: 0.3541 - acc: 0.9600 - 64ms/step
step 50/84 - loss: 0.3643 - acc: 0.9587 - 63ms/step
step 60/84 - loss: 0.3276 - acc: 0.9600 - 62ms/step
step 70/84 - loss: 0.3688 - acc: 0.9596 - 61ms/step
step 80/84 - loss: 0.3467 - acc: 0.9597 - 61ms/step
step 84/84 - loss: 0.3462 - acc: 0.9603 - 59ms/step
Eval samples: 10646
Finally test acc: 0.96027

#保存模型
model.save('./finetuning/lstm/model', training=True)

预测

model= LSTMModel(
        len(vocab),
        len(label_list),
        direction='bidirectional',
        padding_idx=vocab['[PAD]'])

model = paddle.Model(model)

model.load('./finetuning/model')

model.prepare()

label_map = {0: 'negative', 1: 'positive'}
results = model.predict(test_loader, batch_size=128)[0]
predictions = []

for batch_probs in results:
    # 映射分类label
    idx = np.argmax(batch_probs, axis=-1)
    idx = idx.tolist()
    labels = [label_map[i] for i in idx]
    predictions.extend(labels)

# 看看预测数据前5个样例分类结果
for idx, data in enumerate(test_ds.data[:10]):
    print('Data: {} \t Label: {}'.format(data[0], predictions[idx]))

Predict begin...
step 42/42 [==============================] - ETA: 3s - 97ms/ste - ETA: 3s - 104ms/st - ETA: 3s - 108ms/st - ETA: 3s - 97ms/step - ETA: 2s - 89ms/ste - ETA: 2s - 85ms/ste - ETA: 2s - 81ms/ste - ETA: 2s - 78ms/ste - ETA: 1s - 76ms/ste - ETA: 1s - 74ms/ste - ETA: 1s - 73ms/ste - ETA: 1s - 72ms/ste - ETA: 1s - 71ms/ste - ETA: 0s - 70ms/ste - ETA: 0s - 70ms/ste - ETA: 0s - 69ms/ste - ETA: 0s - 69ms/ste - ETA: 0s - 68ms/ste - ETA: 0s - 67ms/ste - ETA: 0s - 64ms/ste - 62ms/step          
Predict samples: 5353
Data: 楼面经理服务态度极差，等位和埋单都差，楼面小妹还挺好 	 Label: negative
Data: 欺负北方人没吃过鲍鱼是怎么着？简直敷衍到可笑的程度，团购连青菜都是两人份？！难吃到死，菜色还特别可笑，什么时候粤菜的小菜改成拍黄瓜了？！把团购客人当傻子，可这满大厅的傻子谁还会再来？！ 	 Label: negative
Data: 如果大家有时间而且不怕麻烦的话可以去这里试试，点一个饭等左2个钟，没错！是两个钟！期间催了n遍都说马上到，结果？呵呵。乳鸽的味道，太咸，可能不新鲜吧……要用重口味盖住异味。上菜超级慢！中途还搞什么表演，麻烦有人手的话就上菜啊，表什么演？！？！要大家饿着看表演吗？最后结账还算错单，我真心服了……有一种店叫不会有下次，大概就是指它吧 	 Label: negative
Data: 偌大的一个大厅就一个人点菜，点菜速度超级慢，菜牌上多个菜停售，连续点了两个没标停售的菜也告知没有，粥上来是凉的，榴莲酥火大了，格格肉超级油腻而且咸?????? 	 Label: negative
Data: 泥撕雞超級好吃！！！吃了一個再叫一個還想打包的節奏！ 	 Label: positive
Data: 作为地道的广州人，从小就跟着家人在西关品尝各式美食，今日带着家中长辈来这个老字号泮溪酒家真实失望透顶，出品差、服务差、洗手间邋遢弥漫着浓郁尿骚味、丢广州人的脸、丢广州老字号的脸。 	 Label: negative
Data: 辣味道很赞哦！猪肚鸡一直是我们的最爱，每次来都必点，服务很给力，环境很好，值得分享哦！西洋菜 	 Label: positive
Data: 第一次吃到這麼脏的火鍋：吃着吃著吃出一條尾指粗的黑毛毛蟲——惡心！脏！！！第一次吃到這麼無誠信的火鍋服務：我們呼喚人員時，某女部長立即使服務員迅速取走蟲所在的碗，任我們多次叫「放下」論理，她們也置若罔聞轉身將蟲毁屍滅跡，還嘻皮笑臉辯稱只是把碗換走,態度行為惡劣——奸詐！毫無誠信！！爛！！！當然還有剛坐下時的情形：第一次吃到這樣的火鍋：所有肉食熟食都上桌了，鍋底遲遲沒上，足足等了半小時才姍姍來遲；---差！！第一次吃到這樣的火鍋：1元雞鍋、1碟6塊小牛肉、1碟小腐皮、1碟5塊裝的普通肥牛、1碟數片的細碎牛肚結帳便2百多元；---不值！！以下省略千字差評......白云路的稻香是最差、最失禮的稻香，天河城、華廈的都比它好上過萬倍！！白云路的稻香是史上最差的餐廳！！！ 	 Label: negative
Data: 文昌鸡份量很少且很咸，其他菜味道很一般！服务态度差差差！还要10%的服务费、 	 Label: negative
Data: 这个网站的评价真是越来越不可信了，搞不懂为什么这么多好评。真的是很一般，不要迷信什么哪里回来的大厨吧。环境和出品若是当作普通茶餐厅来看待就还说得过去，但是价格又不是茶餐厅的价格，这就很尴尬了。。服务也是有待提高。 	 Label: negative

修改seq2vec模型

seq2vec模块

• 输入：文本序列的Embedding Tensor，shape：(batch_size, num_token, emb_dim)

• 输出：文本语义表征Enocded Texts Tensor，shape：(batch_sie,encoding_size)

• 提供了BoWEncoder，CNNEncoder，GRUEncoder，LSTMEncoder，RNNEncoder等模型

  • BoWEncoder 是将输入序列Embedding Tensor在num_token维度上叠加，得到文本语义表征Enocded Texts Tensor。

  • CNNEncoder 是将输入序列Embedding Tensor进行卷积操作，在对卷积结果进行max_pooling，得到文本语义表征Enocded Texts Tensor。

  • GRUEncoder 是对输入序列Embedding Tensor进行GRU运算，在运算结果上进行pooling或者取最后一个step的隐表示，得到文本语义表征Enocded Texts Tensor。

  • LSTMEncoder 是对输入序列Embedding Tensor进行LSTM运算，在运算结果上进行pooling或者取最后一个step的隐表示，得到文本语义表征Enocded Texts Tensor。

  • RNNEncoder 是对输入序列Embedding Tensor进行RNN运算，在运算结果上进行pooling或者取最后一个step的隐表示，得到文本语义表征Enocded Texts Tensor。

• seq2vec提供了许多语义表征方法，那么这些方法有什么特点呢？

  1. BoWEncoder采用Bag of Word Embedding方法，其特点是简单。但其缺点是没有考虑文本的语境，所以对文本语义的表征不足以表意。
  2. CNNEncoder采用卷积操作，提取局部特征，其特点是可以共享权重。但其缺点同样只考虑了局部语义，上下文信息没有充分利用。
  图2：卷积示意图
  3. RNNEnocder采用RNN方法，在计算下一个token语义信息时，利用上一个token语义信息作为其输入。但其缺点容易产生梯度消失和梯度爆炸。

  

  图3：RNN示意图
  4. LSTMEnocder采用LSTM方法，LSTM是RNN的一种变种。为了学到长期依赖关系，LSTM 中引入了门控机制来控制信息的累计速度，包括有选择地加入新的信息，并有选择地遗忘之前累计的信息。

  

  图4：LSTM示意图
  5. GRUEncoder采用GRU方法，GRU也是RNN的一种变种。一个LSTM单元有四个输入 ，因而参数是RNN的四倍，带来的结果是训练速度慢。GRU对LSTM进行了简化，在不影响效果的前提下加快了训练速度。

图5：GRU示意图

关于CNN、LSTM、GRU、RNN等更多信息参考：

• Understanding LSTM Networks: https://colah.github.io/posts/2015-08-Understanding-LSTMs/
• Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling:https://arxiv.org/abs/1412.3555
• A Critical Review of Recurrent Neural Networks
  for Sequence Learning: https://arxiv.org/pdf/1506.00019
• A Convolutional Neural Network for Modelling Sentences: https://arxiv.org/abs/1404.2188

数据准备

import paddlenlp as ppnlp


# 使用paddlenlp内置数据集
# train_ds, dev_ds, test_ds = ppnlp.datasets.ChnSentiCorp.get_datasets(['train', 'dev', 'test'])

# label_list = train_ds.get_labels()
# print(label_list)

# for sent, label in train_ds[:5]:
#     print (sent, label)

# tmp = ppnlp.datasets.ChnSentiCorp('train')
# tmp1 = ppnlp.datasets.MapDatasetWrapper(tmp)
# print(type(tmp), type(tmp1))
# for sent, label in tmp1[:5]:
#     print(sent, label)

