Linking protobuf issue in Docker images with saved model

[bogdanteleaga]

This issue seems to come up a lot, especially with local builds, but it seems to happen also within the Docker images provided. A simple program like

#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/public/session.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/cc/saved_model/loader.h"
#include "tensorflow/cc/saved_model/tag_constants.h"

using namespace tensorflow;

int main(int argc, char* argv[]) {

    auto export_dir = "dir";
    // Load
    SavedModelBundleLite model_bundle;
    SessionOptions session_options = SessionOptions();
    RunOptions run_options = RunOptions();
    Status status = LoadSavedModel(session_options, run_options, export_dir, {kSavedModelTagServe}, &model_bundle);

    // Inference
    const std::vector<string> outputLayer = {"Identity:0"};
    tensorflow::Tensor input_tensor(DT_COMPLEX64, TensorShape({1, 512, 1536, 2}));
    std::vector<tensorflow::Tensor> outputs;
    Status runStatus = model_bundle.GetSession()->Run({{"input:0", input_tensor}}, outputLayer, {}, &outputs);
}

Cannot be linked and results in lots of protobuf errors similar to

/usr/bin/ld: CMakeFiles/example.dir/example.o: in function `google::protobuf::Map<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, tensorflow::SignatureDef>::Init()':
example.cpp:(.text._ZN6google8protobuf3MapINSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEEN10tensorflow12SignatureDefEE4InitEv[_ZN6google8protobuf3MapINSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEEN10tensorflow12SignatureDefEE4InitEv]+0x2ea): undefined reference to `google::protobuf::internal::ArenaImpl::AllocateAligned(unsigned long)'

This is using the floopcz/tensorflow_cc:ubuntu image.

[FloopCZ]

Yes yes, that is an annoying issue. One solution could be to detect and add the protobuf version that bazel uses as an external dependency as mentioned by @cweigel in #246. Unfortunately, I have not made any attempts to do that yet, if anyone is interested, I would surely accept a PR. 🙂

[isra60]

Hi, are there any updates on this? I found the same problem using the ubuntu-Cuda docker version

[lancelot-ch]

I can compile these codes using tensorflow_cc with no errors. However, I am using a fork:
https://github.com/mrdaliri/tensorflow_cc