shen1994 / chatrobot

keras+python3下的seq2seq+attention中文对话系统

ChatRobot

0. 特别提醒  

• 关于keras环境下seq2seq错误修改
  ('_OptionalInputPlaceHolder' object has no attribute 'inbound_nodes')
• 0.0 使用keras2.1.0版本的第三方库(不推荐)
• 0.1 将recurrentshop\engine.py文件中837和842行中inbound_nodes更改为_inbound_nodes

1. 效果展示

1.0 python train.py执行效果图

1.1 python test.py执行效果图

1.2 python chat_robot.py执行效果图

2. 软件安装

• 模型搭建第三方库Keras-2.1.6.tar.gz
  私人地址: 链接: https://pan.baidu.com/s/1ypoEgf6ITjcNalzTRtnvmw 密码: uot8
• recurrentshop下载地址: https://github.com/farizrahman4u/recurrentshop
• seq2seq 下载地址: https://github.com/farizrahman4u/seq2seq
• 微博数据(关于餐饮业,数据未清洗)下载地址
  私人地址: 链接: https://pan.baidu.com/s/1g6l4_IDkLdLAjvrWf5sheQ 密码: fxy3

3. 参考链接

• seq2seq讲解: http://jacoxu.com/encoder_decoder
• seq2seq数据读取: http://suriyadeepan.github.io/2016-06-28-easy-seq2seq
• seq2seq论文地址: https://arxiv.org/abs/1409.3215  
• seq2seq+attention论文地址: https://arxiv.org/abs/1409.0473  
• ChatRobot启发论文: https://arxiv.org/abs/1503.02364
• seq2seq源码: https://github.com/farizrahman4u/seq2seq  
• seq2seq源码需求: https://github.com/farizrahman4u/recurrentshop  
• beamsearch源码参考: https://github.com/yanwii/seq2seq
• bucket源码参考: https://github.com/1228337123/tensorflow-seq2seq-chatbot-zh

4. 执行命令

• 生成序列文件,将文字编码为数字,不足补零
  python data_process.py
• 生成word2vec向量,包括编码向量和解码向量
  python word2vec.py
• 训练网络
  python train.py
• 测试
  python test.py
• 模型评分
  python score.py
• 智能问答
  python chat_robot.py
• 绘制word2vec向量分布图
  python word2vec_plot.py

5. 更新

• Word2cut模型对陌生词汇的分词未解决,有时间搞定一下

chatrobot's Issues

_OptionalInputPlaceHolder' object has no attribute '_inbound_nodes'

There is something looks strange... 0.0
In what i download, engine.py 's 837 and 842 is

def _get_optional_input_placeholder(self, name=None, num=1):
        if name:
            if name not in self._optional_input_placeholders:
                if num > 1:
                    self._optional_input_placeholders[name] = [self._get_optional_input_placeholder() for _ in range(num)]
                else:
                    self._optional_input_placeholders[name] = self._get_optional_input_placeholder()
            return self._optional_input_placeholders[name]
        if num == 1:
            **optional_input_placeholder = _to_list(_OptionalInputPlaceHolder()._inbound_nodes[0].output_tensors)[0]**
            assert self._is_optional_input_placeholder(optional_input_placeholder)
            return optional_input_placeholder
        else:
            y = []
            for _ in range(num):
                **optional_input_placeholder = _to_list(_OptionalInputPlaceHolder()._inbound_nodes[0].output_tensors)[0]**
                assert self._is_optional_input_placeholder(optional_input_placeholder)
                y.append(optional_input_placeholder)
            return y  

thanks !!!

您好，我想问下，假如我语料库改变了，怎么做新的模型啊

我运行命令：
python data_process.py
得到错误：

question about the initial_state of decoder

Hi, I am wonder that how to initialize the state of the encoder and decoder. I tried both encoder_states and [a1, b1] to initialize the state of the decoder_lstm2 but got bad result. I can get good results when I use only decoder_lstm1 with one or several encoders which without initialize the state. Could you please help me? Here is my code.

encoder_inputs = Input(shape=(max_video_length, 4096), dtype='float32')
encoder = LSTM(latent_dim, return_state=True)
encoder_outputs, state_h, state_c = encoder(encoder_outputs2)
encoder_states = [state_h, state_c]

decoder_inputs = Input(shape=(None, len(char_list)), name="decoder_inputs")
decoder_lstm1 = LSTM(latent_dim, return_sequences=True, return_state=True, name="decoder_lstm1")
decoder_lstm2 = LSTM(latent_dim, return_sequences=True, return_state=True, name="decoder_lstm2")
decoder_dense = Dense(len(char_list), activation='softmax')

decoder_outputs, a1, b1 = decoder_lstm1(decoder_inputs, initial_state=encoder_states)
decoder_outputs, a2, b2 = decoder_lstm2(decoder_outputs, initial_state=encoder_states)
decoder_outputs = decoder_dense(decoder_outputs)

model = Model([encoder_inputs, decoder_inputs], decoder_outputs)