I have the following question: I would like to train a model, using TensorFlow, that receives some time series (with different lengths) and extract some features for an estimation task (i.e., I would like to map a time series to a scalar value).

However, I don’t wanna to use the whole time series. Instead, I would like the model to first clip the time series, by learning some start and end point, like a mask. Is this possible?

So far, I tried to manually clip the time series in zones of interest, but I would like the segmentation to be automatic. I’ve also tried random clips, but this did not helped a lot to solve the problem.