💡 Ideas for new features

[cocoa-xu]

Please feel free to share your ideas 💡 for new features!

[RyoWakabayashi]

Easier conversion with Nx tensor

Current

nx_tensor =
  Nx.iota({256, 1}, type: :u8)
  |> Nx.broadcast({256, 256, 3})

nx_tensor
|> Evision.Mat.from_nx_2d()
|> Evision.resize({400, 400})
|> Evision.Mat.to_nx()

Easier

nx_tensor =
  Nx.iota({256, 1}, type: :u8)
  |> Nx.broadcast({256, 256, 3})

Evision.resize(nx_tensor, {400, 400}, to_nx: true)

And, I want to specify the backend.

nx_tensor =
  Nx.iota({256, 1}, type: :u8)
  |> Nx.broadcast({256, 256, 3})

Evision.resize(nx_tensor, {400, 400}, to_nx: EXLA.Backend)

I want to use the nx default backend.

config: [nx: [default_backend: EXLA.Backend]]

[cocoa-xu]

Actually, I think that perhaps I can return a tensor that is on the same backend. For example,

exla_tensor =
  Nx.iota({256, 1}, type: :u8)
  |> Nx.broadcast({256, 256, 3})

result_exla_tensor = 
  Evision.resize(exla_tensor, {400, 400}, from_nx_2d: true)

And to switch to another backend, we can explicitly call Nx.backend_transfer/2

exla_tensor =
  Nx.iota({256, 1}, type: :u8)
  |> Nx.broadcast({256, 256, 3})

result_torch_tensor = 
  Evision.resize(exla_tensor, {400, 400}, from_nx_2d: true)
  |> Nx.backend_transfer(Torchx.Backend)

Or as you proposed, we can do both.

exla_tensor =
  Nx.iota({256, 1}, type: :u8)
  |> Nx.broadcast({256, 256, 3})

torch_tensor_1 = 
  Evision.resize(exla_tensor, {400, 400}, from_nx_2d: true, to_nx: Torchx.Backend)

exla_tensor_2 = 
  Evision.imread("demo.png", to_nx: EXLA.Backend) # this returns an EXLA tensor instead of an Evision.Mat

exla_tensor_3 = 
  Evision.imread("demo.png", to_nx: EXLA.Backend) # this returns an EXLA tensor
  |> Evision.resize({400, 400}, from_nx_2d: true) # and it should stay as an EXLA tensor

exla_tensor_4 = 
  Evision.imread("demo.png")                         # <= this returns an Evision.Mat
  |> Evision.resize({400, 400}, to_nx: EXLA.Backend) # and we can pass `to_nx: EXLA.Backend`
                                                     # so that it returns an EXLA tensor 

What do you think?

[RyoWakabayashi]

Returning the same backend tensor looks very good!

exla_tensor_2 = 
  Evision.imread("demo.png", to_nx: EXLA.Backend) # this returns an EXLA tensor instead of an Evision.Mat

exla_tensor_3 = 
  Evision.imread("demo.png", to_nx: EXLA.Backend) # this returns an EXLA tensor
  |> Evision.resize({400, 400}, from_nx_2d: true) # and it should stay as an EXLA tensor

This style looks very good!

[cocoa-xu]

I found that there are some difficulties in applying this universally. I'll look into this further!

[RyoWakabayashi]

The following warning appears during a build when CUDA is enabled.

nvcc warning : The 'compute_35', 'compute_37', 'compute_50', 'sm_35', 'sm_37' and 'sm_50' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).

Exempting older architectures should reduce build time.

opencv/opencv#20576
https://gist.github.com/woosal1337/eef9cd11a29d50a09851502b22a3e064