Thanks again for your answer. I will try to make clear what I try to accomplish.
I model tries to estimate how many goals a football team will score in the remainder of a game. For this, all the games are split into 100 timeframes, thus t can take values from 0-99. Now there are the features for every timeframe and the label is the goals the team actually scored in the remaining time. So my Idea was that theta is some kind of “scoring intensity” for this timeframe.
Now your Idea that results in a (100, 3000) tensor would lead to the result that the observed label, for example 3, would be the label for all t, although it should only be the observed result for t=30.
I hope that makes clear what I actually try to model. And I apologize for the lack of domain-specific wording