How to deal with Time Sequence Inputs for 1D Convolutional-LSTM networks.

I am trying to combine two approach for Time Sequence Classification using deep learning.

The first one implement LSTM networks and it is described here:

1) https://it.mathworks.com/help/deeplearning/ug/sequence-to-sequence-classification-using-deep-learning.html

The seccond apply convolutional networks and it is described here:

2) https://it.mathworks.com/help/deeplearning/ug/sequence-to-sequence-classification-using-1-d-convolutions.html

Following previous advices on ANSWERS I used the Deep Network builder object to recreate the main convolutional block of 2) as

Now my doubt is how should i format the accelerometry data for the input of this network?

My data are 42 features signals from accelerometry represented as 42xN°observations.

I tried to format the data as a sequence of images as 1x1x42xN°observations, and it seemed work but still my doubt remains.

Is this data format correct ? and if so:

It is correct to define 1x3 as dimension of the filter?

Thank in advance,

[Mirko Job]

UPDATE: So as far as i understand, if i give in input a cell array of sequence of images NxNx3xNobs the fold layer transform the cell array in a 4D-double NxNxNobs1+Nobs2+... Now the question is what happened when I apply the filter along the temporal dimension ? In particular at the end of a sequence and the beginning of another the convolution over 3 samples will aggregate the contributes of multiple sequences? And since the temporal component is the fourth dimension, should my filter be defined as 1x1x1x3?

[krishna Chauhan]

@Mirko Job

did you find your answers sir?

I am dealing with sequence classification using TCN.

[Mirko Job]

@krishna Chauhan Yes, the problem is that you have to write your own custom training loop routine using dlarrays. So first you train TCN using the approach described in the link. Then you extract the Features dlarrays from the net manually before the Fully Connected part. Then you train a lstm net (see lstm function) from the elaborated temporale features obtained through TCN.