PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

2024-01-16 16:16:58

PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课–基于seq2seq的对联生成

对联，是汉族传统文化之一，是写在纸、布上或刻在竹子、木头、柱子上的对偶语句。对联对仗工整，平仄协调，是一字一音的汉语独特的艺术形式，是中国传统文化瑰宝。

这里，我们将根据上联，自动写下联。这是一个典型的序列到序列(sequence2sequence, seq2seq）建模的场景，编码器-解码器（Encoder-Decoder）框架是解决seq2seq问题的经典方法，它能够将一个任意长度的源序列转换成另一个任意长度的目标序列：编码阶段将整个源序列编码成一个向量，解码阶段通过最大化预测序列概率，从中解码出整个目标序列。编码和解码的过程通常都使用RNN实现。

PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

图1：encoder-decoder示意图

这里的Encoder采用LSTM，Decoder采用带有attention机制的LSTM。

PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

图2：带有attention机制的encoder-decoder示意图

我们将以对联的上联作为Encoder的输出，下联作为Decoder的输入，训练模型。

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。

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

Looking in indexes: https://mirror.baidu.com/pypi/simple/
Requirement already up-to-date: paddlenlp in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (2.0.0rc1)
Requirement already satisfied, skipping upgrade: jieba in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp) (0.42.1)
Requirement already satisfied, skipping upgrade: h5py in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp) (2.9.0)
Requirement already satisfied, skipping upgrade: colorama in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp) (0.4.4)
Requirement already satisfied, skipping upgrade: visualdl in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp) (2.1.1)
Requirement already satisfied, skipping upgrade: colorlog in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp) (4.1.0)
Requirement already satisfied, skipping upgrade: seqeval in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from paddlenlp) (1.2.2)
Requirement already satisfied, skipping upgrade: six in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from h5py->paddlenlp) (1.15.0)
Requirement already satisfied, skipping upgrade: numpy>=1.7 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from h5py->paddlenlp) (1.16.4)
Requirement already satisfied, skipping upgrade: pre-commit in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (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) (0.7.1.1)
Requirement already satisfied, skipping upgrade: flake8>=3.7.9 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (3.8.2)
Requirement already satisfied, skipping upgrade: protobuf>=3.11.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (3.14.0)
Requirement already satisfied, skipping upgrade: Pillow>=7.0.0 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (7.1.2)
Requirement already satisfied, skipping upgrade: bce-python-sdk in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (0.8.53)
Requirement already satisfied, skipping upgrade: requests in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (2.22.0)
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) (1.0.0)
Requirement already satisfied, skipping upgrade: flask>=1.1.1 in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from visualdl->paddlenlp) (1.1.1)
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) (0.22.1)
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) (1.4.10)
Requirement already satisfied, skipping upgrade: pyyaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp) (5.1.2)
Requirement already satisfied, skipping upgrade: aspy.yaml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp) (1.3.0)
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) (1.3.4)
Requirement already satisfied, skipping upgrade: toml in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp) (0.10.0)
Requirement already satisfied, skipping upgrade: importlib-metadata; python_version < "3.8" in /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages (from pre-commit->visualdl->paddlenlp) (0.23)
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) (16.7.9)
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) (2.0.1)
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) (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) (2.6.0)
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) (2.2.0)
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) (3.9.9)
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) (0.18.0)
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) (2.8)
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) (3.0.4)
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) (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) (2019.9.11)
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) (2.10.1)
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) (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) (2.8.0)
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) (7.0)
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) (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) (1.1.0)
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) (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) (1.3.0)
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"->pre-commit->visualdl->paddlenlp) (0.6.0)
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) (1.1.1)
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"->pre-commit->visualdl->paddlenlp) (7.2.0)

import paddlenlp
paddlenlp.__version__

'2.0.0rc1'

import io
import os

from functools import partial

import numpy as np

import paddle
import paddle.nn as nn
import paddle.nn.functional as F
from paddlenlp.data import Vocab, Pad
from paddlenlp.metrics import Perplexity
from paddlenlp.datasets import CoupletDataset

