How to Prepare Inflow/Outflow Data as Tensors for PyTorch Dataset and Dataloader?

Question

I’m working with a bike-sharing dataset that tracks trips between stations and I'm wondering if there's a better way to feed the (time series) data into a neural network using PyTorch.
Each record includes:

• rent_station and return_station (start and end points of a trip)
• rent_time and return_time (the hour when the bike was rented/returned for the trip)
• count (number of trips with the same parameters).

My goal is to create inflow and outflow matrices for each time slot (e.g., hourly), where:

• The inflow matrix (n x n) for an hour t whose i, j-th element represents the number of bikes that is rented at station i at time t and later returned at station j, where n is the number of stations.
• The outflow matrix is defined similarly.

Then for a given d, I would then feed the matrices from t-d to t-1 into my model. Here's my current approach:

1. Matrix Representation:

   • For each time slot, I’ll generate an n x n matrix for both inflow and outflow.
   • Then, concatenate these matrices across all time slots to create two tensors:
     • T x n x n for inflow.
     • T x n x n for outflow.
   • Here, T is the total number of periods (hours) in the data.

2. Dataset and Dataloader:

   • The dataset should index into these tensors to retrieve:
     • A sequence of matrices for the past ( k ) time slots (short-term dependency).
     • The same matrices for the same time slots across the past ( d ) days (long-term dependency).
   • The dataloader will batch this data for training a spatial-temporal graph neural network (STGNN).

My Question:

• Is this approach of turning the data into T x n x n tensors the best practice for this type of task?
• What’s the most efficient way to structure the PyTorch Dataset and Dataloader to index these tensors for short-term and long-term dependencies?

Any advice or code snippets would be greatly appreciated!