Confused about The seq_length and window_length ，can I just use window_length to segment ?I try to use Optimized network structure to train data,it seems hard to train ,could you please provide me a single py about mulit-head cnn-lstm model? about enas-pdm HOT 3 CLOSED

HI, 943fansi. In this work, we apply first level time window approach to have 'sequence' and then the 2nd level time window is considered to extract the segments from the sequence. I know this approach is somewhat confusing but this is necessary for the multi-head CNN-LSTM. (for further details, please check our paper, https://www.sciencedirect.com/science/article/pii/S1568494621003975)
In the 'network_training.py' file, you can check how i implement the network in TensorFlow-Kreas.
If you feel this is too complicated and want to avoid, then you can use either 1D-CNN (multi-channel) or LSTM, instead of the CNN-LSTM. In that case, you only need to consider a time window.

from enas-pdm.

Hi, mohyunho.Thanks for your reply. This reply does not contain any questions.
I debug the code ,the original input is 4d tensor ,(batch_szie, n_window, window_lenth, sensor_dim), For each head, the input shape is (batch_szie, n_window, window_lenth, 1),output shape is (batch_szie, n_window, 28), then merge as (batch_szie, n_window, 28*14) to be the input of lstm.
I am a novice in deep learning. I am not familiar with tensorflow and have encountered some difficulties.If it brings you some trouble, say sorry here.
Wish you a happy life.

from enas-pdm.

No problem at all. When you have any questions, please let me know.

from enas-pdm.