数据部分

数据集介绍

采用开源的对联数据集couplet-clean-dataset，该数据集过滤了
couplet-dataset中的低俗、敏感内容。

这个数据集包含70w多条训练样本，1000条验证样本和1000条测试样本。

下面列出一些训练集中对联样例：

上联：晚风摇树树还挺 下联：晨露润花花更红

上联：愿景天成无墨迹 下联：万方乐奏有于阗

上联：丹枫江冷人初去 下联：绿柳堤新燕复来

上联：闲来野钓人稀处 下联：兴起高歌酒醉中

加载数据集

paddlenlp.datasets中内置了多个常见数据集，包括这里的对联数据集CoupletDataset。


paddlenlp.datasets均继承paddle.io.Dataset，支持paddle.io.Dataset的所有功能：

  • 通过len()函数返回数据集长度，即样本数量。
  • 下标索引：通过下标索引[n]获取第n条样本。
  • 遍历数据集，获取所有样本。

此外，paddlenlp.datasets，还支持如下操作：

  • 调用get_datasets()函数，传入list或者string，获取相对应的train_dataset、development_dataset、test_dataset等。其中train为训练集，用于模型训练； development为开发集，也称验证集validation_dataset，用于模型参数调优；test为测试集，用于评估算法的性能，但不会根据测试集上的表现再去调整模型或参数。
  • 调用apply()函数，对数据集进行指定操作。

这里的CoupletDataset数据集继承TranslationDataset，继承自paddlenlp.datasets，除以上通用用法外，还有一些个性设计：

  • 在CoupletDataset class中，还定义了transform函数，用于在每个句子的前后加上起始符<s>和结束符</s>，并将原始数据映射成id序列。
PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

图3：token-to-id示意图 
train_ds, dev_ds, test_ds = CoupletDataset.get_datasets(['train', 'dev', 'test'])

100%|██████████| 21421/21421 [00:00<00:00, 23224.53it/s]

来看看数据集有多大，长什么样：

print (len(train_ds), len(test_ds), len(dev_ds))
for i in range(5):
    print (train_ds[i])

print ('\n')
for i in range(5):
    print (test_ds[i])

702594 999 1000
([1, 447, 3, 509, 153, 153, 279, 1517, 2], [1, 816, 294, 378, 9, 9, 142, 32, 2])
([1, 594, 185, 10, 71, 18, 158, 912, 2], [1, 14, 105, 107, 835, 20, 268, 3855, 2])
([1, 335, 830, 68, 425, 4, 482, 246, 2], [1, 94, 51, 1115, 23, 141, 761, 17, 2])
([1, 126, 17, 217, 802, 4, 1103, 118, 2], [1, 125, 205, 47, 55, 57, 78, 15, 2])
([1, 1203, 228, 390, 10, 1921, 827, 474, 2], [1, 1699, 89, 426, 317, 314, 43, 374, 2])


([1, 6, 201, 350, 54, 1156, 2], [1, 64, 522, 305, 543, 102, 2])
([1, 168, 1402, 61, 270, 11, 195, 253, 2], [1, 435, 782, 1046, 36, 188, 1016, 56, 2])
([1, 744, 185, 744, 6, 18, 452, 16, 1410, 2], [1, 286, 102, 286, 74, 20, 669, 280, 261, 2])
([1, 2577, 496, 1133, 60, 107, 2], [1, 1533, 318, 625, 1401, 172, 2])
([1, 163, 261, 6, 64, 116, 350, 253, 2], [1, 96, 579, 13, 463, 16, 774, 586, 2])

vocab, _ = CoupletDataset.get_vocab()
trg_idx2word = vocab.idx_to_token
vocab_size = len(vocab)

pad_id = vocab[CoupletDataset.EOS_TOKEN]
bos_id = vocab[CoupletDataset.BOS_TOKEN]
eos_id = vocab[CoupletDataset.EOS_TOKEN]
print (pad_id, bos_id, eos_id)

2 1 2

构造dataloder

使用paddle.io.DataLoader来创建训练和预测时所需要的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。支持单进程或多进程加载数据，快！


