[SOLVED] passing one image from batch to function in c++

passing one image from batch to function in c++

I'm trying to change a code in this repo from 2D to 3D; https://github.com/sadeepj/crfasrnn_keras

However, my C++ is super rusty and I'm having a hard time with one problem. I'm trying to pass the Tensor& out to this function;

void ModifiedPermutohedral::compute(Tensor& out, const Tensor& in, int value_size, bool reverse, bool add) const

with this line

ModifiedPermutohedral mp;
const Tensor& input_tensor = context->input(0);
Tensor* output_tensor = NULL;
OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(), &output_tensor));
mp.compute(output_tensor->SubSlice(b), input_tensor.SubSlice(b), channels, backwards_);

where b is the batch index I want to pass to this compute function. However, it gives me an error that

error: cannot bind non-const lvalue reference of type ‘tensorflow::Tensor&’ to an rvalue of type ‘tensorflow::Tensor’

I believe it's because output_tensor is a pointer, but how am I supposed to pass it then? Shouldn't the SubSlice function return a tensor? I've tried

&(output_tensor->SubSlice(b))
*(output_tensor->SubSlice(b))

too but all creates a different error. Can anyone provide insights on how I should pass this? Thank you!

Solution

Although I am not familiar with tensorflow C++, I noticed that in mp.compute(output_tensor->SubSlice(b), input_tensor.SubSlice(b), channels, backwards_);, first argument is a rvalue such that which cannot be used for value assigning purpose. I suggest:

auto sliced_putput = output_tensor->SubSlice(b);
mp.compute(sliced_output, input_tensor.SubSlice(b), channels, backwards_);
//Assign sliced output to the output_tensor's appropriate dimensiton