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


# Reads data and generates mini-batches.
trans_fn = partial(
    convert_example,
    vocab=vocab,
    unk_token_id=vocab.get('[UNK]', 1),
    is_test=False)

# 将读入的数据batch化处理，便于模型batch化运算。
# batch中的每个句子将会padding到这个batch中的文本最大长度batch_max_seq_len。
# 当文本长度大于batch_max_seq时，将会截断到batch_max_seq_len；当文本长度小于batch_max_seq时，将会padding补齐到batch_max_seq_len.
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=vocab['[PAD]']),  # input_ids
    Stack(dtype="int64"),  # seq len
    Stack(dtype="int64")  # label
): [data for data in fn(samples)]

train_loader = create_dataloader(
    train_ds,
    trans_fn=trans_fn,
    batch_size=128,
    mode='train',
    batchify_fn=batchify_fn)
    
dev_loader = create_dataloader(
    dev_ds,
    trans_fn=trans_fn,
    batch_size=128,
    mode='validation',
    batchify_fn=batchify_fn)

test_loader = create_dataloader(
    test_ds,
    trans_fn=trans_fn,
    batch_size=128,
    mode='test',
    batchify_fn=batchify_fn)

模型建立

class GRUModel(nn.Layer):
    def __init__(self,
                 vocab_size,
                 num_classes,
                 emb_dim=128,
                 padding_idx=0,
                 hidden_size=198,
                 direction='forward',
                 num_layers=1,
                 dropout_rate=0,
                 pooling_type=None,
                 fc_hidden_size=96):
        super().__init__()

        # 首先将输入word id 查表后映射成 word embedding
        self.embedder = nn.Embedding(
            num_embeddings=vocab_size,
            embedding_dim=emb_dim,
            padding_idx=padding_idx)

        # 将word embedding经过LSTMEncoder变换到文本语义表征空间中
        self.gru_encoder = ppnlp.seq2vec.GRUEncoder(
            emb_dim,
            hidden_size,
            num_layers=num_layers,
            direction=direction,
            dropout=dropout_rate,
            pooling_type=pooling_type)

        # LSTMEncoder.get_output_dim()方法可以获取经过encoder之后的文本表示hidden_size
        self.fc = nn.Linear(self.gru_encoder.get_output_dim(), fc_hidden_size)

        # 最后的分类器
        self.output_layer = nn.Linear(fc_hidden_size, num_classes)

    def forward(self, text, seq_len):
        # text shape: (batch_size, num_tokens)
        # print('input :', text.shape)
        
        # Shape: (batch_size, num_tokens, embedding_dim)
        embedded_text = self.embedder(text)
        # print('after word-embeding:', embedded_text.shape)

        # Shape: (batch_size, num_tokens, num_directions*lstm_hidden_size)
        # num_directions = 2 if direction is 'bidirectional' else 1
        text_repr = self.gru_encoder(embedded_text, sequence_length=seq_len)
        # print('after lstm:', text_repr.shape)


        # Shape: (batch_size, fc_hidden_size)
        fc_out = paddle.tanh(self.fc(text_repr))
        # print('after Linear classifier:', fc_out.shape)

        # Shape: (batch_size, num_classes)
        logits = self.output_layer(fc_out)
        # print('output:', logits.shape)
        
        # probs 分类概率值
        probs = F.softmax(logits, axis=-1)
        # print('output probability:', probs.shape)
        return probs

模型配置

model= GRUModel(
        len(vocab),
        len(label_list),
        direction='bidirectional',
        padding_idx=vocab['[PAD]'])
        
model = paddle.Model(model)

optimizer = paddle.optimizer.Adam(
        parameters=model.parameters(), learning_rate=5e-5)

loss = paddle.nn.CrossEntropyLoss()

metric = paddle.metric.Accuracy()

model.prepare(optimizer, loss, metric)

# 设置visualdl路径
log_dir = './visualdl/gru_v1'
callbacks = paddle.callbacks.VisualDL(log_dir=log_dir)

model.fit(train_loader, 
            dev_loader, 
            epochs=10, 
            save_dir='./checkpoints/gru_v1', 
            save_freq=5, 
            callbacks=callbacks)

