Autodiffbackend

[wangjiawen2013]

Hi,
What's the difference between NdArray and Autodiff<NdArray ? Are they different backends? I thought the backend would be determined automatically when setting device as NdArrayDevice::Cpu, but it seems that we also need to set NdArray or Autodiff<NdArray> explicitly even if the device is NdArrayDevice::Cpu.

[loganbnielsen]

Autodiff implements Backend and so does NdArray. But Autodiff backend has an inner backend that knows how to do the computation. (Autodiff is just for computing gradients). A device is not a backend but is associated with a backend

pub trait Backend:  // <-- A backend
    FloatTensorOps<Self>
    + BoolTensorOps<Self>
    + IntTensorOps<Self>
    ...
    + 'static
{
    /// Device type.
    type Device: DeviceOps; // <-- has an associated device

        type Backend = Autodiff<Wgpu>;  // <-- specify backend
        let device = WgpuDevice::default(); // <-- specify device 

        let x: Tensor<Backend, 1> = Tensor::from_floats([3.0], &device);
        let mut y: Tensor<Backend, 1> = Tensor::from_floats([2.0], &device);

[wangjiawen2013]

As "A device is not a backend but is associated with a backend", can burn infer the backend from the device ?
In the example above, the device is WgpuDevice::default(); , is the compiler smart enough to infer that the backend is wgpu because WgpuDevice is used ?

[wangjiawen2013]

I tested the following code:

let device = NdArrayDevice::default();
let x: Tensor<NdArray, 1> = Tensor::from_floats([3.0], &device);
let x: Tensor<Autodiff<NdArray>, 1> = Tensor::from_floats([3.0], &device);

Both NdArray and Autodiff<NdArray> worked. So the compiler cannot infer the backend is NdArray or Autodiff<NdArray>, we must set it explicitly, even the NdArrayDevice::default() is given because NdArray and Autodiff<NdArray> used the same device. Am I right ?

[laggui]

Yep! The device can be inferred from the backend, but not the other way around. NdArray is a concrete backend implementation, and Autodiff also implements Backend but uses the inner backend to perform the tensor operations and adds auto-differentiation capabilities to the backend used. Autodiff is intended to be used as a decorator/wrapper, which means you have to be explicit about when to use it. This is intended so you specify explicitly when you want to track the operations graph, otherwise you'd be wasting resources for nothing 🙂

See also: #2415