Convert variable to list of variables without breaking Autograd?

1

Hi there,

do you know if there is a way to convert variable to list of variables without breaking Autograd?

for example, I have variable T whose shape is (3,5), and T.require_grad is True. Now I want to convert T to a list of list. I know I could use T.data.tolist(), but this will break the autograd computation graph. I also do not want to do two for loops to get the elements from T and put them into the list one by one since the efficiency would be very bad.

Is there anyone know how to do this efficiently without breaking the Autograd?

Thanks!

2

You are looking for torch.chunk or torch.split. For example, you have tensor x of size [8, 2, 2], you want to convert it into two tensor of size [4, 2, 2] along the dim 0, you can use the following:

torch.split(x, 4, dim=0)
# or you can use torch.chunk(x, 2, dim=0)
3

Thanks! It works for me.