The loss value printed in the log is the current step, and the metric is the average value of previous step.
Epoch 1/10
step  10/125 - loss: 1.1046 - acc: 0.0000e+00 - 111ms/step
step  20/125 - loss: 1.0817 - acc: 0.2332 - 99ms/step
step  30/125 - loss: 1.0604 - acc: 0.3221 - 95ms/step
step  40/125 - loss: 1.0446 - acc: 0.3602 - 94ms/step
step  50/125 - loss: 1.0306 - acc: 0.3811 - 92ms/step
step  60/125 - loss: 1.0166 - acc: 0.4025 - 91ms/step
step  70/125 - loss: 1.0039 - acc: 0.4552 - 91ms/step
step  80/125 - loss: 0.9935 - acc: 0.4743 - 91ms/step
step  90/125 - loss: 0.9841 - acc: 0.5078 - 91ms/step
step 100/125 - loss: 0.9752 - acc: 0.5260 - 90ms/step
step 110/125 - loss: 0.9695 - acc: 0.5223 - 90ms/step
step 120/125 - loss: 0.9620 - acc: 0.5324 - 90ms/step
step 125/125 - loss: 0.9583 - acc: 0.5418 - 88ms/step
save checkpoint at /home/aistudio/checkpoints/gru_v1/0
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.9590 - acc: 0.7398 - 95ms/step
step 20/84 - loss: 0.9589 - acc: 0.7324 - 75ms/step
step 30/84 - loss: 0.9585 - acc: 0.7328 - 69ms/step
step 40/84 - loss: 0.9579 - acc: 0.7295 - 67ms/step
step 50/84 - loss: 0.9593 - acc: 0.7316 - 67ms/step
step 60/84 - loss: 0.9556 - acc: 0.7346 - 67ms/step
step 70/84 - loss: 0.9594 - acc: 0.7323 - 67ms/step
step 80/84 - loss: 0.9589 - acc: 0.7323 - 66ms/step
step 84/84 - loss: 0.9616 - acc: 0.7330 - 63ms/step
Eval samples: 10646
Epoch 2/10
step  10/125 - loss: 0.9529 - acc: 0.7086 - 114ms/step
step  20/125 - loss: 0.9500 - acc: 0.7051 - 104ms/step
step  30/125 - loss: 0.9377 - acc: 0.7341 - 98ms/step
step  40/125 - loss: 0.9367 - acc: 0.7318 - 96ms/step
step  50/125 - loss: 0.9213 - acc: 0.7439 - 94ms/step
step  60/125 - loss: 0.8881 - acc: 0.7379 - 93ms/step
step  70/125 - loss: 0.8537 - acc: 0.7412 - 92ms/step
step  80/125 - loss: 0.8674 - acc: 0.7401 - 93ms/step
step  90/125 - loss: 0.8386 - acc: 0.7376 - 92ms/step
step 100/125 - loss: 0.8068 - acc: 0.7366 - 92ms/step
step 110/125 - loss: 0.8452 - acc: 0.7364 - 92ms/step
step 120/125 - loss: 0.7921 - acc: 0.7374 - 91ms/step
step 125/125 - loss: 0.8001 - acc: 0.7378 - 89ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.7727 - acc: 0.7500 - 88ms/step
step 20/84 - loss: 0.7982 - acc: 0.7555 - 71ms/step
step 30/84 - loss: 0.7934 - acc: 0.7549 - 67ms/step
step 40/84 - loss: 0.7760 - acc: 0.7557 - 65ms/step
step 50/84 - loss: 0.7713 - acc: 0.7572 - 64ms/step
step 60/84 - loss: 0.7942 - acc: 0.7568 - 62ms/step
step 70/84 - loss: 0.8002 - acc: 0.7545 - 62ms/step
step 80/84 - loss: 0.7784 - acc: 0.7554 - 61ms/step
step 84/84 - loss: 0.8570 - acc: 0.7559 - 59ms/step
Eval samples: 10646
Epoch 3/10
step  10/125 - loss: 0.7971 - acc: 0.7773 - 109ms/step
step  20/125 - loss: 0.7422 - acc: 0.7855 - 100ms/step
step  30/125 - loss: 0.7125 - acc: 0.7951 - 97ms/step
step  40/125 - loss: 0.7575 - acc: 0.8016 - 94ms/step
step  50/125 - loss: 0.6859 - acc: 0.8089 - 93ms/step
step  60/125 - loss: 0.6763 - acc: 0.8214 - 93ms/step
step  70/125 - loss: 0.6569 - acc: 0.8319 - 94ms/step
step  80/125 - loss: 0.6440 - acc: 0.8431 - 93ms/step
step  90/125 - loss: 0.6573 - acc: 0.8521 - 95ms/step
step 100/125 - loss: 0.6323 - acc: 0.8586 - 94ms/step
step 110/125 - loss: 0.6311 - acc: 0.8650 - 94ms/step
step 120/125 - loss: 0.6425 - acc: 0.8699 - 94ms/step
step 125/125 - loss: 0.6360 - acc: 0.8723 - 92ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6429 - acc: 0.9258 - 87ms/step
step 20/84 - loss: 0.6527 - acc: 0.9227 - 70ms/step
step 30/84 - loss: 0.6373 - acc: 0.9237 - 66ms/step
step 40/84 - loss: 0.6258 - acc: 0.9271 - 64ms/step
step 50/84 - loss: 0.6468 - acc: 0.9269 - 63ms/step
step 60/84 - loss: 0.6249 - acc: 0.9279 - 63ms/step
step 70/84 - loss: 0.6554 - acc: 0.9261 - 62ms/step
step 80/84 - loss: 0.6408 - acc: 0.9262 - 61ms/step
step 84/84 - loss: 0.6208 - acc: 0.9280 - 58ms/step
Eval samples: 10646
Epoch 4/10
step  10/125 - loss: 0.6157 - acc: 0.9445 - 107ms/step
step  20/125 - loss: 0.6325 - acc: 0.9371 - 97ms/step
step  30/125 - loss: 0.6596 - acc: 0.9365 - 95ms/step
step  40/125 - loss: 0.6046 - acc: 0.9361 - 93ms/step
step  50/125 - loss: 0.6530 - acc: 0.9320 - 92ms/step
step  60/125 - loss: 0.6364 - acc: 0.9341 - 91ms/step
step  70/125 - loss: 0.6242 - acc: 0.9348 - 90ms/step
step  80/125 - loss: 0.6116 - acc: 0.9347 - 90ms/step
step  90/125 - loss: 0.6371 - acc: 0.9350 - 90ms/step
step 100/125 - loss: 0.6125 - acc: 0.9355 - 90ms/step
step 110/125 - loss: 0.5946 - acc: 0.9368 - 90ms/step
step 120/125 - loss: 0.6274 - acc: 0.9363 - 89ms/step
step 125/125 - loss: 0.5729 - acc: 0.9368 - 87ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6248 - acc: 0.9398 - 84ms/step
step 20/84 - loss: 0.6306 - acc: 0.9383 - 69ms/step
step 30/84 - loss: 0.6160 - acc: 0.9401 - 65ms/step
step 40/84 - loss: 0.6045 - acc: 0.9424 - 64ms/step
step 50/84 - loss: 0.6228 - acc: 0.9414 - 62ms/step
step 60/84 - loss: 0.5969 - acc: 0.9419 - 61ms/step
step 70/84 - loss: 0.6313 - acc: 0.9408 - 60ms/step
step 80/84 - loss: 0.6134 - acc: 0.9413 - 59ms/step
step 84/84 - loss: 0.6030 - acc: 0.9427 - 56ms/step
Eval samples: 10646
Epoch 5/10
step  10/125 - loss: 0.5897 - acc: 0.9445 - 107ms/step
step  20/125 - loss: 0.6003 - acc: 0.9520 - 98ms/step
step  30/125 - loss: 0.6177 - acc: 0.9516 - 96ms/step
step  40/125 - loss: 0.6054 - acc: 0.9525 - 94ms/step
step  50/125 - loss: 0.6303 - acc: 0.9537 - 93ms/step
step  60/125 - loss: 0.6067 - acc: 0.9530 - 94ms/step
step  70/125 - loss: 0.5856 - acc: 0.9540 - 93ms/step
step  80/125 - loss: 0.5976 - acc: 0.9545 - 92ms/step
step  90/125 - loss: 0.6220 - acc: 0.9537 - 91ms/step
step 100/125 - loss: 0.6000 - acc: 0.9540 - 91ms/step
step 110/125 - loss: 0.6048 - acc: 0.9522 - 91ms/step
step 120/125 - loss: 0.6144 - acc: 0.9520 - 91ms/step
step 125/125 - loss: 0.6159 - acc: 0.9520 - 89ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6338 - acc: 0.9406 - 87ms/step
step 20/84 - loss: 0.6171 - acc: 0.9441 - 71ms/step
step 30/84 - loss: 0.6064 - acc: 0.9461 - 67ms/step
step 40/84 - loss: 0.5934 - acc: 0.9490 - 65ms/step
step 50/84 - loss: 0.6212 - acc: 0.9481 - 64ms/step
step 60/84 - loss: 0.5881 - acc: 0.9484 - 64ms/step
step 70/84 - loss: 0.6167 - acc: 0.9477 - 63ms/step
step 80/84 - loss: 0.5918 - acc: 0.9479 - 61ms/step
step 84/84 - loss: 0.6166 - acc: 0.9488 - 59ms/step
Eval samples: 10646
Epoch 6/10
step  10/125 - loss: 0.5790 - acc: 0.9539 - 111ms/step
step  20/125 - loss: 0.5977 - acc: 0.9559 - 99ms/step
step  30/125 - loss: 0.5986 - acc: 0.9557 - 96ms/step
step  40/125 - loss: 0.5777 - acc: 0.9551 - 95ms/step
step  50/125 - loss: 0.5906 - acc: 0.9563 - 93ms/step
step  60/125 - loss: 0.5921 - acc: 0.9577 - 93ms/step
step  70/125 - loss: 0.5816 - acc: 0.9589 - 92ms/step
step  80/125 - loss: 0.6051 - acc: 0.9584 - 92ms/step
step  90/125 - loss: 0.5874 - acc: 0.9579 - 91ms/step
step 100/125 - loss: 0.5957 - acc: 0.9590 - 91ms/step
step 110/125 - loss: 0.6152 - acc: 0.9592 - 91ms/step
step 120/125 - loss: 0.5884 - acc: 0.9595 - 90ms/step
step 125/125 - loss: 0.6076 - acc: 0.9595 - 88ms/step
save checkpoint at /home/aistudio/checkpoints/gru_v1/5
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6259 - acc: 0.9492 - 86ms/step
step 20/84 - loss: 0.6103 - acc: 0.9500 - 69ms/step
step 30/84 - loss: 0.5972 - acc: 0.9516 - 65ms/step
step 40/84 - loss: 0.5912 - acc: 0.9543 - 63ms/step
step 50/84 - loss: 0.6174 - acc: 0.9533 - 62ms/step
step 60/84 - loss: 0.5815 - acc: 0.9540 - 61ms/step
step 70/84 - loss: 0.6090 - acc: 0.9535 - 60ms/step
step 80/84 - loss: 0.5879 - acc: 0.9540 - 58ms/step
step 84/84 - loss: 0.6208 - acc: 0.9546 - 56ms/step
Eval samples: 10646
Epoch 7/10
step  10/125 - loss: 0.5885 - acc: 0.9609 - 111ms/step
step  20/125 - loss: 0.5848 - acc: 0.9637 - 98ms/step
step  30/125 - loss: 0.5899 - acc: 0.9648 - 95ms/step
step  40/125 - loss: 0.6268 - acc: 0.9602 - 93ms/step
step  50/125 - loss: 0.5984 - acc: 0.9617 - 92ms/step
step  60/125 - loss: 0.6060 - acc: 0.9624 - 91ms/step
step  70/125 - loss: 0.6193 - acc: 0.9633 - 90ms/step
step  80/125 - loss: 0.6107 - acc: 0.9638 - 90ms/step
step  90/125 - loss: 0.6120 - acc: 0.9636 - 90ms/step
step 100/125 - loss: 0.5850 - acc: 0.9638 - 90ms/step
step 110/125 - loss: 0.6033 - acc: 0.9646 - 90ms/step
step 120/125 - loss: 0.5901 - acc: 0.9643 - 89ms/step
step 125/125 - loss: 0.5671 - acc: 0.9646 - 87ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6079 - acc: 0.9547 - 87ms/step
step 20/84 - loss: 0.5994 - acc: 0.9566 - 71ms/step
step 30/84 - loss: 0.5997 - acc: 0.9573 - 67ms/step
step 40/84 - loss: 0.5883 - acc: 0.9590 - 65ms/step
step 50/84 - loss: 0.6115 - acc: 0.9583 - 64ms/step
step 60/84 - loss: 0.5703 - acc: 0.9592 - 62ms/step
step 70/84 - loss: 0.6082 - acc: 0.9585 - 61ms/step
step 80/84 - loss: 0.5885 - acc: 0.9590 - 60ms/step
step 84/84 - loss: 0.5901 - acc: 0.9596 - 58ms/step
Eval samples: 10646
Epoch 8/10
step  10/125 - loss: 0.6017 - acc: 0.9648 - 105ms/step
step  20/125 - loss: 0.5801 - acc: 0.9695 - 97ms/step
step  30/125 - loss: 0.5824 - acc: 0.9693 - 92ms/step
step  40/125 - loss: 0.6109 - acc: 0.9689 - 91ms/step
step  50/125 - loss: 0.5899 - acc: 0.9686 - 90ms/step
step  60/125 - loss: 0.5632 - acc: 0.9691 - 89ms/step
step  70/125 - loss: 0.6070 - acc: 0.9694 - 89ms/step
step  80/125 - loss: 0.5853 - acc: 0.9693 - 89ms/step
step  90/125 - loss: 0.6099 - acc: 0.9687 - 90ms/step
step 100/125 - loss: 0.5829 - acc: 0.9694 - 90ms/step
step 110/125 - loss: 0.5884 - acc: 0.9694 - 90ms/step
step 120/125 - loss: 0.6085 - acc: 0.9689 - 90ms/step
step 125/125 - loss: 0.5566 - acc: 0.9688 - 88ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6086 - acc: 0.9563 - 89ms/step
step 20/84 - loss: 0.5964 - acc: 0.9582 - 72ms/step
step 30/84 - loss: 0.5983 - acc: 0.9591 - 67ms/step
step 40/84 - loss: 0.5864 - acc: 0.9605 - 64ms/step
step 50/84 - loss: 0.6144 - acc: 0.9600 - 63ms/step
step 60/84 - loss: 0.5730 - acc: 0.9604 - 63ms/step
step 70/84 - loss: 0.6029 - acc: 0.9597 - 61ms/step
step 80/84 - loss: 0.5849 - acc: 0.9601 - 60ms/step
step 84/84 - loss: 0.5951 - acc: 0.9606 - 58ms/step
Eval samples: 10646
Epoch 9/10
step  10/125 - loss: 0.5861 - acc: 0.9633 - 110ms/step
step  20/125 - loss: 0.5853 - acc: 0.9652 - 100ms/step
step  30/125 - loss: 0.5744 - acc: 0.9695 - 95ms/step
step  40/125 - loss: 0.5724 - acc: 0.9689 - 93ms/step
step  50/125 - loss: 0.5910 - acc: 0.9702 - 92ms/step
step  60/125 - loss: 0.5868 - acc: 0.9701 - 91ms/step
step  70/125 - loss: 0.5931 - acc: 0.9705 - 91ms/step
step  80/125 - loss: 0.5752 - acc: 0.9715 - 91ms/step
step  90/125 - loss: 0.5677 - acc: 0.9720 - 91ms/step
step 100/125 - loss: 0.5638 - acc: 0.9712 - 90ms/step
step 110/125 - loss: 0.5650 - acc: 0.9715 - 90ms/step
step 120/125 - loss: 0.5765 - acc: 0.9715 - 90ms/step
step 125/125 - loss: 0.5677 - acc: 0.9713 - 88ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6051 - acc: 0.9578 - 89ms/step
step 20/84 - loss: 0.5957 - acc: 0.9602 - 73ms/step
step 30/84 - loss: 0.5952 - acc: 0.9607 - 68ms/step
step 40/84 - loss: 0.5860 - acc: 0.9625 - 65ms/step
step 50/84 - loss: 0.6130 - acc: 0.9616 - 64ms/step
step 60/84 - loss: 0.5709 - acc: 0.9620 - 62ms/step
step 70/84 - loss: 0.5985 - acc: 0.9615 - 61ms/step
step 80/84 - loss: 0.5859 - acc: 0.9617 - 60ms/step
step 84/84 - loss: 0.5923 - acc: 0.9622 - 57ms/step
Eval samples: 10646
Epoch 10/10
step  10/125 - loss: 0.5898 - acc: 0.9703 - 109ms/step
step  20/125 - loss: 0.5808 - acc: 0.9699 - 96ms/step
step  30/125 - loss: 0.5680 - acc: 0.9706 - 94ms/step
step  40/125 - loss: 0.5849 - acc: 0.9723 - 94ms/step
step  50/125 - loss: 0.5791 - acc: 0.9741 - 94ms/step
step  60/125 - loss: 0.5722 - acc: 0.9742 - 93ms/step
step  70/125 - loss: 0.5931 - acc: 0.9735 - 92ms/step
step  80/125 - loss: 0.5773 - acc: 0.9725 - 91ms/step
step  90/125 - loss: 0.5967 - acc: 0.9731 - 91ms/step
step 100/125 - loss: 0.5766 - acc: 0.9735 - 90ms/step
step 110/125 - loss: 0.5773 - acc: 0.9739 - 90ms/step
step 120/125 - loss: 0.5805 - acc: 0.9740 - 89ms/step
step 125/125 - loss: 0.5547 - acc: 0.9741 - 88ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6034 - acc: 0.9563 - 85ms/step
step 20/84 - loss: 0.5972 - acc: 0.9598 - 69ms/step
step 30/84 - loss: 0.5895 - acc: 0.9604 - 65ms/step
step 40/84 - loss: 0.5856 - acc: 0.9623 - 63ms/step
step 50/84 - loss: 0.6117 - acc: 0.9614 - 62ms/step
step 60/84 - loss: 0.5724 - acc: 0.9618 - 61ms/step
step 70/84 - loss: 0.5927 - acc: 0.9614 - 60ms/step
step 80/84 - loss: 0.5870 - acc: 0.9614 - 59ms/step
step 84/84 - loss: 0.5899 - acc: 0.9620 - 57ms/step
Eval samples: 10646
save checkpoint at /home/aistudio/checkpoints/gru_v1/final

