Which clone method should I use to create a deep copy of tensor

C++
1

How to create a copy of tensorInputs with the same values but different addresses for each data point in the following code? a search of the documentation brought me to many possible clone methods, and none of them have detail documentation about how to use them.

The reason for clone is because I want to run the following code in a loop and I need to avoid data racing

            at::Tensor tensorInputs = torch::from_blob((void*)wavBuffer, {1, bufLength}, at::kFloat);


            float* floatInput = tensorInputs.data_ptr<float>();
            if (!floatInput) {
                return nil;
            }
   
            c10::InferenceMode guard;
            CFTimeInterval startTime = CACurrentMediaTime();
            auto result = _impl.forward({ tensorInputs }).toStringRef();
            CFTimeInterval elapsedTime = CACurrentMediaTime() - startTime;
            NSLog(@"inference time:%f", elapsedTime);
                
            return [NSString stringWithCString:result.c_str() encoding:[NSString defaultCStringEncoding]];

I tried the following from this post and it doesn’t work:
tensorCopy = tensorInputs.detach().clone(LEGACY_CONTIGUOUS_MEMORY_FORMAT);

2

Calling clone on the tensor sounds like the right approach. Could you describe the issues you are seeing using it?

4

data race still happened because I’m cloning tensorInputs while modifying it in the _impl.forward function. I got the error:

Thread 17: EXC_BAD_ACCESS (code=1, address=0x14a585000)

5

If you clone tensorInputs via:

at::Tensor tensorInputs = torch::from_blob((void*)wavBuffer, {1, bufLength}, at::kFloat).clone();

the object will stay alive until it gets out of scope.
I don’t understand how you are modifying it inside the forward method but maybe you want to clone it there so the object stays alive?