接收如下重要参数：

  • batch_sampler：批采样器实例，用于在paddle.io.DataLoader 中迭代式获取mini-batch的样本下标数组，数组长度与 batch_size 一致。
  • 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。

PaddleNLP提供了许多NLP任务中，用于数据处理、组batch数据的相关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
paddlenlp.data.Pad 将长度不同的多个句子padding到统一长度，取N个输入数据中的最大长度
paddlenlp.data.Tuple 将多个batchify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

def create_data_loader(dataset):
    data_loader = paddle.io.DataLoader(
        dataset,
        batch_sampler=None,
        batch_size = batch_size,
        collate_fn=partial(prepare_input, pad_id=pad_id))
    return data_loader

def prepare_input(insts, pad_id):
    src, src_length = Pad(pad_val=pad_id, ret_length=True)([inst[0] for inst in insts])
    tgt, tgt_length = Pad(pad_val=pad_id, ret_length=True)([inst[1] for inst in insts])
    tgt_mask = (tgt[:, :-1] != pad_id).astype(paddle.get_default_dtype())
    return src, src_length, tgt[:, :-1], tgt[:, 1:, np.newaxis], tgt_mask

use_gpu = False
device = paddle.set_device("gpu" if use_gpu else "cpu")

batch_size = 128
num_layers = 2
dropout = 0.2
hidden_size =256
max_grad_norm = 5.0
learning_rate = 0.001
max_epoch = 20
model_path = './couplet_models'
log_freq = 200

# Define dataloader
train_loader = create_data_loader(train_ds)
test_loader = create_data_loader(test_ds)

print(len(train_ds), len(train_loader), batch_size)
# 702594 5490 128  共5490个batch

for i in train_loader:
    print (len(i))
    for ind, each in enumerate(i):
        print (ind, each.shape)
    break

702594 5490 128
5
0 [128, 18]
1 [128]
2 [128, 17]
3 [128, 17, 1]
4 [128, 17]

模型部分

下图是带有Attention的Seq2Seq模型结构。下面我们分别定义网络的每个部分，最后构建Seq2Seq主网络。

PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

图5：带有attention机制的encoder-decoder原理示意图

定义Encoder

Encoder部分非常简单，可以直接利用PaddlePaddle2.0提供的RNN系列API的nn.LSTM。

  1. nn.Embedding：该接口用于构建 Embedding 的一个可调用对象，根据输入的size (vocab_size, embedding_dim)自动构造一个二维embedding矩阵，用于table-lookup。查表过程如下：
PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

图5：token-to-id & 查表获取向量示意图
  1. nn.LSTM：提供序列，得到encoder_output和encoder_state。
    参数：
  • input_size (int) 输入的大小。
  • hidden_size (int) - 隐藏状态大小。
  • num_layers (int，可选) - 网络层数。默认为1。
  • direction (str，可选) - 网络迭代方向，可设置为forward或bidirect（或bidirectional）。默认为forward。
  • time_major (bool，可选) - 指定input的第一个维度是否是time steps。默认为False。
  • dropout (float，可选) - dropout概率，指的是出第一层外每层输入时的dropout概率。默认为0。

https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/nn/layer/rnn/LSTM_cn.html

输出:

outputs (Tensor) - 输出，由前向和后向cell的输出拼接得到。如果time_major为True，则Tensor的形状为[time_steps,batch_size,num_directions * hidden_size]，如果time_major为False，则Tensor的形状为[batch_size,time_steps,num_directions * hidden_size]，当direction设置为bidirectional时，num_directions等于2，否则等于1。

final_states (tuple) - 最终状态,一个包含h和c的元组。形状为[num_lauers * num_directions, batch_size, hidden_size],当direction设置为bidirectional时，num_directions等于2，否则等于1。

class Seq2SeqEncoder(nn.Layer):
    def __init__(self, vocab_size, embed_dim, hidden_size, num_layers):
        super(Seq2SeqEncoder, self).__init__()
        self.embedder = nn.Embedding(vocab_size, embed_dim)
        self.lstm = nn.LSTM(
            input_size=embed_dim,
            hidden_size=hidden_size,
            num_layers=num_layers,
            dropout=0.2 if num_layers > 1 else 0.)

