NDShape shape({ 3, 2 });
Variable W = InputVariable(shape, DataType::Float, true);
auto data = make_shared<vector<float>>(3, float(2));
for (int j = 0; j < shape[0]; j++)
{
    data->push_back(2 + 1);
}
auto val = MakeSharedObject<Value>(MakeSharedObject<NDArrayView>(shape, data->data(), data->size(), DeviceDescriptor::CPUDevice()));

// | 2 2 2 | when transposed after
// | 3 3 3 |

Variable x = InputVariable(NDShape({ 3, 1 }), DataType::Float, true);
auto data = make_shared<vector<float>>(3, float(1));
auto valx = MakeSharedObject<Value>(MakeSharedObject<NDArrayView>(NDShape({ 3, 1 }), data->data(), data->size(), DeviceDescriptor::CPUDevice()));

// [1 1 1]

auto left = Transpose(W);
auto op = Times(left, x);

std::unordered_map<Variable, ValuePtr> outputValues = { { op->Output(), nullptr } };
auto back = op->Forward({ {W, val}, {x, valx} }, outputValues, DeviceDescriptor::UseDefaultDevice(), { op->Output() });
auto outVal = outputValues[op->Output()];
wcout << outVal->AsString() << "n";
printMatrix(outVal->Data()->DeepClone(DeviceDescriptor::CPUDevice()));
fflush(stdout);

//prints out correctly [6 9]

ValuePtr rootGrad = MakeSharedObject<Value>(MakeSharedObject<NDArrayView>(DataType::Float, op->Output().Shape(), DeviceDescriptor::UseDefaultDevice()));
rootGrad->Data()->SetValue(1.0f);
unordered_map<Variable, ValuePtr> gradientOut = { { x, nullptr } };
op->Backward(back, { { op->Output(), rootGrad } }, gradientOut);
auto outGrad = gradientOut[x];
wcout << outGrad->AsString() << "n";
printMatrix(outGrad->Data()->DeepClone(DeviceDescriptor::CPUDevice()));
fflush(stdout);

//[5 5 5] !!!
// and not
// |2 2 2|
// |3 3 3| !!!