libtorch: Why am I getting "c10::Error at memory location 0x000000C5539CC010." when calling torch::nll_loss()?

Question

Update: I added a Minimal Reproducible Example below the original question.

I've been debugging this error for some time now. It occurs when I call, torch::nll_loss(). I thought it was probably because my tensors didn't match, but I spent some time making sure they're the same size, and I'm not sure what else to check. The libtorch errors I'm getting are pretty unhelpful, unlike the errors provided in the python pytorch. (All of my libtorch errors so far have been Error at memory location 0x............)

I have two Tensors, prediction and target.

Most of the implementation came from this example.

The input Tensor is generated with this method:

void NeuralNetwork::Load(Waveform waveform)
{
    // Convert waveform.samples into torch::Tensor
    c10::DeviceType deviceType;
    if (torch::cuda::is_available()) {
        deviceType = torch::kCUDA;
    }
    else {
        deviceType = torch::kCPU;
    }
    float newArr[1][WAVEFORM_SIZE] = { {0} };
    for(int i=0;i<WAVEFORM_SIZE;i++)
        newArr[0][i] = waveform.samples[i];
    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(deviceType);
    torch::Tensor inputTensor = torch::from_blob(newArr, { 1, WAVEFORM_SIZE }, options);
    inputTensor.set_requires_grad(true);
    TrainingSample newData = TrainingSample(inputTensor, waveform);
    trainingData.push_back(newData);
}

prediction is generated with this network:

struct NeuralNetwork::Net : torch::nn::Module {
    int _inputSize;
    Net(int inputSize) {
        _inputSize = inputSize;
        // Construct and register two Linear submodules.
        fc1 = register_module("fc1", torch::nn::Linear(_inputSize, 64));
        fc2 = register_module("fc2", torch::nn::Linear(64, 32));
        fc3 = register_module("fc3", torch::nn::Linear(32, 3));
    }

    // Implement the Net's algorithm.
    torch::Tensor forward(torch::Tensor x) {
        // If tensor is one dimensional, change to batch_size,seq_len (2D)
        int tensorDims = 0;
        for (int i : x.sizes())
            tensorDims++;
        if(tensorDims==1)
            x = torch::unsqueeze(x,0);
        DBG("x.sizes: " + NeuralNetwork::Item2String<c10::IntArrayRef>(x.sizes()));
        DBG("\nx.sizes size: " + NeuralNetwork::Item2String<c10::IntArrayRef>(sizeof(x.sizes()) / sizeof(x.sizes()[0])));
        // Use one of many tensor manipulation functions.
        DBG("\ninput tensor: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::relu(fc1->forward(x));
        DBG("\nafter fc1: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::dropout(x, /*p=*/0.5, /*train=*/is_training());
        DBG("\nafter dropout: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::relu(fc2->forward(x));
        DBG("\nafter fc2: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::log_softmax(fc3->forward(x), /*dim=*/1);
        DBG("\nafter fc3: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        return x;
    }

    // Use one of many "standard library" modules.
    torch::nn::Linear fc1{ nullptr }, fc2{ nullptr }, fc3{ nullptr };
};

target is generated with this method:

torch::Tensor TrainingSample::getRatingTensor()
{
    c10::DeviceType deviceType;
    if (torch::cuda::is_available()) {
        deviceType = torch::kCUDA;
    }
    else {
        deviceType = torch::kCPU;
    }
    float ratingArray[1][3] = { {0} };
    ratingArray[0][(int)waveform.rating] = 1;
    ostringstream os0;
    for (int i = 0;i<(sizeof(ratingArray[0])/sizeof(ratingArray[0][0]));i++) {
        os0 << ratingArray[0][i];
        os0 << ",";
    }
    DBG("ratingArray: \n" + os0.str());
    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(deviceType);
    torch::Tensor ratingTensor = torch::from_blob(ratingArray, { 1, 3 }, options);
    ostringstream os1;
    os1 << ratingTensor[0];
    DBG("ratingTensor: \n" + os1.str());
    return ratingTensor.clone();
}

