Why

[lminjy]

[XiaoyongNI]

Hi,

This is probably due to bad initialization of GRU. The easiest solution is to re-start the training several times until GRU is randomly initialized to a relatively good starting point. Of course, you could also try some more advanced initialization methods to stablize the starting phase of training.

Best regards, Xiaoyong

[lminjy]

Ok, after many tries, the training worked. I am new to neural networks and Kalman filtering, and I have some questions. Would you please explain.

1. Why cross-validate first？
2. Do you use the neural network to train first and get the x state directly?

3.How do I train with your particular network？I mean where do they run?

[XiaoyongNI]

1 To check whether the network initialize to a good point.

2 Yes.

3 In the main files.

[lminjy]

I checked all the calls and found no training for kg There are indeed steps to run here, but I didn't find out how they work

[lminjy]

I run main_linear and main_linear_estH, but do not see calls to these functions

[XiaoyongNI]

They are called inside the pipeline file. where you do forward step of the KNet model.

[lminjy]

oh, thank you very much