As of MATLAB R2021a, defining a convolution layer for time series data is not directly supported in Deep Network Designer.
Currently, the 1-D convolution operation is available through the custom training loop workflow with a "model as function" approach and you can find an example available in the doc here: https://www.mathworks.com/help/deeplearning/ug/sequence-to-sequence-classification-using-1-d-convolutions.html
The Development team is aware of this limitation and will consider to address it in future releases.
Solutions that you can try:
- Recurrent Neural Network (lstmLayer, gruLayer, ...), available in Deep Network Designer App
- Adaptation of Feed-Forward Networks, i.e. folding/unfolding of sequences and a fullyConnectedLayer in between, available in Deep Network Designer App.
- Convolutional Neural Network approach as described in the aforementioned example, using a custom training loop (hence not using Deep Network Designer)
Hope this helps!