It's implemented like:

...
    for each(TrainingSample trainingSample in trainingData) {
        // Check trainingSample.sampleTensor's length to make sure it works
        if (trainingSample.sampleTensor.size(1) != WAVEFORM_SIZE) {
            throw std::logic_error("Input must match WAVEFORM_SIZE");            
        }
        // Reset gradients.
        optimizer.zero_grad();
        // Execute the model on the input data.
        torch::Tensor prediction = net->forward(trainingSample.sampleTensor);
        // Compute a loss value to judge the prediction of our model.
        torch::Tensor target = trainingSample.getRatingTensor();

        std::ostringstream os_tensor0;
        os_tensor0 << target;
        DBG("target_val: \n" + os_tensor0.str());

        std::ostringstream os_tensor1;
        os_tensor1 << prediction;
        DBG("prediction_val: \n" + os_tensor1.str());

        torch::Tensor loss = torch::nll_loss(prediction, target);
...

When run, I get the following console output:

x.sizes: [1, 450]

x.sizes size: [2]

input tensor: 
Columns 1 to 6-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 7 to 12-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 13 to 18-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 19 to 24-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 25 to 30-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 31 to 36-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 37 to 42-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 43 to 48-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 49 to 54-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 55 to 60-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 61 to 66-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08

Columns 67 to 72-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08 -6.1297e+27  1.4370e-41

Columns 73 to 78-1.0737e+08 -1.0737e+08  1.3479e+12  2.7606e-43  1.3483e+12  2.7606e-43

Columns 79 to 84 1.5097e+35  4.5908e-41  1.3479e+12  2.7606e-43 -1.0737e+08 -1.0737e+08

Columns 85 to 90-1.0737e+08 -1.0737e+08 -1.0737e+08 -1.0737e+08  1.3479e+12  2.7606e-43

Columns 91 to 96-5.2868e-34  4.5908e-41  1.4063e+24  2.3647e-41  1.3479e+12  2.7606e-43

Columns 97 to 102-1.0737e+08 -1.0737e+08  1.3483e+12  2.7606e-43  1.3483e+12  2.7606e-43

Columns 103 to 108-5.2840e-34  4.5908e-41 -2.5697e+36  6.8103e-43  1.3479e+12  2.7606e-43

Columns 109 to 114 1.3479e+12  2.7606e-43  7.8240e-02  7.2303e-02  5.2409e-02  2.3487e-02

Columns 115 to 120-9.2460e-03 -3.9394e-02  0.0000e+00  0.0000e+00 -9.9383e-02  6.3058e-43

Columns 121 to 126-1.4176e-01 -1.6761e-01 -1.9406e-01 -2.1916e-01 -2.4059e-01 -2.5746e-01

Columns 127 to 132-2.7041e-01 -2.8198e-01 -2.9581e-01 -3.1235e-01 -3.2742e-01 -3.3695e-01

Columns 133 to 138-3.4230e-01 -3.4594e-01 -2.5143e+36  6.8103e-43 -3.5119e-01 -3.5661e-01

Columns 139 to 144-3.6110e-01 -3.6576e-01 -2.5697e+36  6.8103e-43 -2.5697e+36  6.8103e-43

Columns 145 to 150 1.3480e+12  2.7606e-43  1.3480e+12  2.7606e-43  1.3480e+12  2.7606e-43

Columns 151 to 156 1.3480e+12  2.7606e-43  1.3479e+12  2.7606e-43  1.3479e+12  2.7606e-43

Columns 157 to 162 1.3479e+12  2.7606e-43  1.3479e+12  2.7606e-43 -5.4939e-01 -5.4665e-01

Columns 163 to 168-5.3675e-01 -5.1979e-01 -4.9963e-01 -4.8594e-01 -4.8443e-01 -4.9146e-01

Columns 169 to 174-4.9957e-01 -5.0272e-01 -4.9933e-01 -4.9043e-01 -4.7711e-01 -4.6370e-01

Columns 175 to 180-4.5665e-01 -4.5746e-01 -4.6186e-01 -4.6114e-01 -4.4190e-01 -3.9672e-01

