I wonder if the dynonet can be modified to a time varying transfer function coefficient estimator?

[lk1983823]

Thanks for your wonderful work on estimating the static coefficient of a transfer function using an end-to-end deep learning method.

Based on your previous work， I just recently modified the source code to let the b_coeff and a_coeff be a time varying result. That is, I assume there is an exogenous variable, u_ex which could have any feature and affect the coefficients. In addition, the u_ex have the same sequence length as u_in. The shape of u_ex could be (batch_size, seq_len, input_dim). So, the math formula could be "b_coeff, a_coeff = NN(u_ex) ". Finally, I choose GRU cell to do this. However, it comes into a problem:

I find the time-varying coeff from my GRU network has the batch_size dimension, that is the b_coeff's shape changes into (batch_size, out_channels, in_channels, n_b). This can't be accepted by your MimoLinearDynamicalOperatorFun class.

So, I would like to ask you for some suggestions, as:

1. Is it proper to do this from the viewpoint of system identification theory ?
2. if the answer is yes for question1, then how should I modify the forward and backward function in MimoLinearDynamicalOperatorFun class?
   Thanks for your reply.