    def forward(self, sequence, sequence_length):
        inputs = self.embedder(sequence)
        encoder_output, encoder_state = self.lstm(
            inputs, sequence_length=sequence_length)
        
        # encoder_output [128, 18, 256]  [batch_size,time_steps,hidden_size]
        # encoder_state (tuple) - 最终状态,一个包含h和c的元组。 [2, 128, 256] [2, 128, 256] [num_lauers * num_directions, batch_size, hidden_size]
        return encoder_output, encoder_state

定义Decoder

定义AttentionLayer

  1. nn.Linear线性变换层传入2个参数
  • in_features (int) – 线性变换层输入单元的数目。
  • out_features (int) – 线性变换层输出单元的数目。

PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成

  1. paddle.matmul用于计算两个Tensor的乘积，遵循完整的广播规则，关于广播规则，请参考广播 (broadcasting) 。 并且其行为与 numpy.matmul 一致。
  • x (Tensor) : 输入变量，类型为 Tensor，数据类型为float32， float64。
  • y (Tensor) : 输入变量，类型为 Tensor，数据类型为float32， float64。
  • transpose_x (bool，可选) : 相乘前是否转置 x，默认值为False。
  • transpose_y (bool，可选) : 相乘前是否转置 y，默认值为False。

  1. paddle.unsqueeze用于向输入Tensor的Shape中一个或多个位置（axis）插入尺寸为1的维度

  2. paddle.add逐元素相加算子，输入 x 与输入 y 逐元素相加，并将各个位置的输出元素保存到返回结果中。

输入 x 与输入 y 必须可以广播为相同形状。

class AttentionLayer(nn.Layer):
    def __init__(self, hidden_size):
        super(AttentionLayer, self).__init__()
        self.input_proj = nn.Linear(hidden_size, hidden_size)
        self.output_proj = nn.Linear(hidden_size + hidden_size, hidden_size)

    def forward(self, hidden, encoder_output, encoder_padding_mask):
        encoder_output = self.input_proj(encoder_output)
        attn_scores = paddle.matmul(
            paddle.unsqueeze(hidden, [1]), encoder_output, transpose_y=True)
        # print('attention score', attn_scores.shape) #[128, 1, 18]

        if encoder_padding_mask is not None:
            attn_scores = paddle.add(attn_scores, encoder_padding_mask)

        attn_scores = F.softmax(attn_scores)
        attn_out = paddle.squeeze(
            paddle.matmul(attn_scores, encoder_output), [1])
        # print('1 attn_out', attn_out.shape) #[128, 256]

        attn_out = paddle.concat([attn_out, hidden], 1)
        # print('2 attn_out', attn_out.shape) #[128, 512]

        attn_out = self.output_proj(attn_out)
        # print('3 attn_out', attn_out.shape) #[128, 256]
        return attn_out

定义Seq2SeqDecoderCell

由于Decoder部分是带有attention的LSTM，我们不能复用nn.LSTM，所以需要定义Seq2SeqDecoderCell

  1. nn.LayerList 用于保存子层列表，它包含的子层将被正确地注册和添加。列表中的子层可以像常规python列表一样被索引。这里添加了num_layers=2层lstm。
class Seq2SeqDecoderCell(nn.RNNCellBase):
    def __init__(self, num_layers, input_size, hidden_size):
        super(Seq2SeqDecoderCell, self).__init__()
        self.dropout = nn.Dropout(0.2)
        self.lstm_cells = nn.LayerList([
            nn.LSTMCell(
                input_size=input_size + hidden_size if i == 0 else hidden_size,
                hidden_size=hidden_size) for i in range(num_layers)
        ])

        self.attention_layer = AttentionLayer(hidden_size)
    
    def forward(self,
                step_input,
                states,
                encoder_output,
                encoder_padding_mask=None):
        lstm_states, input_feed = states
        new_lstm_states = []
        step_input = paddle.concat([step_input, input_feed], 1)
        for i, lstm_cell in enumerate(self.lstm_cells):
            out, new_lstm_state = lstm_cell(step_input, lstm_states[i])
            step_input = self.dropout(out)
            new_lstm_states.append(new_lstm_state)
        out = self.attention_layer(step_input, encoder_output,
                                   encoder_padding_mask)
        return out, [new_lstm_states, out]

