How to run just a raw tensor through ncnn? I'm getting "incompatible constructor arguments"

[redthing1]

Hello,

I want to pass a raw array/tensor to my model. My model works fine in onnxruntime, wonnx, etc. I want to run it in ncnn:

Tencent#3447

Here is where they said you can do it as such:

float* data_ptr;
Mat in(data_length, data_ptr);

using this constructor:

// external vec
Mat(int w, void* data, size_t elemsize = 4u, Allocator* allocator = 0);

However, I'm trying to do this in python:

input_mat = ncnn.Mat(flat_input_value.shape[0], flat_input_value)

But I get this:

failed to execute ncnn: __init__(): incompatible constructor arguments. The following argument types are supported:
    1. ncnn_vulkan.ncnn.Mat()
    2. ncnn_vulkan.ncnn.Mat(shape: tuple, *, elemsize: int = 4, elempack: int = 1, allocator: ncnn_vulkan.ncnn.Allocator = None)
    3. ncnn_vulkan.ncnn.Mat(w: int, *, elemsize: int = 4, elempack: int = 1, allocator: ncnn_vulkan.ncnn.Allocator = None)
    4. ncnn_vulkan.ncnn.Mat(w: int, h: int, *, elemsize: int = 4, elempack: int = 1, allocator: ncnn_vulkan.ncnn.Allocator = None)
    5. ncnn_vulkan.ncnn.Mat(w: int, h: int, c: int, *, elemsize: int = 4, elempack: int = 1, allocator: ncnn_vulkan.ncnn.Allocator = None)
    6. ncnn_vulkan.ncnn.Mat(w: int, h: int, d: int, c: int, *, elemsize: int = 4, elempack: int = 1, allocator: ncnn_vulkan.ncnn.Allocator = None)
    7. ncnn_vulkan.ncnn.Mat(m: ncnn_vulkan.ncnn.Mat)
    8. ncnn_vulkan.ncnn.Mat(array: buffer)

Invoked with: 150528, array([104.32   , 129.22101, 151.061  , ..., -85.68   , -83.779  ,
       -82.939  ], dtype=float32)
[1]

Any idea how I can do this kind of constructor from Python?

[joeyballentine]

Hey, sorry for the late response on this, but the official ncnn package now supports Vulkan in the python bindings apparently, so I'd recommend switching to that.