model.save('finetuning/gru/model', training=True)

VisualDL可视化

模型评估

• 调用model.evaluate一键评估模型

• 参数：

• eval_data (Dataset|DataLoader) - 一个可迭代的数据源，推荐给定一个 paddle.io.Dataset 或 paddle.io.Dataloader 的实例。默认值：None。

results = model.evaluate(dev_loader)
print("Finally test acc: %.5f" % results['acc'])

Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/84 - loss: 0.6034 - acc: 0.9563 - 88ms/step
step 20/84 - loss: 0.5972 - acc: 0.9598 - 71ms/step
step 30/84 - loss: 0.5895 - acc: 0.9604 - 67ms/step
step 40/84 - loss: 0.5856 - acc: 0.9623 - 64ms/step
step 50/84 - loss: 0.6117 - acc: 0.9614 - 62ms/step
step 60/84 - loss: 0.5724 - acc: 0.9618 - 61ms/step
step 70/84 - loss: 0.5927 - acc: 0.9614 - 60ms/step
step 80/84 - loss: 0.5870 - acc: 0.9614 - 58ms/step
step 84/84 - loss: 0.5899 - acc: 0.9620 - 56ms/step
Eval samples: 10646
Finally test acc: 0.96196

模型

• 调用model.predict进行预测。

• 参数

• test_data (Dataset|DataLoader): 一个可迭代的数据源，推荐给定一个paddle.io.Dataset 或 paddle.io.Dataloader 的实例。默认值：None。

import numpy as np

label_map = {0: 'negative', 1: 'positive'}
results = model.predict(test_loader, batch_size=64)[0]
predictions = []

for batch_probs in results:
    # 映射分类label
    idx = np.argmax(batch_probs, axis=-1)
    idx = idx.tolist()
    labels = [label_map[i] for i in idx]
    predictions.extend(labels)

# 看看预测数据前5个样例分类结果
for idx, data in enumerate(test_ds.data[:5]):
    print('Data: {} \t Label: {}'.format(data[0], predictions[idx]))

Predict begin...
step 42/42 [==============================] - ETA: 3s - 89ms/ste - ETA: 3s - 86ms/ste - ETA: 3s - 96ms/ste - ETA: 3s - 95ms/ste - ETA: 2s - 87ms/ste - ETA: 2s - 83ms/ste - ETA: 2s - 80ms/ste - ETA: 2s - 78ms/ste - ETA: 1s - 75ms/ste - ETA: 1s - 74ms/ste - ETA: 1s - 74ms/ste - ETA: 1s - 70ms/ste - ETA: 1s - 68ms/ste - ETA: 0s - 70ms/ste - ETA: 0s - 71ms/ste - ETA: 0s - 71ms/ste - ETA: 0s - 71ms/ste - ETA: 0s - 70ms/ste - ETA: 0s - 68ms/ste - ETA: 0s - 66ms/ste - 64ms/step          
Predict samples: 5353
Data: 楼面经理服务态度极差，等位和埋单都差，楼面小妹还挺好 	 Label: negative
Data: 欺负北方人没吃过鲍鱼是怎么着？简直敷衍到可笑的程度，团购连青菜都是两人份？！难吃到死，菜色还特别可笑，什么时候粤菜的小菜改成拍黄瓜了？！把团购客人当傻子，可这满大厅的傻子谁还会再来？！ 	 Label: negative
Data: 如果大家有时间而且不怕麻烦的话可以去这里试试，点一个饭等左2个钟，没错！是两个钟！期间催了n遍都说马上到，结果？呵呵。乳鸽的味道，太咸，可能不新鲜吧……要用重口味盖住异味。上菜超级慢！中途还搞什么表演，麻烦有人手的话就上菜啊，表什么演？！？！要大家饿着看表演吗？最后结账还算错单，我真心服了……有一种店叫不会有下次，大概就是指它吧 	 Label: negative
Data: 偌大的一个大厅就一个人点菜，点菜速度超级慢，菜牌上多个菜停售，连续点了两个没标停售的菜也告知没有，粥上来是凉的，榴莲酥火大了，格格肉超级油腻而且咸?????? 	 Label: negative
Data: 泥撕雞超級好吃！！！吃了一個再叫一個還想打包的節奏！ 	 Label: positive

