Sequence region classification with RNN/LSTM?

Question

Neural networks

There is data from the sensors in the form of a sequence (exercise). The repetition of each exercise is marked with boundaries. Each repetition has a different length from 0.4 sec to 15 sec and above (cannot be tied to a fixed length).

How to count the number of repetitions of an action using LSTM classification?
Just to classify what action is currently taking place - it turns out with a probability of 92%. But the task is precisely to find how many times the exercise was performed.
As far as I understand, the task is similar to speech recognition.

Answer 1

Can't you just calculate the minimums? You have the boundaries of the exercise coincide with the minima of the functions, it seems.
If you attach a neuron here, then you will have another problem :)

Answer 2

Is it necessary to do this with the help of RNN or LSTM?
Convolutional neurons, after all, have the ability to train the patterns themselves (convolutions), i.e. just the kind of thing that fires the output neuron when the pattern correlates the most with the data below it. You can simply count the number of firings of the convolutional neuron, thereby determining the number of repetitions.