Hi Bram!
As you've noticed, the globalAveragePooling1dLayer plays a critical role here. This is because that layer removes the time dimension by pooling over it globally (i.e. keeping only the largest value in the sequence). This layer is useful for sequence-to-one tasks, where the output isn't a sequence. The output here would be a numClasses x numObservations array.
However, because yours is a sequence-to-sequence problem, you want the output to be a numClasses x numObservations x sequenceLength array (with the sequence dimension conserved). So in that case, the globalAveragePooling1dLayer isn't appropriate for your workflow, because of the fact it removes the sequence dimension.
Hope this helps, and let me know if you have any other questions!
Best,
Tomaso