Columns 181 to 186-3.4028e-01 -2.9929e-01 -2.8463e-01 -2.8733e-01  1.3480e+12  2.7606e-43

Columns 187 to 192-2.5944e+36  6.8103e-43  1.3479e+12  2.7606e-43 -2.3795e+36  6.8103e-43

Columns 193 to 198-2.9645e+36  6.8103e-43 -2.5944e+36  6.8103e-43  0.0000e+00  0.0000e+00

Columns 199 to 204-2.3795e+36  6.8103e-43  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00

Columns 205 to 210-2.3795e+36  6.8103e-43  0.0000e+00  0.0000e+00  1.3479e+12  2.7606e-43

Columns 211 to 216-2.3795e+36  6.8103e-43  1.4329e-02  1.9215e-02  2.6336e-02  3.6743e-02

Columns 217 to 222 4.7983e-02  5.8675e-02  6.9367e-02  7.8963e-02 -2.9645e+36  6.8103e-43

Columns 223 to 228 0.0000e+00  1.1853e-01 -1.0677e-23  6.8103e-43 -2.8907e-21  6.8103e-43

Columns 229 to 234-4.5850e-21  6.8103e-43 -4.5858e-21  6.8103e-43 -4.5858e-21  6.8103e-43

Columns 235 to 240 1.8583e+19  0.0000e+00 -1.8184e-36  4.5908e-41 -2.8923e-21  6.8103e-43

Columns 241 to 246-1.0742e-23  6.8103e-43 -1.0742e-23  6.8103e-43 -1.0742e-23  6.8103e-43

Columns 247 to 252 4.2970e-01  4.2393e-01 -1.1069e-23  6.8103e-43  0.0000e+00  0.0000e+00

Columns 253 to 258 5.0396e-01  5.0000e-01  0.0000e+00  0.0000e+00 -1.0742e-23  6.8103e-43

Columns 259 to 264 5.1114e-01  5.1451e-01  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00

Columns 265 to 270 0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00 -1.8184e-36  4.5908e-41

Columns 271 to 276 8.9129e+04  1.0350e+00  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00

Columns 277 to 282 0.0000e+00  0.0000e+00  4.6471e-01  4.5735e-01  4.5031e-01  4.4295e-01

Columns 283 to 288 4.3358e-01  4.2181e-01  4.0897e-01  3.9672e-01  3.8583e-01  3.7529e-01

Columns 289 to 294 3.6445e-01  3.5527e-01  3.4973e-01  3.4699e-01  3.4594e-01  3.4637e-01

Columns 295 to 300 3.4666e-01  3.4488e-01  3.4171e-01  3.3831e-01  3.3367e-01  3.2617e-01

Columns 301 to 306 3.1636e-01  3.0843e-01  3.0667e-01  3.1162e-01  3.2197e-01  3.3437e-01

Columns 307 to 312 3.4179e-01  3.3873e-01  3.2635e-01  3.1051e-01  2.9666e-01  2.8553e-01

Columns 313 to 318 2.7514e-01  2.6404e-01  2.5120e-01  2.3579e-01  2.1667e-01  1.9390e-01

Columns 319 to 324 1.7077e-01  1.5234e-01  1.4434e-01  1.4809e-01  1.5381e-01  1.5017e-01

Columns 325 to 330 1.3569e-01  1.1446e-01  9.1649e-02  7.1120e-02  5.3066e-02  3.6940e-02

Columns 331 to 336 2.2677e-02  1.1021e-02  2.9797e-03 -3.2208e-03 -9.0269e-03 -1.4329e-02

Columns 337 to 342-2.2239e-02 -3.3632e-02 -4.4148e-02 -5.2343e-02 -5.9880e-02 -6.9323e-02

Columns 343 to 348-8.2929e-02 -9.8573e-02 -1.1382e-01 -1.2942e-01 -1.4561e-01 -1.6303e-01

Columns 349 to 354-1.8139e-01 -1.9697e-01 -2.0819e-01 -2.1678e-01 -2.2688e-01 -2.4581e-01

