Here, there is a variable ‘num_unrolls’ in the line 63 and 64 https://github.com/chikuanlin/re3-pytorch/blob/master/training/training.py

What does it mean exactly?

In the paper “Real-Time Recurrent Regression Networks for Visual Tracking of Generic Objects” it describe what it is as the following

for tracking,each input is directly paired with an immediate output. Thus,we can use a training curriculum that begins with few unrolls,and slowly increases the time horizon that the network sees to teach it longer-term relationships. Without the shorter unroll step, the network may take exponentially longer to train, or
may simply never converge. Specifically, we initially train the network with only two unrolls and a mini-batch size of 64. After the loss plateaus, we double the number of unrolls and halve the mini-batch size until a maximum unroll of 32 timesteps and a mini-batch size of 4. Using this curriculum, we do not find it necessary to clip gradients.

I could not understand what kind of process it is I need to apply on Dataset