SequenceinputLayer and Convolution2dLayer

I was told by the deep network designer that the SequenceinputLayer and Convolution2dLayer that these two layers are incompatible. I would like to know why and how I could have an input of a 1x29 double set up with a Convolution2dLayer and have two classifications.

Answer 1

[Zhiyu WANG]

It`s recommended to transform your data to formatted dlarray objects at first.

As for the dlarray object, you can refer to this:

https://ww2.mathworks.cn/help/deeplearning/ref/dlarray.html?s_tid=doc_ta

"Convolution2dLayer" needs at least 2 "spatial" dimensions

"SequenceinputLayer " needs at least 1 "time" dimension

Answer 2

[Katja Mogalle]

From what you write, Chris, my understanding is that you want to perform 1D convolutions on sequence data (e.g. time series). From R2021b onwards, you can do this via sequenceInputLayer combined with convolution1dLayer. For an example, see here: https://uk.mathworks.com/help/deeplearning/ug/sequence-classification-using-1-d-convolutions.html

If this is not what you're trying to do, could you please provide some more context about your problem and what your data dimensions mean? You can also include a screenshot from the network analysis (e.g. using analyzeNetwork or Deep Network Designer).

As Zhiyu WANG mentioned, a convolution2dLayer generally expects two dimensions to convolve over. Typically, these are two spatial dimensions. But from R2022a you can also convolve over a mix of dimensions (e.g. spatial and sequence dimensions).