Columns 355 to 360-2.7731e-01 -3.1671e-01 -3.5641e-01 -3.9151e-01 -4.2142e-01 -4.4712e-01

Columns 361 to 366-4.7063e-01 -4.9630e-01 -5.2582e-01 -5.5605e-01 -5.8340e-01 -6.0428e-01

Columns 367 to 372-6.1668e-01 -6.2303e-01 -6.2938e-01 -6.4352e-01 -6.6841e-01 -6.9415e-01

Columns 373 to 378-7.0000e-01 -6.7009e-01 -6.0787e-01 -5.2685e-01 -4.4050e-01 -3.7221e-01

Columns 379 to 384-3.4480e-01 -3.5365e-01 -3.7503e-01 -3.9094e-01 -3.9547e-01 -3.9074e-01

Columns 385 to 390-3.8215e-01 -3.7611e-01 -3.7830e-01 -3.9017e-01 -4.0926e-01 -4.2847e-01

Columns 391 to 396-4.3616e-01 -4.2378e-01 -3.9458e-01 -3.6263e-01 -3.4188e-01 -3.3779e-01

Columns 397 to 402-3.4927e-01 -3.6984e-01 -3.8943e-01 -4.0187e-01 -4.0720e-01 -4.0921e-01

Columns 403 to 408-4.1265e-01 -4.1879e-01 -4.2694e-01 -4.3649e-01 -4.4532e-01 -4.5174e-01

Columns 409 to 414-4.5421e-01 -4.5082e-01 -4.4352e-01 -4.3686e-01 -4.3415e-01 -4.3489e-01

Columns 415 to 420-4.3342e-01 -4.2635e-01 -4.1559e-01 -4.0253e-01 -3.8982e-01 -3.8171e-01

Columns 421 to 426-3.7727e-01 -3.7273e-01 -3.6684e-01 -3.6184e-01 -3.6070e-01 -3.6278e-01

Columns 427 to 432-3.6539e-01 -3.6697e-01 -3.6699e-01 -3.6660e-01 -3.6660e-01 -3.6614e-01

Columns 433 to 438-3.6438e-01 -3.6037e-01 -3.5455e-01 -3.4859e-01 -3.4298e-01 -3.3779e-01

Columns 439 to 444-3.3229e-01 -3.2536e-01 -3.1872e-01 -3.1430e-01 -3.1184e-01 -3.0891e-01

Columns 445 to 450-3.0017e-01 -2.8205e-01 -2.5545e-01 -2.2501e-01 -1.9967e-01  0.0000e+00
[ CPUFloatType{1,450} ]

after fc1: 
Columns 1 to 6 0.0000e+00  7.8770e+34  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00

Columns 7 to 12 0.0000e+00  0.0000e+00  1.8537e+35  0.0000e+00  0.0000e+00  0.0000e+00

Columns 13 to 18 1.3528e+35  0.0000e+00  5.2141e+35  0.0000e+00  3.3142e+34  1.8497e+35

Columns 19 to 24 0.0000e+00  0.0000e+00  2.7433e+35  0.0000e+00  2.7988e+35  4.7219e+33

Columns 25 to 30 0.0000e+00  6.9118e+34  1.2590e+35  0.0000e+00  0.0000e+00  0.0000e+00

Columns 31 to 36 1.8690e+35  0.0000e+00  5.8213e+34  0.0000e+00  0.0000e+00  0.0000e+00

Columns 37 to 42 1.7089e+35  0.0000e+00  4.6019e+35  4.6340e+34  7.8131e+34  0.0000e+00

Columns 43 to 48 0.0000e+00  0.0000e+00  5.5950e+35  2.5632e+35  1.1927e+35  0.0000e+00

Columns 49 to 54 0.0000e+00  0.0000e+00  3.2757e+34  0.0000e+00  2.1214e+35  0.0000e+00

Columns 55 to 60 4.9851e+35  0.0000e+00  6.9772e+34  4.5953e+35  1.6406e+34  0.0000e+00

