If you can, how can you get a SeriesNetwork object from a Sequence-to-Sequence Classification using 1-D Convolutions?

I am developing a mobile application that uses neural networks to detect a certain pattern in a sequence.

In order to train the network, the example of "Sequence-to-Sequence Classification Using 1-D Convolutions" was largely followed.

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

My goal is to convert the SeriesNetwork object of the trained network to .onnx format, using exportONNXNetwork, but in this example this SeriesNetwork object is never created. Instead, for the training and inference phase we rely on parameters and hyperparameters identified during the training phase.

Is there a known strategy to perform the same sequence of operations (Training and Inference), but having as a result of training a Series Network object representing the neural network?

Answer 1

[Anshika Chaurasia]

Hi,

The example you mentioned had created and trained a deep learning network by using functions rather than a layer graph or a dlnetwork.

As per my knowledge, there is no way to convert model functions into layer graph or SeriesNetwork object. The only way is to implement the network using deep learning layers as shown in this example.

In the Sequence-to-Sequence Classification Using 1-D Convolutions example, network performs convolution over the time dimension of the input data (also known as 1-D convolutions), using the dlconv function. Currently, we support convolution2dLayer and convolution3dLayer.

Hope it helps!