multiple batches inference using C++ api

[arthur-s]

Good day, I try to process audio using onnxruntime, the scheme is:
I get first 4000 elements from audio data (it is a chunk), create tensor, and add it into batch vector (witch max size = 200), get next 4000 elements, and so on. So, for 1 minute audio I have 10 batches. I want to process each batch in one session in a loop:

auto session = .. // init session
while(hasSamples){
    batch_input_tensors.clear()
    for (i; batch_input_tensors.size() <= 200; i += 500) {// sliding window
        ... // add samples to windowSamples
        // then add windowSamples to the batch
        batch_input_tensors.push_back(
            Ort::Value::CreateTensor<float>(memoryInfo, windowSamples.data(),windowSamples.size(), input_shape.data(), input_shape.size())
        )

    }
    // batch has size 200, I try to process it:
    session.Run(Ort::RunOptions{nullptr},
                    inputNames.data(),
                    batch_input_tensors.data(),
                    //batch_num,
                    1,
                    outputNames.data(),
                    outputTensors.data(),
                    1);
    
    auto& outputTensor = outputTensors[0];
    const auto* floats = outputTensor.GetTensorData<float>();
    ...
    for (int _i=0; _i < (int)floatsCount; _i += 2)
    {
        std::cout << floats[_i] << std::endl;
    }
}

For the first batch I have realistic values, e.g. I have in stdout

0.998717
0.998713
0.998714
0.998713
0.998715
0.999086
0.999051

But starting from second batch I have the same repeating value 0.94601 (and sometimes 0.946011):

0.94601
0.94601
0.94601
0.94601
0.94601
0.94601
0.94601
0.946011
0.94601
0.94601

Why is this? I tried other audiofiles, result is same - starting from second batch I have 0.94601.

[arthur-s]

Ok, I've fixed it (it's turned out that I wrongly understood tensors).