lstm_attention's Introduction

LSTM_Attention

X = Input Sequence of length n.
H = LSTM(X); Note that here the LSTM has return_sequences = True,
    so H is a sequence of vectors of length n.
s is the hidden state of the LSTM (h and c)

h is a weighted sum over H: 加权和
h = sigma(j = 0 to n-1)  alpha(j) * H(j)

weight alpha[i, j] for each hj is computed as follows:
H = [h1,h2,...,hn]
M = tanh(H)
alhpa = softmax(w.transpose * M)

h# = tanh(h)
y = softmax(W * h# + b)

J(theta) = negative_log_likelihood + regularity

attModel1

GitHub 项目

datalogue/keras-attention

Attention_Recurrent

GitHub 项目

ningshixian/LSTM_Attention

attModel2

GitHub 项目

Keras Attention Mechanism

Example: Attention block*

Attention defined per time series (each TS has its own attention)

attModel3

Github 项目

keras-language-modeling

https://github.com/roebius/deeplearning_keras2/blob/master/nbs2/attention_wrapper.py

attModel4

Github 项目

CDRextraction

hierarchical-attention-networks

Github：

synthesio/hierarchical-attention-networks

self-attention-networks

参考

lstm_attention's Issues

AttentionLSTM and AttentionLSTM_t

Thanks and I have following questions:

what is the difference of AttentionLSTM and AttentionLSTM_t in your code attention_lstm.py? (i.e. what do you mean by 'attention_vec'?)
Is the output at each timestep is the input to the next timestep in AttentionLSTM_t?

self_attention 第167行的问题，谢谢！

你好，我在运行你的这个程序时，第167行报错：TypeError: add_weight() got multiple values for argument 'name'，我用的是anaconda3+python3.6+keras2.3.1+tensorflow2.1.0，不知道你有什么建议吗？

process_data

Hello ningshixian
i fail to reproduce your fantastic work, it report as "NO module like process_data". could you give me some advice

AttentionDecoder 输入的time_step必须要和encoder的time_step一样？

您好，看了您的Git受益匪浅，有个小疑问：AttentionDecoder 输入的time_step必须要和encoder的time_step一样，那么AttentionDecoder 输出的time_step也就等于encoder的time_step，可是输入和输出的time_step数量一般情况下是不一样的啊？比如输入的是个问题，输出的是个答案，问题和答案的词数量一般都不一样。