更换三分类数据集进行测试

三分类除了涉及到positive和negative两种情感外，还有一种neural情感，从原始数据集中可以提取到有语义转折的句子，“然而”，“但”都是关键词。从而可以得到3份不同语义的数据集。

数据准备

class Classifier3Dataset(paddle.io.Dataset):
    def __init__(self, data):
        super(Classifier3Dataset, self).__init__()
        self.data = data

    def __getitem__(self, idx):
        return self.data[idx]

    def __len__(self):
        return len(self.data)
        
    def get_labels(self):
        return ["0", "1", "2"]

def txt_to_list(file_name):
    res_list = []
    for line in open(file_name):
        res_list.append(line.strip().split('\t'))
    return res_list

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

trainlst = txt_to_list('./my_data/train.txt')
devlst = txt_to_list('./my_data/dev.txt')
testlst = txt_to_list('./my_data/test.txt')

# 通过get_datasets()函数，将list数据转换为dataset。
# get_datasets()可接收[list]参数，或[str]参数，根据自定义数据集的写法自由选择。
# train_ds, dev_ds, test_ds = ppnlp.datasets.ChnSentiCorp.get_datasets(['train', 'dev', 'test'])
train_ds, dev_ds, test_ds = Classifier3Dataset.get_datasets([trainlst, devlst, testlst])

label_list = train_ds.get_labels()
print(label_list)

for sent, label in train_ds[:5]:
    print (sent, label)

tmp = ppnlp.datasets.ChnSentiCorp('train')
tmp1 = ppnlp.datasets.MapDatasetWrapper(tmp)
print(type(tmp), type(tmp1))
for sent, label in tmp1[:5]:
    print(sent, label)

# Reads data and generates mini-batches.
trans_fn = partial(
    convert_example,
    vocab=vocab,
    unk_token_id=vocab.get('[UNK]', 1),
    is_test=False)

# 将读入的数据batch化处理，便于模型batch化运算。
# batch中的每个句子将会padding到这个batch中的文本最大长度batch_max_seq_len。
# 当文本长度大于batch_max_seq时，将会截断到batch_max_seq_len；当文本长度小于batch_max_seq时，将会padding补齐到batch_max_seq_len.
batchify_fn = lambda samples, fn=Tuple(
    Pad(axis=0, pad_val=vocab['[PAD]']),  # input_ids
    Stack(dtype="int64"),  # seq len
    Stack(dtype="int64")  # label
): [data for data in fn(samples)]

# batch_size改为256
train_loader = create_dataloader(
    train_ds,
    trans_fn=trans_fn,
    batch_size=256,
    mode='train',
    batchify_fn=batchify_fn)
    
dev_loader = create_dataloader(
    dev_ds,
    trans_fn=trans_fn,
    batch_size=256,
    mode='validation',
    batchify_fn=batchify_fn)

test_loader = create_dataloader(
    test_ds,
    trans_fn=trans_fn,
    batch_size=256,
    mode='test',
    batchify_fn=batchify_fn)

['0', '1', '2']
环境不错，叉烧包小孩爱吃，三色煎糕很一般 2
刚来的时候让我们等位置，我们就在门口等了十分钟左右，没有见到有人离开，然后有工作人员让我们上来二楼，上来后看到有十来桌是没有人的，既然有位置，为什么非得让我们在门口等位！！！本以为有座位后就可以马上吃饭了，让人内心崩溃的是点菜就等了十分钟才有人有空过来理我们。上菜更是郁闷，端来了一锅猪肚鸡，但是没人开火，要点调味料也没人管，服务质量太差，之前来过一次觉得还可以，这次让我再也不想来这家店了！真心失望 0
口味还可以服务真的差到爆啊我来过45次真的次次都只给差评东西确实不错但你看看你们的服务还收服务费我的天干蒸什么的60块比太古汇翠园还贵主要是没人收台没人倒水谁还要自己倒我的天给你服务费还什么都自己干我接受不了钱花了服务没有实在不行 0
出品不错老字号就是好有山有水有树有鱼赞 1
在江南大道这间～服务态度好差，食物出品平凡，应该唔会再去了 1
 
选择珠江花园的原因就是方便，有电动扶梯直接到达海边，周围餐馆、食廊、商场、超市、摊位一应俱全。酒店装修一般，但还算整洁。 泳池在大堂的屋顶，因此很小，不过女儿倒是喜欢。 包的早餐是西式的，还算丰富。 服务吗，一般 1
15.4寸笔记本的键盘确实爽，基本跟台式机差不多了，蛮喜欢数字小键盘，输数字特方便，样子也很美观，做工也相当不错 1
房间太小。其他的都一般。。。。。。。。。 0
1.接电源没有几分钟,电源适配器热的不行. 2.摄像头用不起来. 3.机盖的钢琴漆，手不能摸，一摸一个印. 4.硬盘分区不好办. 0
今天才知道这书还有第6卷,真有点郁闷:为什么同一套书有两种版本呢?当当网是不是该跟出版社商量商量,单独出个第6卷,让我们的孩子不会有所遗憾。 1

训练感觉有点过拟合 加大了batch_size 128->256, hidden_size 96->128 同时引入了dropout=0.2

vocab = load_vocab('./senta_word_dict.txt')

model = GRUModel(
        vocab_size=len(vocab),
        num_classes=len(label_list),
        direction='bidirectional',
        padding_idx=vocab['[PAD]'],
        dropout_rate=0.2,
        fc_hidden_size=128) # out -> 128 -> 3
        
model = paddle.Model(model)

optimizer = paddle.optimizer.Adam(
        parameters=model.parameters(), learning_rate=5e-5)

loss = paddle.nn.CrossEntropyLoss()

metric = paddle.metric.Accuracy()

model.prepare(optimizer, loss, metric)

# 设置visualdl路径
log_dir = './visualdl/gru_3'
callbacks = paddle.callbacks.VisualDL(log_dir=log_dir)

model.fit(train_loader, 
            dev_loader, 
            epochs=10, 
            save_dir='./checkpoints/gru_3', 
            save_freq=5, 
            callbacks=callbacks)