定义Seq2SeqDecoder

有了Seq2SeqDecoderCell，就可以构建Seq2SeqDecoder了


  1. paddle.nn.RNN 该OP是循环神经网络（RNN）的封装，将输入的Cell封装为一个循环神经网络。它能够重复执行 cell.forward() 直到遍历完input中的所有Tensor。
  • cell (RNNCellBase) - RNNCellBase类的一个实例。
class Seq2SeqDecoder(nn.Layer):
    def __init__(self, vocab_size, embed_dim, hidden_size, num_layers):
        super(Seq2SeqDecoder, self).__init__()
        self.embedder = nn.Embedding(vocab_size, embed_dim)
        self.lstm_attention = nn.RNN(
            Seq2SeqDecoderCell(num_layers, embed_dim, hidden_size))
        self.output_layer = nn.Linear(hidden_size, vocab_size)

    def forward(self, trg, decoder_initial_states, encoder_output,
                encoder_padding_mask):
        inputs = self.embedder(trg)

        decoder_output, _ = self.lstm_attention(
            inputs,
            initial_states=decoder_initial_states,
            encoder_output=encoder_output,
            encoder_padding_mask=encoder_padding_mask)
        predict = self.output_layer(decoder_output)

        return predict

构建主网络Seq2SeqAttnModel

Encoder和Decoder定义好之后，网络就可以构建起来了

class Seq2SeqAttnModel(nn.Layer):
    def __init__(self, vocab_size, embed_dim, hidden_size, num_layers,
                 eos_id=1):
        super(Seq2SeqAttnModel, self).__init__()
        self.hidden_size = hidden_size
        self.eos_id = eos_id
        self.num_layers = num_layers
        self.INF = 1e9
        self.encoder = Seq2SeqEncoder(vocab_size, embed_dim, hidden_size,
                                      num_layers)
        self.decoder = Seq2SeqDecoder(vocab_size, embed_dim, hidden_size,
                                      num_layers)

    def forward(self, src, src_length, trg):
        # encoder_output 各时刻的输出h
        # encoder_final_state 最后时刻的输出h，和记忆信号c
        encoder_output, encoder_final_state = self.encoder(src, src_length)
        print('encoder_output shape', encoder_output.shape)  #  [128, 18, 256]  [batch_size,time_steps,hidden_size]
        print('encoder_final_states shape', encoder_final_state[0].shape, encoder_final_state[1].shape) #[2, 128, 256] [2, 128, 256] [num_lauers * num_directions, batch_size, hidden_size]

        # Transfer shape of encoder_final_states to [num_layers, 2, batch_size, hidden_size]？？？
        encoder_final_states = [
            (encoder_final_state[0][i], encoder_final_state[1][i])
            for i in range(self.num_layers)
        ]
        print('encoder_final_states shape', encoder_final_states[0][0].shape, encoder_final_states[0][1].shape) #[128, 256] [128, 256]


        # Construct decoder initial states: use input_feed and the shape is
        # [[h,c] * num_layers, input_feed], consistent with Seq2SeqDecoderCell.states
        decoder_initial_states = [
            encoder_final_states,
            self.decoder.lstm_attention.cell.get_initial_states(
                batch_ref=encoder_output, shape=[self.hidden_size])
        ]

