I have a C++ library function addTensor(tensor, type) in Tensor class, tensor being an xarray and type being an enum with FP32, FP16, INT, I8 and U8F as possible types. I also have overloads for each of the types, e.g. addTensorFP32(tensor) etc. I’m trying to get python bindings so that users may only call n.addTensor(a.astype(np.float16)) to add a tensor. I’ve bind the various overloads such as addTensorF32() and the other ones no problem, and I’m writing a dispatch function that detects the type of values in a numpy array and calls the right overload based on that. My progress thus far:

.def("addTensor", [](Tensor&obj, py::array tensor){
        if(py::isinstance<py::array_t<float>>(tensor) || 
           py::isinstance<py::array_t<double>>(tensor)){
            obj.addTensorFP32(tensor.cast<xt::xarray<float>>());
        } else if(py::isinstance<py::array_t<int>>(tensor)) {
            obj.addTensorINT(tensor.cast<xt::xarray<int>>());
        } else if(py::isinstance<py::array_t<float16>>(tensor)) {     // Problem
            obj.addTensorFP16(name, tensor.cast<xt::xarray<float16>>());  // Also problematic
        } else {
            throw std::runtime_error("Unsupported datatype");
        }
    })

Checking the instance of numpy arrays works great for C++ built-ins, but doesn’t work for non-standard types such as float16. Is there an alternative way to check the underlying type? I also tried using dtype but there’s also no option to detect float16.