The loss value printed in the log is the current step, and the metric is the average value of previous step.
Epoch 1/10
step  10/126 - loss: 1.0978 - acc: 0.3398 - 208ms/step
step  20/126 - loss: 1.0996 - acc: 0.3350 - 181ms/step
step  30/126 - loss: 1.0976 - acc: 0.3408 - 173ms/step
step  40/126 - loss: 1.0953 - acc: 0.3434 - 169ms/step
step  50/126 - loss: 1.0956 - acc: 0.3451 - 165ms/step
step  60/126 - loss: 1.0965 - acc: 0.3449 - 162ms/step
step  70/126 - loss: 1.0983 - acc: 0.3444 - 161ms/step
step  80/126 - loss: 1.0960 - acc: 0.3454 - 160ms/step
step  90/126 - loss: 1.0955 - acc: 0.3456 - 159ms/step
step 100/126 - loss: 1.0925 - acc: 0.3445 - 158ms/step
step 110/126 - loss: 1.0915 - acc: 0.3480 - 158ms/step
step 120/126 - loss: 1.0898 - acc: 0.3595 - 158ms/step
step 126/126 - loss: 1.0895 - acc: 0.3675 - 154ms/step
save checkpoint at /home/aistudio/checkpoints/gru_3/0
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 1.0886 - acc: 0.5535 - 129ms/step
step 16/16 - loss: 1.0892 - acc: 0.5592 - 117ms/step
Eval samples: 3968
Epoch 2/10
step  10/126 - loss: 1.0875 - acc: 0.5953 - 200ms/step
step  20/126 - loss: 1.0836 - acc: 0.5965 - 177ms/step
step  30/126 - loss: 1.0774 - acc: 0.5948 - 171ms/step
step  40/126 - loss: 1.0721 - acc: 0.5929 - 166ms/step
step  50/126 - loss: 1.0589 - acc: 0.5895 - 164ms/step
step  60/126 - loss: 1.0572 - acc: 0.5824 - 161ms/step
step  70/126 - loss: 1.0510 - acc: 0.5797 - 161ms/step
step  80/126 - loss: 1.0356 - acc: 0.5745 - 160ms/step
step  90/126 - loss: 0.9759 - acc: 0.5755 - 160ms/step
step 100/126 - loss: 0.9620 - acc: 0.5741 - 160ms/step
step 110/126 - loss: 0.9256 - acc: 0.5756 - 159ms/step
step 120/126 - loss: 0.9095 - acc: 0.5785 - 158ms/step
step 126/126 - loss: 0.8743 - acc: 0.5817 - 154ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.8846 - acc: 0.6320 - 138ms/step
step 16/16 - loss: 0.9308 - acc: 0.6308 - 123ms/step
Eval samples: 3968
Epoch 3/10
step  10/126 - loss: 0.9291 - acc: 0.6621 - 197ms/step
step  20/126 - loss: 0.8573 - acc: 0.6590 - 176ms/step
step  30/126 - loss: 0.9236 - acc: 0.6594 - 168ms/step
step  40/126 - loss: 0.8650 - acc: 0.6684 - 164ms/step
step  50/126 - loss: 0.8187 - acc: 0.6738 - 162ms/step
step  60/126 - loss: 0.8715 - acc: 0.6797 - 161ms/step
step  70/126 - loss: 0.8505 - acc: 0.6854 - 160ms/step
step  80/126 - loss: 0.8257 - acc: 0.6878 - 159ms/step
step  90/126 - loss: 0.8558 - acc: 0.6905 - 159ms/step
step 100/126 - loss: 0.8403 - acc: 0.6926 - 159ms/step
step 110/126 - loss: 0.7945 - acc: 0.6958 - 158ms/step
step 120/126 - loss: 0.8609 - acc: 0.6980 - 158ms/step
step 126/126 - loss: 0.8882 - acc: 0.6992 - 154ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.8116 - acc: 0.7141 - 134ms/step
step 16/16 - loss: 0.8394 - acc: 0.7193 - 119ms/step
Eval samples: 3968
Epoch 4/10
step  10/126 - loss: 0.7935 - acc: 0.7508 - 231ms/step
step  20/126 - loss: 0.7982 - acc: 0.7494 - 190ms/step
step  30/126 - loss: 0.8068 - acc: 0.7480 - 177ms/step
step  40/126 - loss: 0.7855 - acc: 0.7446 - 172ms/step
step  50/126 - loss: 0.8038 - acc: 0.7469 - 168ms/step
step  60/126 - loss: 0.8076 - acc: 0.7500 - 165ms/step
step  70/126 - loss: 0.7809 - acc: 0.7518 - 165ms/step
step  80/126 - loss: 0.7585 - acc: 0.7547 - 163ms/step
step  90/126 - loss: 0.7997 - acc: 0.7564 - 162ms/step
step 100/126 - loss: 0.8227 - acc: 0.7561 - 161ms/step
step 110/126 - loss: 0.7757 - acc: 0.7562 - 160ms/step
step 120/126 - loss: 0.7927 - acc: 0.7562 - 160ms/step
step 126/126 - loss: 0.7383 - acc: 0.7566 - 155ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7797 - acc: 0.7344 - 131ms/step
step 16/16 - loss: 0.8309 - acc: 0.7424 - 118ms/step
Eval samples: 3968
Epoch 5/10
step  10/126 - loss: 0.7490 - acc: 0.7945 - 204ms/step
step  20/126 - loss: 0.7892 - acc: 0.7848 - 178ms/step
step  30/126 - loss: 0.7733 - acc: 0.7818 - 170ms/step
step  40/126 - loss: 0.7219 - acc: 0.7829 - 165ms/step
step  50/126 - loss: 0.7361 - acc: 0.7833 - 163ms/step
step  60/126 - loss: 0.7994 - acc: 0.7804 - 161ms/step
step  70/126 - loss: 0.7618 - acc: 0.7810 - 161ms/step
step  80/126 - loss: 0.7607 - acc: 0.7832 - 161ms/step
step  90/126 - loss: 0.7378 - acc: 0.7851 - 160ms/step
step 100/126 - loss: 0.7430 - acc: 0.7852 - 159ms/step
step 110/126 - loss: 0.7676 - acc: 0.7856 - 159ms/step
step 120/126 - loss: 0.7475 - acc: 0.7865 - 158ms/step
step 126/126 - loss: 0.7938 - acc: 0.7866 - 154ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7679 - acc: 0.7441 - 132ms/step
step 16/16 - loss: 0.8193 - acc: 0.7505 - 118ms/step
Eval samples: 3968
Epoch 6/10
step  10/126 - loss: 0.7762 - acc: 0.8020 - 200ms/step
step  20/126 - loss: 0.7412 - acc: 0.8004 - 177ms/step
step  30/126 - loss: 0.7627 - acc: 0.8049 - 169ms/step
step  40/126 - loss: 0.7367 - acc: 0.8074 - 165ms/step
step  50/126 - loss: 0.7610 - acc: 0.8068 - 163ms/step
step  60/126 - loss: 0.7663 - acc: 0.8056 - 161ms/step
step  70/126 - loss: 0.7318 - acc: 0.8050 - 160ms/step
step  80/126 - loss: 0.7516 - acc: 0.8081 - 159ms/step
step  90/126 - loss: 0.7567 - acc: 0.8073 - 158ms/step
step 100/126 - loss: 0.7430 - acc: 0.8080 - 159ms/step
step 110/126 - loss: 0.7549 - acc: 0.8069 - 159ms/step
step 120/126 - loss: 0.7199 - acc: 0.8068 - 159ms/step
step 126/126 - loss: 0.7319 - acc: 0.8067 - 155ms/step
save checkpoint at /home/aistudio/checkpoints/gru_3/5
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7777 - acc: 0.7492 - 131ms/step
step 16/16 - loss: 0.8116 - acc: 0.7538 - 118ms/step
Eval samples: 3968
Epoch 7/10
step  10/126 - loss: 0.7169 - acc: 0.8000 - 195ms/step
step  20/126 - loss: 0.7229 - acc: 0.8127 - 176ms/step
step  30/126 - loss: 0.7203 - acc: 0.8156 - 170ms/step
step  40/126 - loss: 0.7103 - acc: 0.8191 - 166ms/step
step  50/126 - loss: 0.6762 - acc: 0.8223 - 163ms/step
step  60/126 - loss: 0.7651 - acc: 0.8215 - 161ms/step
step  70/126 - loss: 0.7337 - acc: 0.8228 - 160ms/step
step  80/126 - loss: 0.7348 - acc: 0.8228 - 160ms/step
step  90/126 - loss: 0.7023 - acc: 0.8231 - 160ms/step
step 100/126 - loss: 0.7326 - acc: 0.8229 - 159ms/step
step 110/126 - loss: 0.7133 - acc: 0.8237 - 159ms/step
step 120/126 - loss: 0.7220 - acc: 0.8237 - 159ms/step
step 126/126 - loss: 0.6864 - acc: 0.8238 - 155ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7588 - acc: 0.7543 - 133ms/step
step 16/16 - loss: 0.8104 - acc: 0.7581 - 118ms/step
Eval samples: 3968
Epoch 8/10
step  10/126 - loss: 0.7200 - acc: 0.8258 - 196ms/step
step  20/126 - loss: 0.7397 - acc: 0.8305 - 176ms/step
step  30/126 - loss: 0.7124 - acc: 0.8366 - 170ms/step
step  40/126 - loss: 0.7328 - acc: 0.8372 - 167ms/step
step  50/126 - loss: 0.7100 - acc: 0.8402 - 164ms/step
step  60/126 - loss: 0.6964 - acc: 0.8393 - 162ms/step
step  70/126 - loss: 0.7164 - acc: 0.8388 - 163ms/step
step  80/126 - loss: 0.7102 - acc: 0.8390 - 165ms/step
step  90/126 - loss: 0.7163 - acc: 0.8385 - 165ms/step
step 100/126 - loss: 0.7125 - acc: 0.8386 - 165ms/step
step 110/126 - loss: 0.7342 - acc: 0.8384 - 164ms/step
step 120/126 - loss: 0.7394 - acc: 0.8379 - 163ms/step
step 126/126 - loss: 0.6978 - acc: 0.8382 - 158ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7574 - acc: 0.7547 - 140ms/step
step 16/16 - loss: 0.8107 - acc: 0.7558 - 127ms/step
Eval samples: 3968
Epoch 9/10
step  10/126 - loss: 0.6952 - acc: 0.8582 - 208ms/step
step  20/126 - loss: 0.7105 - acc: 0.8520 - 178ms/step
step  30/126 - loss: 0.7099 - acc: 0.8492 - 170ms/step
step  40/126 - loss: 0.7073 - acc: 0.8480 - 167ms/step
step  50/126 - loss: 0.7057 - acc: 0.8480 - 163ms/step
step  60/126 - loss: 0.6870 - acc: 0.8474 - 160ms/step
step  70/126 - loss: 0.6980 - acc: 0.8481 - 159ms/step
step  80/126 - loss: 0.6783 - acc: 0.8481 - 158ms/step
step  90/126 - loss: 0.7179 - acc: 0.8467 - 158ms/step
step 100/126 - loss: 0.6865 - acc: 0.8482 - 158ms/step
step 110/126 - loss: 0.6836 - acc: 0.8488 - 158ms/step
step 120/126 - loss: 0.7049 - acc: 0.8488 - 158ms/step
step 126/126 - loss: 0.7239 - acc: 0.8485 - 154ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7657 - acc: 0.7578 - 134ms/step
step 16/16 - loss: 0.8159 - acc: 0.7608 - 121ms/step
Eval samples: 3968
Epoch 10/10
step  10/126 - loss: 0.6735 - acc: 0.8539 - 202ms/step
step  20/126 - loss: 0.7280 - acc: 0.8547 - 178ms/step
step  30/126 - loss: 0.7055 - acc: 0.8539 - 170ms/step
step  40/126 - loss: 0.7137 - acc: 0.8554 - 165ms/step
step  50/126 - loss: 0.6825 - acc: 0.8568 - 162ms/step
step  60/126 - loss: 0.7012 - acc: 0.8569 - 161ms/step
step  70/126 - loss: 0.7151 - acc: 0.8573 - 160ms/step
step  80/126 - loss: 0.7342 - acc: 0.8571 - 158ms/step
step  90/126 - loss: 0.6726 - acc: 0.8569 - 158ms/step
step 100/126 - loss: 0.7126 - acc: 0.8577 - 157ms/step
step 110/126 - loss: 0.6896 - acc: 0.8578 - 158ms/step
step 120/126 - loss: 0.6932 - acc: 0.8581 - 158ms/step
step 126/126 - loss: 0.7185 - acc: 0.8580 - 153ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7604 - acc: 0.7590 - 134ms/step
step 16/16 - loss: 0.8171 - acc: 0.7626 - 121ms/step
Eval samples: 3968
save checkpoint at /home/aistudio/checkpoints/gru_3/final