        # Build attention mask to avoid paying attention on padddings
        src_mask = (src != self.eos_id).astype(paddle.get_default_dtype())
        print ('src_mask shape', src_mask.shape)  #[128, 18]
        print(src_mask[0, :])

        encoder_padding_mask = (src_mask - 1.0) * self.INF
        print ('encoder_padding_mask', encoder_padding_mask.shape)  #[128, 18]
        print(encoder_padding_mask[0, :])

        encoder_padding_mask = paddle.unsqueeze(encoder_padding_mask, [1])
        print('encoder_padding_mask', encoder_padding_mask.shape)  #[128, 1, 18]

        predict = self.decoder(trg, decoder_initial_states, encoder_output,
                               encoder_padding_mask)
        print('predict', predict.shape)   #[128, 17, 7931]

        return predict

定义损失函数

这里使用的是交叉熵损失函数，我们需要将padding位置的loss置为0，因此需要在损失函数中引入trg_mask参数，由于PaddlePaddle框架提供的paddle.nn.CrossEntropyLoss不能接受trg_mask参数，因此在这里需要重新定义：

class CrossEntropyCriterion(nn.Layer):
    def __init__(self):
        super(CrossEntropyCriterion, self).__init__()

    def forward(self, predict, label, trg_mask):
        cost = F.softmax_with_cross_entropy(
            logits=predict, label=label, soft_label=False)
        cost = paddle.squeeze(cost, axis=[2])
        masked_cost = cost * trg_mask
        batch_mean_cost = paddle.mean(masked_cost, axis=[0])
        seq_cost = paddle.sum(batch_mean_cost)

        return seq_cost

执行过程

训练过程

使用高层API执行训练，需要调用prepare和fit函数。

在prepare函数中，配置优化器、损失函数，以及评价指标。其中评价指标使用的是PaddleNLP提供的困惑度计算API paddlenlp.metrics.Perplexity。

如果你安装了VisualDL，可以在fit中添加一个callbacks参数使用VisualDL观测你的训练过程，如下：

model.fit(train_data=train_loader,
            epochs=max_epoch,
            eval_freq=1,
            save_freq=1,
            save_dir=model_path,
            log_freq=log_freq,
            callbacks=[paddle.callbacks.VisualDL('./log')])

在这里，由于对联生成任务没有明确的评价指标，因此，可以在保存的多个模型中，通过人工评判生成结果选择最好的模型。

本项目中，为了便于演示，已经将训练好的模型参数载入模型，并省略了训练过程。读者自己实验的时候，可以尝试自行修改超参数，调用下面被注释掉的fit函数，重新进行训练。

如果读者想要在更短的时间内得到效果不错的模型，可以使用预训练模型技术，例如《预训练模型ERNIE-GEN自动写诗》项目为大家展示了如何利用预训练的生成模型进行训练。

model = paddle.Model(
    Seq2SeqAttnModel(vocab_size, hidden_size, hidden_size,
                        num_layers, pad_id))

optimizer = paddle.optimizer.Adam(
    learning_rate=learning_rate, parameters=model.parameters())
ppl_metric = Perplexity()
model.prepare(optimizer, CrossEntropyCriterion(), ppl_metric)

model.fit(train_data=train_loader,
            epochs=max_epoch,
            eval_freq=1,
            save_freq=1,
            save_dir=model_path,
            log_freq=log_freq)

predict [128, 17, 7931]

模型预测

定义预测网络Seq2SeqAttnInferModel

预测网络继承上面的主网络Seq2SeqAttnModel，定义子类Seq2SeqAttnInferModel

class Seq2SeqAttnInferModel(Seq2SeqAttnModel):
    def __init__(self,
                 vocab_size,
                 embed_dim,
                 hidden_size,
                 num_layers,
                 bos_id=0,
                 eos_id=1,
                 beam_size=4,
                 max_out_len=256):
        self.bos_id = bos_id
        self.beam_size = beam_size
        self.max_out_len = max_out_len
        self.num_layers = num_layers
        super(Seq2SeqAttnInferModel, self).__init__(
            vocab_size, embed_dim, hidden_size, num_layers, eos_id)

