add TString deallocation step

[dskkato]

When I compare TString usage in cppflow and tensorflow's unittest, it seems that deallocation step may be required. But I did not check actual implementation of TString, so if I am wrong, please let it go.

https://github.com/tensorflow/tensorflow/blob/2305f9d0a58921987788fbd1d8fa43eda888233f/tensorflow/core/platform/ctstring_test.cc#L34-L68

[ljn917]

I believe this should be merged, though not tested. Thanks.

In addition, we also need to add test cases for the string typed tensor.

[ns-wxin]

My simple test doesn't complain memory leak any more. However, when running against TF_TString {1} model, I got "heap-use-after-free". I can share with you my model if you would like to try it yourself.

The following code doesn't complain memory leak anymore.

int main(int argc, char **argv) {
// string modelPath = argv[1];
string sent1 {"I enjoy taking long walks\n along the beach with my dog."};
string sent2 {"I taking long walks\n along the beach with my dog."};
// auto sentence = cppflow::tensor(text);
auto sentence = cppflow::tensor(sent1);
auto sentence2 = cppflow::tensor(sent2);
std::cout << "tensor: " << sentence << std::endl;
// vector<TF_TString> data = sentence.get_data<TF_TString>();
// cppflow::model model(modelPath);
return 0;
}

However, the following code caused the "heap-use-after-free".
"tensor.h":

 template<>
inline tensor::tensor(const std::string& value) {
TF_TString tstr[1];
TF_TString_Init(&tstr[0]);
TF_TString_Copy(&tstr[0], value.c_str(), value.size());
// *this = tensor(static_cast<enum TF_DataType>(TF_STRING), (void *) tstr, sizeof(tstr), {});
*this = tensor(static_cast<enum TF_DataType>(TF_STRING), (void *) tstr, sizeof(tstr), {1});
TF_TString_Dealloc(&tstr[0]);
}

my test code:

vector<float> generateEmbedding(cppflow::model model, string text) {
// vector<string> textVec(1);
// textVec.push_back(text);
cppflow::tensor sentence = cppflow::tensor(text);
// cppflow::tensor sentence = cppflow::tensor(textVec);
std::vector<cppflow::tensor> output = model({{"serving_default_sentences:0", sentence}},{"StatefulPartitionedCall_2:0"});
// vector<TF_TString> rawData = sentence.get_data<TF_TString>();
return convertModelResult(output);
}

[serizba]

This solution does not work, as it deallocates before using the tensor.