# 由于acc还在上升 再训练5个epoch

model.fit(train_loader, 
            dev_loader, 
            epochs=5, 
            save_dir='./checkpoints/gru_3', 
            save_freq=5, 
            callbacks=callbacks)

The loss value printed in the log is the current step, and the metric is the average value of previous step.
Epoch 1/5
step  10/126 - loss: 0.6639 - acc: 0.8621 - 205ms/step
step  20/126 - loss: 0.6666 - acc: 0.8641 - 184ms/step
step  30/126 - loss: 0.6868 - acc: 0.8629 - 178ms/step
step  40/126 - loss: 0.6952 - acc: 0.8665 - 174ms/step
step  50/126 - loss: 0.6786 - acc: 0.8670 - 172ms/step
step  60/126 - loss: 0.6718 - acc: 0.8674 - 172ms/step
step  70/126 - loss: 0.7139 - acc: 0.8652 - 170ms/step
step  80/126 - loss: 0.7146 - acc: 0.8647 - 168ms/step
step  90/126 - loss: 0.6890 - acc: 0.8660 - 167ms/step
step 100/126 - loss: 0.7247 - acc: 0.8655 - 165ms/step
step 110/126 - loss: 0.6748 - acc: 0.8660 - 164ms/step
step 120/126 - loss: 0.6856 - acc: 0.8658 - 163ms/step
step 126/126 - loss: 0.7343 - acc: 0.8657 - 159ms/step
save checkpoint at /home/aistudio/checkpoints/gru_3/0
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7668 - acc: 0.7523 - 130ms/step
step 16/16 - loss: 0.8106 - acc: 0.7581 - 116ms/step
Eval samples: 3968
Epoch 2/5
step  10/126 - loss: 0.6910 - acc: 0.8793 - 201ms/step
step  20/126 - loss: 0.6806 - acc: 0.8732 - 178ms/step
step  30/126 - loss: 0.6818 - acc: 0.8702 - 168ms/step
step  40/126 - loss: 0.7119 - acc: 0.8700 - 164ms/step
step  50/126 - loss: 0.6921 - acc: 0.8701 - 163ms/step
step  60/126 - loss: 0.6909 - acc: 0.8715 - 162ms/step
step  70/126 - loss: 0.7123 - acc: 0.8719 - 162ms/step
step  80/126 - loss: 0.6695 - acc: 0.8720 - 161ms/step
step  90/126 - loss: 0.6982 - acc: 0.8723 - 160ms/step
step 100/126 - loss: 0.7042 - acc: 0.8723 - 160ms/step
step 110/126 - loss: 0.6810 - acc: 0.8721 - 160ms/step
step 120/126 - loss: 0.7072 - acc: 0.8723 - 159ms/step
step 126/126 - loss: 0.6921 - acc: 0.8724 - 155ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7636 - acc: 0.7559 - 144ms/step
step 16/16 - loss: 0.8205 - acc: 0.7601 - 127ms/step
Eval samples: 3968
Epoch 3/5
step  10/126 - loss: 0.6746 - acc: 0.8824 - 208ms/step
step  20/126 - loss: 0.6689 - acc: 0.8826 - 183ms/step
step  30/126 - loss: 0.6737 - acc: 0.8811 - 174ms/step
step  40/126 - loss: 0.6970 - acc: 0.8804 - 169ms/step
step  50/126 - loss: 0.6662 - acc: 0.8800 - 165ms/step
step  60/126 - loss: 0.6893 - acc: 0.8801 - 163ms/step
step  70/126 - loss: 0.6819 - acc: 0.8804 - 163ms/step
step  80/126 - loss: 0.6664 - acc: 0.8803 - 162ms/step
step  90/126 - loss: 0.6590 - acc: 0.8793 - 161ms/step
step 100/126 - loss: 0.6681 - acc: 0.8790 - 161ms/step
step 110/126 - loss: 0.6805 - acc: 0.8778 - 160ms/step
step 120/126 - loss: 0.7028 - acc: 0.8781 - 160ms/step
step 126/126 - loss: 0.6371 - acc: 0.8775 - 156ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7640 - acc: 0.7527 - 134ms/step
step 16/16 - loss: 0.8242 - acc: 0.7586 - 120ms/step
Eval samples: 3968
Epoch 4/5
step  10/126 - loss: 0.6708 - acc: 0.8879 - 200ms/step
step  20/126 - loss: 0.6879 - acc: 0.8863 - 177ms/step
step  30/126 - loss: 0.6686 - acc: 0.8828 - 168ms/step
step  40/126 - loss: 0.6895 - acc: 0.8804 - 164ms/step
step  50/126 - loss: 0.6907 - acc: 0.8794 - 162ms/step
step  60/126 - loss: 0.6387 - acc: 0.8808 - 161ms/step
step  70/126 - loss: 0.6723 - acc: 0.8815 - 162ms/step
step  80/126 - loss: 0.6770 - acc: 0.8814 - 161ms/step
step  90/126 - loss: 0.6581 - acc: 0.8829 - 160ms/step
step 100/126 - loss: 0.6441 - acc: 0.8826 - 160ms/step
step 110/126 - loss: 0.6712 - acc: 0.8829 - 159ms/step
step 120/126 - loss: 0.6454 - acc: 0.8835 - 159ms/step
step 126/126 - loss: 0.6550 - acc: 0.8836 - 155ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7569 - acc: 0.7500 - 139ms/step
step 16/16 - loss: 0.8168 - acc: 0.7573 - 123ms/step
Eval samples: 3968
Epoch 5/5
step  10/126 - loss: 0.6532 - acc: 0.8828 - 193ms/step
step  20/126 - loss: 0.6534 - acc: 0.8863 - 175ms/step
step  30/126 - loss: 0.6523 - acc: 0.8850 - 169ms/step
step  40/126 - loss: 0.6625 - acc: 0.8868 - 166ms/step
step  50/126 - loss: 0.6752 - acc: 0.8874 - 164ms/step
step  60/126 - loss: 0.6603 - acc: 0.8874 - 163ms/step
step  70/126 - loss: 0.6711 - acc: 0.8879 - 162ms/step
step  80/126 - loss: 0.6867 - acc: 0.8875 - 161ms/step
step  90/126 - loss: 0.6505 - acc: 0.8882 - 160ms/step
step 100/126 - loss: 0.6640 - acc: 0.8889 - 158ms/step
step 110/126 - loss: 0.6786 - acc: 0.8891 - 158ms/step
step 120/126 - loss: 0.6796 - acc: 0.8883 - 159ms/step
step 126/126 - loss: 0.6795 - acc: 0.8881 - 155ms/step
Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7610 - acc: 0.7520 - 136ms/step
step 16/16 - loss: 0.8282 - acc: 0.7571 - 121ms/step
Eval samples: 3968
save checkpoint at /home/aistudio/checkpoints/gru_3/final

VisualDL可视化

评估

results = model.evaluate(dev_loader)
print("Finally test acc: %.5f" % results['acc'])

Eval begin...
The loss value printed in the log is the current batch, and the metric is the average value of previous step.
step 10/16 - loss: 0.7610 - acc: 0.7520 - 131ms/step
step 16/16 - loss: 0.8282 - acc: 0.7571 - 117ms/step
Eval samples: 3968
Finally test acc: 0.75706

预测

import numpy as np

label_map = {0: 'negative', 1: 'positive', 2:'neural'}
results = model.predict(test_loader, batch_size=64)[0]
predictions = []

for batch_probs in results:
    # 映射分类label
    idx = np.argmax(batch_probs, axis=-1)
    idx = idx.tolist()
    labels = [label_map[i] for i in idx]
    predictions.extend(labels)

# 看看预测数据前5个样例分类结果
for idx, data in enumerate(test_ds.data[:10]):
    print('Data: {} \t Label: {}'.format(data[0], predictions[idx]))