        # Dynamic decoder for inference
        self.beam_search_decoder = nn.BeamSearchDecoder(
            self.decoder.lstm_attention.cell,
            start_token=bos_id,
            end_token=eos_id,
            beam_size=beam_size,
            embedding_fn=self.decoder.embedder,
            output_fn=self.decoder.output_layer)

    def forward(self, src, src_length):
        encoder_output, encoder_final_state = self.encoder(src, src_length)

        encoder_final_state = [
            (encoder_final_state[0][i], encoder_final_state[1][i])
            for i in range(self.num_layers)
        ]

        # Initial decoder initial states
        decoder_initial_states = [
            encoder_final_state,
            self.decoder.lstm_attention.cell.get_initial_states(
                batch_ref=encoder_output, shape=[self.hidden_size])
        ]
        # Build attention mask to avoid paying attention on paddings
        src_mask = (src != self.eos_id).astype(paddle.get_default_dtype())

        encoder_padding_mask = (src_mask - 1.0) * self.INF
        encoder_padding_mask = paddle.unsqueeze(encoder_padding_mask, [1])

        # Tile the batch dimension with beam_size
        encoder_output = nn.BeamSearchDecoder.tile_beam_merge_with_batch(
            encoder_output, self.beam_size)
        encoder_padding_mask = nn.BeamSearchDecoder.tile_beam_merge_with_batch(
            encoder_padding_mask, self.beam_size)

        # Dynamic decoding with beam search
        seq_output, _ = nn.dynamic_decode(
            decoder=self.beam_search_decoder,
            inits=decoder_initial_states,
            max_step_num=self.max_out_len,
            encoder_output=encoder_output,
            encoder_padding_mask=encoder_padding_mask)
        return seq_output

解码部分

接下来对我们的任务选择beam search解码方式，可以指定beam_size为10。

def post_process_seq(seq, bos_idx, eos_idx, output_bos=False, output_eos=False):
    """
    Post-process the decoded sequence.
    """
    eos_pos = len(seq) - 1
    for i, idx in enumerate(seq):
        if idx == eos_idx:
            eos_pos = i
            break
    seq = [
        idx for idx in seq[:eos_pos + 1]
        if (output_bos or idx != bos_idx) and (output_eos or idx != eos_idx)
    ]
    return seq

beam_size = 10
# init_from_ckpt = './couplet_models/0' # for test
# infer_output_file = './infer_output.txt'

# test_loader, vocab_size, pad_id, bos_id, eos_id = create_data_loader(test_ds, batch_size)
# vocab, _ = CoupletDataset.get_vocab()
# trg_idx2word = vocab.idx_to_token

model = paddle.Model(
    Seq2SeqAttnInferModel(
        vocab_size,
        hidden_size,
        hidden_size,
        num_layers,
        bos_id=bos_id,
        eos_id=eos_id,
        beam_size=beam_size,
        max_out_len=256))

model.prepare()

在预测之前，我们需要将训练好的模型参数load进预测网络，之后我们就可以根据对联的上联，生成对联的下联啦！

model.load('couplet_models/model_18')

test_ds = CoupletDataset.get_datasets(['test'])
idx = 0
for data in test_loader():
    inputs = data[:2]
    finished_seq = model.predict_batch(inputs=list(inputs))[0]
    finished_seq = finished_seq[:, :, np.newaxis] if len(
        finished_seq.shape) == 2 else finished_seq
    finished_seq = np.transpose(finished_seq, [0, 2, 1])
    for ins in finished_seq:
        for beam in ins:
            id_list = post_process_seq(beam, bos_id, eos_id)
            word_list_l = [trg_idx2word[id] for id in test_ds[idx][0]][1:-1]
            word_list_r = [trg_idx2word[id] for id in id_list]
            sequence = "上联: "+" ".join(word_list_l)+"\t下联: "+" ".join(word_list_r) + "\n"
            print(sequence)
            idx += 1
            break

/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

PaddleNLP更多教程

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！

PaddlePaddle飞桨《高层API助你快速上手深度学习》『深度学习7日打卡营』第五课--基于seq2seq的对联生成
上一篇：WSI 病理图像预处理（python）安装spams (失败)


下一篇：离线状态迁移Anaconda虚拟环境