[WIP] Implement Tensor::unfold(dim, size, step): 1-dim unfolded window views.

[crutcher]

This manipulates layout to unfold a window along a dimension.

See: https://docs.pytorch.org/docs/2.8/generated/torch.Tensor.unfold.html

TODO: I need help with backprop.rs::Tensor::backward()

[crutcher]

I'm working to thread unfold through rust tensor backends, to support lower-memory spatial transforms; particularly those being used by burn.

I'm new to candle, and don't know how to write the backprop code; could use a suggested impl or some pointers.

[crutcher]

A heavier handed approach would be something like Tensor::with_layout(); which I understand is a scary API to provide, though it is actually pretty simple to validate. Any given layout has an easy to compute max and min ravel offset; and the test is just "is that inside the storage size?"

[crutcher]

I have a broken impl of backprop; I'm trying to debug it, but it never runs? there's an assert in it that never triggers.

[crutcher]

@ivarflakstad ping?

[greenrazer]

Hi @crutcher,

In your test instead of:

let data = &[[0f32, 1., 2., 3., 4.], [5f32, 6., 7., 8., 9.]];
    let x = Tensor::new(data, device)?;

you should do:

    let data = &[[0f32, 1., 2., 3., 4.], [5f32, 6., 7., 8., 9.]];
    let v = Var::new(data, device)?;
    let x = v.as_tensor();

This sets the variable field on the tensor to true, which in turn makes x.track_op() return true. As a result, BackpropOp::new1(self, |arg| Op::Unfold(arg, dim, size, step)) returns Some instead of None, which allows sorted_nodes to return a non-empty list.

With this changeassert!(false, "never runs"); should execute