shawnyxiao / textclassification-keras Goto Github PK

Attention类实现感觉有问题,本人使用TextBiRNN模型来跑,出现以下错误:

Traceback (most recent call last):
  File "train.py", line 86, in <module>
    last_activation='softmax').get_model()
  File "/media/gaoya/disk/Applications/keras/短文本分类/model.py", line 282, in get_model
    x_word = Attention(self.maxlen_word)(x_word)
  File "/media/gaoya/disk/Applications/anaconda3/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py", line 75, in symbolic_fn_wrapper
    return func(*args, **kwargs)
  File "/media/gaoya/disk/Applications/anaconda3/lib/python3.7/site-packages/keras/engine/base_layer.py", line 463, in __call__
    self.build(unpack_singleton(input_shapes))
  File "/media/gaoya/disk/Applications/keras/短文本分类/model.py", line 231, in build
    constraint=self.b_constraint)
TypeError: add_weight() got multiple values for argument 'name'

请问怎么解决呢

不知道有没有人调通HAN，我在最后fit的时候报了一个维度不不匹配的错误，错误显示：Input输入时需要三维，但是实际获取的是二维（句子数目，句长）

请问大神：在TextAttBiRNN中怎么输出Encoder中某个词对最后预测的权重值啊？

Hi Shawny,

I run your code RCNN, and I found the accuracy was 0.5. Do you know why this happened? Looking forward to your reply. Thanks!

IMDB dataset is tokenized by words, so after reshape in https://github.com/ShawnyXiao/TextClassification-Keras/blob/master/model/HAN/main.py#L20-L23 word-level dimension contain the whole words tokens, not chars/wordparts. Does it make sense at all? If it is used just for illustration, then maybe it worth to add a comment near it.

如上line22-25这4行代码，所示编码过程好像如下：
Step1: 强行在document(所有句子)后面padding一次，而不是在每个句子后面都padding一次，形如：（---表示句子）
-----------,------,--- ------------,-------- --,000000000000000000 00000000000000000000

Step2: 强行把document按maxlen_sentence(假设为20)划分看，而非原本句子的自然划分，形如：（|表示向量划分）
-----------,------,---|------------,--------|--,000000000000000000|00000000000000000000

我认为，应该是每个句子内先进行Word Level的编码，然后再进行句子间的Sentence Level编码？形如：
----------- 000000 000|------000000 00000000|-- -------------00000|----------0000000000

大家如何看待？

您好，在基于您的代码进行学习的过程中，想查看一下注意力机制为我的文本数据分配的权重，即TextAttBiRNN中attention计算中的a。
但是我在成功输出a后发现，我每句话的每个词都被分配了同样的权重，例如一句话23个单词，每个单词的权重都是1/23。
百思不得其解，在这里想向您请教一下有可能是什么原因造成的呢？数据集按理说不应该有问题的，直接基于RNN做效果也不错。非常抱歉问题这么基础，希望您能解答，感谢！！

Hi,
Thank you for these codes.
I wonder if it is possible to get attention weights vector on word-level and also sentence-level attention in Han model?
I want to plot these weights vector after predicting its category.

Hello!

Could you tell me, please, how can I use your code on my own data? For example I have a vectors of words organised by tf-idf scheme. What should I do if I wanna test it instead of imdb data?