Columns 61 to 64 3.4096e+35  1.2904e+35  0.0000e+00  1.3653e+35
[ CPUFloatType{1,64} ]

after dropout: 
Columns 1 to 6 0.0000e+00  1.5754e+35  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00

Columns 7 to 12 0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00

Columns 13 to 18 2.7056e+35  0.0000e+00  1.0428e+36  0.0000e+00  0.0000e+00  3.6993e+35

Columns 19 to 24 0.0000e+00  0.0000e+00  0.0000e+00  0.0000e+00  5.5975e+35  0.0000e+00

Columns 25 to 30 0.0000e+00  1.3824e+35  2.5180e+35  0.0000e+00  0.0000e+00  0.0000e+00

Columns 31 to 36 0.0000e+00  0.0000e+00  1.1643e+35  0.0000e+00  0.0000e+00  0.0000e+00

Columns 37 to 42 3.4178e+35  0.0000e+00  9.2039e+35  9.2679e+34  1.5626e+35  0.0000e+00

Columns 43 to 48 0.0000e+00  0.0000e+00  0.0000e+00  5.1264e+35  2.3855e+35  0.0000e+00

Columns 49 to 54 0.0000e+00  0.0000e+00  6.5513e+34  0.0000e+00  0.0000e+00  0.0000e+00

Columns 55 to 60 9.9702e+35  0.0000e+00  1.3954e+35  0.0000e+00  0.0000e+00  0.0000e+00

Columns 61 to 64 6.8193e+35  2.5809e+35  0.0000e+00  0.0000e+00
[ CPUFloatType{1,64} ]

after fc2: 
Columns 1 to 6 1.2524e+35  0.0000e+00  0.0000e+00  2.4940e+35  0.0000e+00  2.4364e+35

Columns 7 to 12 2.6166e+35  0.0000e+00  1.7298e+35  0.0000e+00  0.0000e+00  0.0000e+00

Columns 13 to 18 1.2573e+35  7.5406e+34  0.0000e+00  8.5736e+33  0.0000e+00  1.8340e+35

Columns 19 to 24 1.4026e+35  6.5115e+34  0.0000e+00  0.0000e+00  1.6587e+35  0.0000e+00

Columns 25 to 30 6.3401e+34  1.1294e+35  0.0000e+00  8.4040e+34  0.0000e+00  3.9660e+34

Columns 31 to 32 0.0000e+00  0.0000e+00
[ CPUFloatType{1,32} ]

after fc3: 
-6.2916e+34  0.0000e+00 -1.1594e+35
[ CPUFloatType{1,3} ]
ratingArray: 
1,0,0,
ratingTensor: 
 1
 0
 0
[ CPUFloatType{3} ]
target_val: 
 1  0  0
[ CPUFloatType{1,3} ]
prediction_val: 
-6.2916e+34  0.0000e+00 -1.1594e+35
[ CPUFloatType{1,3} ]
Exception thrown at 0x00007FFA0B7D3E49 in AudioPluginHost.exe: Microsoft C++ exception: c10::Error at memory location 0x000000C5539CC010.
Unhandled exception at 0x00007FFA0B7D3E49 in AudioPluginHost.exe: Microsoft C++ exception: c10::Error at memory location 0x000000C5539CC010.

Looking at this, the Tensor that is fed into the network is correctly sized and all of the other sizes appear to be correct as well. The output also appears to be correctly sized (1,3).

Also, target_val -- the Tensor produced by getRatingTensor -- also appears to be correctly sized and its values are correct as well.

I'm working on a JUCE audio plugin, so I'm using projucer to include/link the libtorch libraries, and I'm using Visual Studio 2019 to compile it.

