Corey Lowman

Yeah you're right - I was finding that for some reason doing `Tensor::of_slice` and `.view` was much slower than just one call to the above, and was looking to maximize...

Whatever the dataset iterator api is, it should also enable things like torchvision.transforms. E.g. I should be able to do something like: ```rust let transforms = (CenterCrop(10, 10), Normalize(...)); for...

Needs to account for things like the RedCaps dataset which only contains URLs for images, and large datasets which many files and indices

Nice work @quietlychris, looks very useful! Will let you know if there's features that would be useful. Definitely need to think through a `trait Dataset` to make writing training loops...

Cool, I think generic associated types will be useful for this: ```rust trait Dataset { type Item; type BatchedItem; } ```

Related to #9

Resource I'll post here from pytorch land: https://dev-discuss.pytorch.org/t/keeping-pytorchs-ops-maintainable-the-jiterator/468

I think this can be done with two separate devices: `CudaJIT` and `Cuda`. They can share the underlyilng kernel code, but their impls can construct them differently

rust-cuda currently does not include a binding to nvrtc, so this will have to be added

Should all device traits be `pub(crate)`?