Predict begin...
step 16/16 [==============================] - ETA: 2s - 191ms/st - ETA: 1s - 162ms/st - ETA: 1s - 155ms/st - ETA: 1s - 146ms/st - ETA: 0s - 144ms/st - ETA: 0s - 141ms/st - ETA: 0s - 138ms/st - 128ms/step          
Predict samples: 3986
Data: 出品与环境都算可以吧，服务亦过得去。 	 Label: positive
Data: 菜真是一般般滴，除了贵没啥优点… 	 Label: negative
Data: 第一次是朋友约的喝茶，两层楼都是坐满满客人，我们一行四个人，东西味道很正宗的广式茶点，最爱红米肠，还有乳鸽...网上团购真心觉得划算，在惠福路靠近北京路步行街总觉得东西会很贵，买单时让我感觉很意外四个人才一百多两百，旁边的一些客人也在说好划算哦，因为味道正宗，服务好，性价比高，这就是它为什么现在很多人选择它的原因，朋友家人聚餐的又一个不错的选择... 	 Label: positive
Data: 抱住期待来…有d失望咯…点左bb猪，乳鸽松，上汤豆苗，薄撑，燕窝鹧鸪粥bb猪一般，无好好吃乳鸽松其实不错，不过实在太咸啦！！！！而且好多味精！吃完点饮水都唔够！上汤豆苗d豆苗唔知系乜豆苗，d汤底几好，不过都系咸左小小薄撑边系得得地啊，系唔多得！皮不够烟韧，又唔够脆，特別系在上边的，馅糖太多，花生系咸的！！！！！燕窝粥我无食…环境ok，就系有点嘈。服务一般，不太值10%。总结一句无乜动力令我再帮衬，实在太咸太甜啦！！！！！我宜家仲想买支水一口气喝掉半瓶！！！！！ 	 Label: positive
Data: 点了一个小时的单，还没有上，上了的东西还是生的，跟服务员理论还一脸臭相。差！ 	 Label: negative
Data: 服务贴心。不过鲍鱼没有想象中的那么大，团购划算不少！每人一份感觉挺丰盛的。已经团了两次了，自己觉得好还带家人去尝试。 	 Label: neural
Data: 环境可以就是吃饭时间人比较多，出品比较精致分量男生可能少点 	 Label: positive
Data: 差评！非繁忙时段，两位不可以坐卡座，卡座还有一大片是空着的，二人桌只剩一张而且对着门口不想坐，看到别人两位可以坐卡座，就我们两位不可以坐！差评！果断走！不是只有你一间吃饭的！ 	 Label: negative
Data: 好久都没去过广州塔和珠江边附近走走了，又刚好有个灯光节，在加上抽奖抽中了穿粤传奇的【4D魔幻灯光秀】，就必须要过去看看咯，感受一下灯光节的夜景??不知不觉看完4D魔幻灯光秀都快到九点了，就不如来个宵夜吧。走着走着看到了广州塔下面，有一件叫做赏点点心喝茶的好地方。哈哈…于是我们两个人决定去那吃宵夜咯，增肥的节奏哇！一进去很有广州喝茶的感觉，连装修风格也很像之前茶酒的感觉，好喜欢这种风格哦，不错不错。我们一坐下就看了菜单，看到价格都不是很贵，还是挺实惠的。早上是十一点前埋单，还有得打折呢。我们两个人一共点了五样食物。说真的，他家的出品还真的不错，和我经常去的“点都德”相比来说，他家的份量多点，味道也好点，有一些价格上还便宜先。不管白天还是晚上，喝完茶以后，走走珠江边，看看广州塔，还真是个不错的选择。 	 Label: neural
Data: 晚市点了一条鱼，一份翡翠饺子，一个咖喱牛筋。饺子很一般很一般，鱼很难吃，牛筋少且难吃。两个人花了178没一道菜及格的也是醉了，因为太难吃了到最后菜都没吃完，不知道是不是我们太***丝了到这里都要点上百的菜。反正除了环境好，感觉就只剩招牌了。 	 Label: negative

预训练模型

近年来随着深度学习的发展，模型参数数量飞速增长，为了训练这些参数，需要更大的数据集来避免过拟合。然而，对于大部分NLP任务来说，构建大规模的标注数据集成本过高，非常困难，特别是对于句法和语义相关的任务。相比之下，大规模的未标注语料库的构建则相对容易。最近的研究表明，基于大规模未标注语料库的预训练模型（Pretrained Models, PTM) 能够习得通用的语言表示，将预训练模型Fine-tune到下游任务，能够获得出色的表现。另外，预训练模型能够避免从零开始训练模型。

参考如何通过预训练模型Fine-tune下游任务

图2：预训练模型一览，图片来源：https://github.com/thunlp/PLMpapers

PaddleNLP一键加载预训练模型

情感分析本质是一个文本分类任务，PaddleNLP对于各种预训练模型已经内置了对于下游任务-文本分类的Fine-tune网络。以下教程ERNIE为例，介绍如何将预训练模型Fine-tune完成文本分类任务。

• paddlenlp.transformers.ErnieModel()

  一行代码即可加载预训练模型ERNIE。

• paddlenlp.transformers.ErnieForSequenceClassification()

  一行代码即可加载预训练模型ERNIE用于文本分类任务的Fine-tune网络。
  其在ERNIE模型后拼接上一个全连接网络（Full Connected）进行分类。

• paddlenlp.transformers.ErnieForSequenceClassification.from_pretrained()

  只需指定想要使用的模型名称和文本分类的类别数即可完成网络定义。

  PaddleNLP不仅支持ERNIE预训练模型，还支持BERT、RoBERTa、Electra等预训练模型。

调用ppnlp.transformers.ErnieTokenizer进行数据处理

预训练模型ERNIE对中文数据的处理是以字为单位。PaddleNLP对于各种预训练模型已经内置了相应的tokenizer。指定想要使用的模型名字即可加载对应的tokenizer。

tokenizer作用为将原始输入文本转化成模型model可以接受的输入数据形式。

ERNIE模型框架示意图

# 使用ernie预训练模型
# ernie
model = ppnlp.transformers.ErnieForSequenceClassification.from_pretrained('ernie-1.0', num_classes=2)
tokenizer = ppnlp.transformers.ErnieTokenizer.from_pretrained('ernie-1.0')

# ernie-tiny
# model = ppnlp.transformers.ErnieForSequenceClassification.rom_pretrained('ernie-tiny',num_classes=2))
# tokenizer = ppnlp.transformers.ErnieTinyTokenizer.from_pretrained('ernie-tiny')


# 使用bert预训练模型
# bert-base-chinese
# model = ppnlp.transformers.BertForSequenceClassification.from_pretrained('bert-base-chinese', num_class=2)
# tokenizer = ppnlp.transformers.BertTokenizer.from_pretrained('bert-base-chinese')

# bert-wwm-chinese
# model = ppnlp.transformers.BertForSequenceClassification.from_pretrained('bert-wwm-chinese', num_class=2)
# tokenizer = ppnlp.transformers.BertTokenizer.from_pretrained('bert-wwm-chinese')

# bert-wwm-ext-chinese
# model = ppnlp.transformers.BertForSequenceClassification.from_pretrained('bert-wwm-ext-chinese', num_class=2)
# tokenizer = ppnlp.transformers.BertTokenizer.from_pretrained('bert-wwm-ext-chinese')


# 使用roberta预训练模型
# roberta-wwm-ext
# model = ppnlp.transformers.RobertaForSequenceClassification.from_pretrained('roberta-wwm-ext', num_class=2)
# tokenizer = ppnlp.transformers.RobertaTokenizer.from_pretrained('roberta-wwm-ext')

# roberta-wwm-ext
# model = ppnlp.transformers.RobertaForSequenceClassification.from_pretrained('roberta-wwm-ext-large', num_class=2)
# tokenizer = ppnlp.transformers.RobertaTokenizer.from_pretrained('roberta-wwm-ext-large')

ForSequenceClassification.from_pretrained('roberta-wwm-ext', num_class=2)
# tokenizer = ppnlp.transformers.RobertaTokenizer.from_pretrained('roberta-wwm-ext')

# roberta-wwm-ext
# model = ppnlp.transformers.RobertaForSequenceClassification.from_pretrained('roberta-wwm-ext-large', num_class=2)
# tokenizer = ppnlp.transformers.RobertaTokenizer.from_pretrained('roberta-wwm-ext-large')

[2021-02-07 16:58:12,351] [    INFO] - Already cached /home/aistudio/.paddlenlp/models/ernie-1.0/ernie_v1_chn_base.pdparams
[2021-02-07 16:58:14,238] [    INFO] - Found /home/aistudio/.paddlenlp/models/ernie-1.0/vocab.txt

github源码：https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/examples/text_classification/pretrained_models

PaddleNLP 更多项目

• 瞧瞧怎么使用PaddleNLP内置数据集-基于seq2vec的情感分析
• 如何通过预训练模型Fine-tune下游任务
• 使用BiGRU-CRF模型完成快递单信息抽取
• 使用预训练模型ERNIE优化快递单信息抽取
• 使用Seq2Seq模型完成自动对联
• 使用预训练模型ERNIE-GEN实现智能写诗
• 使用TCN网络完成新冠疫情病例数预测
• 使用预训练模型完成阅读理解

加入交流群，一起学习吧

现在就加入PaddleNLP的QQ技术交流群，一起交流NLP技术吧！