My projucer settings are:

External Libraries to Link:

E:\Programming\Downloads\libtorch\lib\c10.lib
E:\Programming\Downloads\libtorch\lib\c10_cuda.lib
E:\Programming\Downloads\libtorch\lib\caffe2_nvrtc.lib
E:\Programming\Downloads\libtorch\lib\torch.lib
E:\Programming\Downloads\libtorch\lib\torch_cpu.lib
E:\Programming\Downloads\libtorch\lib\torch_cuda.lib

Header Search Paths:

E:\Programming\Downloads\libtorch\include\
E:\Programming\Downloads\libtorch\include\torch\csrc\api\include

Extra Library Search Paths:

E:\Programming\Downloads\libtorch\lib

---

Minimal Reproducible Example

I'm using JUCE 6.0.1 on Windows 10, with Visual Studio 2019

1. In projucer, create new audio plugin, name it Debug
2. In Exporters>Visual Studio 2019: In the field, External Libraries to Link, paste the following:

path\to\libtorch\lib\c10.lib
path\to\libtorch\lib\c10_cuda.lib
path\to\libtorch\lib\caffe2_nvrtc.lib

path\to\libtorch\lib\torch.lib
path\to\libtorch\lib\torch_cpu.lib
path\to\libtorch\lib\torch_cuda.lib

3. Click on "Debug." In the field, Header Search Paths, enter:

path\to\libtorch\include\
path\to\libtorch\include\torch\csrc\api\include

In the field, Extra Library Search Paths, enter:

path\to\libtorch\lib

4. In projucer, click on File Exporter, right click the Source folder and select "Add New CPP & Header File" called, Debug.
5. Now in your audio plugin folder, there's a subfolder at: Debug\Builds\VisualStudio2019\x64\Debug\Standalone Plugin. Create a hardlink from the .dll files in path\to\libtorch\lib to the Standalone Plugin folder.
6. Using the button in projucer, open your project in Visual Studio 2019.
7. In Debug.h, enter:

#pragma once
#include <torch/torch.h>
#include <JuceHeader.h>
#define WAVEFORM_SIZE 450
using namespace std;
using namespace juce;


class NeuralNetwork {
public:
    NeuralNetwork();
    struct Net;
    float trainingData[WAVEFORM_SIZE];
    int epoch = 0;
    void Train();
    torch::Tensor getRatingTensor();
    template <class T>
    static string Item2String(T x);
};

8. In Debug.cpp, enter:

#include "Debug.h"

// Define a new Module.
struct NeuralNetwork::Net : torch::nn::Module {
    int _inputSize;
    Net(int inputSize) {
        _inputSize = inputSize;
        // Construct and register two Linear submodules.
        fc1 = register_module("fc1", torch::nn::Linear(_inputSize, 64));
        fc2 = register_module("fc2", torch::nn::Linear(64, 32));
        fc3 = register_module("fc3", torch::nn::Linear(32, 3));
    }

    // Implement the Net's algorithm.
    torch::Tensor forward(torch::Tensor x) {
        int tensorDims = 0;
        for (int i : x.sizes())
            tensorDims++;
        if (tensorDims == 1)
            x = torch::unsqueeze(x, 0);
        DBG("x.sizes: " + NeuralNetwork::Item2String<c10::IntArrayRef>(x.sizes()));
        DBG("\nx.sizes size: " + NeuralNetwork::Item2String<c10::IntArrayRef>(sizeof(x.sizes()) / sizeof(x.sizes()[0])));
        // Use one of many tensor manipulation functions.
        DBG("\ninput tensor: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::relu(fc1->forward(x));
        DBG("\nafter fc1: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::dropout(x, /*p=*/0.5, /*train=*/is_training());
        DBG("\nafter dropout: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::relu(fc2->forward(x));
        DBG("\nafter fc2: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        x = torch::log_softmax(fc3->forward(x), /*dim=*/1);
        DBG("\nafter fc3: \n" + NeuralNetwork::Item2String<torch::Tensor>(x));
        return x;
    }

    // Use one of many "standard library" modules.
    torch::nn::Linear fc1{ nullptr }, fc2{ nullptr }, fc3{ nullptr };
};

NeuralNetwork::NeuralNetwork()
{
}

void NeuralNetwork::Train()
{
    // Create saw-tooth wave, WAVEFORM_SIZE samples
    for (int i = 0; i < WAVEFORM_SIZE; i++) {
        trainingData[i] = (i - floor((float)WAVEFORM_SIZE / 2)) / ((float)WAVEFORM_SIZE/2);
    }

