Add TensorFrom trait to create tensors from both vectors and arrays. #414
Conversation
|
Would it be possible to have stack-allocated tensors if they're small enough? |
|
Currently, no, because StridedArrays store tensors as an |
examples/06-mnist.rs
Outdated
| let mut lbl = dev.zeros(); | ||
| lbl.copy_from(&lbl_data); | ||
| (img, lbl) | ||
| (dev.tensor_from_vec(img_data), dev.tensor_from_vec(lbl_data)) |
src/tensor/storage_traits.rs
Outdated
| /// # let dev: Cpu = Default::default(); | ||
| /// let _ = dev.dynamic_tensor_from_vec(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], [2, 3]); | ||
| /// ``` | ||
| fn dynamic_tensor_from_vec<S: Shape>(&self, src: Vec<E>, shape: S) -> Tensor<S, E, Self> { |
There was a problem hiding this comment.
Thoughts on these names?
tensor_from_vec_with_shapetensor_from_shaped_vectensor_from_vec_like(similar tozeros_like/ones_like/etc, however I do think that is less clear in this case)
I think the with_shape one is the most clear about how its different from the other calls of this trait, and how you call it. Usage of dynamic doesn't necessarily imply you have to give a shape, but I do like that it implies it should be used with runtime shapes.
There was a problem hiding this comment.
Yeah, I've gone with tensor_from_vec_with_shape, but I'm not super happy that the names are as long as they are. It's kind of a shame that rust doesn't seem to have the features necessary to have the features necessary for us to merge the TensorFromVec and TensorFromArray traits while still allowing users to use these traits with a generic Device<E>.
There was a problem hiding this comment.
Hmm could we do something like this?
trait TryTensor<Src, S: Shape, E: Unit>: DeviceStorage {
fn tensor(&self, src: Src) -> Tensor<S, E, Self>;
}
impl<E, S: ConstShape> TryTensor<Vec<E>, S, E> for ... {}
impl<E, S: Shape> TryTensor<(Vec<E>, S), S, E> for ... {}
impl<E, S: ConstShape, Src: RustArray<Shape=S>> TryTensor<Src, S, E> for ... {}which would look like:
let a: Tensor<Rank1<3>, ...> = dev.tensor([1.0, 2.0, 3.0]);
let a: Tensor<Rank1<3>, ...> = dev.tensor(vec![1.0, 2, 3]);
// note: tuple
let a: Tensor<(usize, ), ...> = dev.tensor((vec![1.0, 2.0, 3.0], (3, ));There was a problem hiding this comment.
That's a very interesting idea actually, but I don't know that its possible to have all of these implemented for a generic Device, so I'll have to experiment.
|
Per your previous comment, I've greatly simplified the api, and it is now possible to create tensors from arrays and vectors, even with a generic Device, with only the |
Draft of a solution for #399, which splits TensorFromArray into ConstShapeTensorFrom, DynamicTensorFrom, and the public-facing TensorFrom. Currently, this doesn't work how I'd like it to, as this cannot be used when the device is a generic
Devicebecause the traits need to be parameterized by a Shape, which means that it can't be added to the definition ofDevice.