    // Convert waveform.samples into torch::Tensor
    c10::DeviceType deviceType;
    if (torch::cuda::is_available()) {
        deviceType = torch::kCUDA;
    }
    else {
        deviceType = torch::kCPU;
    }
    // Adjust dimensions (unsqueeze array)
    float newArr[1][WAVEFORM_SIZE] = { {0} };
    for (int i = 0; i < WAVEFORM_SIZE; i++)
        newArr[0][i] = trainingData[i];
    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(deviceType);
    torch::Tensor inputTensor = torch::from_blob(newArr, { 1, WAVEFORM_SIZE }, options);
    inputTensor.set_requires_grad(true);

    // Now we have inputTensor
    // Get target Tensor
    torch::Tensor targetTensor = getRatingTensor();

    // Train model
    // Create a new Net.
    auto net = std::make_shared<Net>(WAVEFORM_SIZE);

    // Instantiate an SGD optimization algorithm to update our Net's parameters.
    torch::optim::SGD optimizer(net->parameters(), /*lr=*/0.01);


    size_t batch_index = 0;

    // Check trainingSample.sampleTensor's length to make sure it works
    if (inputTensor.size(1) != WAVEFORM_SIZE) {
        throw std::logic_error("Input must match WAVEFORM_SIZE");
    }
    // Reset gradients.
    optimizer.zero_grad();
    // Execute the model on the input data.
    torch::Tensor prediction = net->forward(inputTensor);
    // Compute a loss value to judge the prediction of our model.

    std::ostringstream os_tensor0;
    os_tensor0 << targetTensor;
    DBG("target_val: \n" + os_tensor0.str());

    std::ostringstream os_tensor1;
    os_tensor1 << prediction;
    DBG("prediction_val: \n" + os_tensor1.str());

    torch::Tensor loss = torch::nll_loss(prediction, targetTensor);
    // Compute gradients of the loss w.r.t. the parameters of our model.
    loss.backward();
    // Update the parameters based on the calculated gradients.
    optimizer.step();
    // Output the loss and checkpoint every 100 batches.
    if (++batch_index % 100 == 0) {
        std::cout << "Epoch: " << epoch << " | Batch: " << batch_index
            << " | Loss: " << loss.item<float>() << std::endl;
        // Serialize your model periodically as a checkpoint.
        torch::save(net, "net.pt");
    }

}



torch::Tensor NeuralNetwork::getRatingTensor()
{
    int rating = 0; 
    c10::DeviceType deviceType;
    if (torch::cuda::is_available()) {
        deviceType = torch::kCUDA;
    }
    else {
        deviceType = torch::kCPU;
    }
    float ratingArray[1][3] = { {0} };
    ratingArray[0][rating] = 1;
    ostringstream os0;
    for (int i = 0; i < (sizeof(ratingArray[0]) / sizeof(ratingArray[0][0])); i++) {
        os0 << ratingArray[0][i];
        os0 << ",";
    }
    DBG("ratingArray: \n" + os0.str());
    auto options = torch::TensorOptions().dtype(torch::kFloat32).device(deviceType);
    torch::Tensor ratingTensor = torch::from_blob(ratingArray, { 1, 3 }, options);
    ostringstream os1;
    os1 << ratingTensor[0];
    DBG("ratingTensor: \n" + os1.str());
    return ratingTensor.clone();
}

template <class T>
string NeuralNetwork::Item2String(T x) {
    std::ostringstream os;
    os << x;
    return os.str();
}

9. In PluginEditor.cpp, next to the other two #includes, add #include "Debug.h"
10. In the same file, in the constructor, DebugAudioProcessorEditor::DebugAudioProcessorEditor, add the following code:

    NeuralNetwork neuralNet = NeuralNetwork();
    neuralNet.Train();

Answer 1

I have encountered the same error. nll_loss parameter target must be tensor of torch::kLong you have torch::kFloat.

Hope this helps you.