FacenetClassifier::FacenetClassifier (string operation, string model_path, string classifier_model_path, string labels_file_path) {
this->operation = operation;
this->model_path = model_path;
this->classifier_model_path = classifier_model_path;
this->labels_file_path = labels_file_path;
ReadBinaryProto(tensorflow::Env::Default(), model_path.c_str(), &graph_def);
tensorflow::SessionOptions options;
tensorflow::NewSession (options, &session);
session->Create (graph_def);
}
void TensorFlowEngine::load() {
const auto networkFilename = getFilename();
tensorflow::SessionOptions options;
tensorflow::ConfigProto &config = options.config;
tensorflow::GPUOptions* gpuOptions = config.mutable_gpu_options();
gpuOptions->set_allow_growth(true); // Set this so that tensorflow will not use up all GPU memory
//gpuOptions->set_per_process_gpu_memory_fraction(0.5);
mSession.reset(tensorflow::NewSession(options));
tensorflow::GraphDef tensorflow_graph;
{
reportInfo() << "Loading network file: " << networkFilename << reportEnd();
tensorflow::Status s = ReadBinaryProto(tensorflow::Env::Default(), networkFilename, &tensorflow_graph);
if (!s.ok()) {
throw Exception("Could not read TensorFlow graph file " + networkFilename);
}
}
bool nodesSpecified = true;
int inputCounter = 0;
if(mInputNodes.size() == 0) {
nodesSpecified = false;
}
for(int i = 0; i < tensorflow_graph.node_size(); ++i) {
tensorflow::NodeDef node = tensorflow_graph.node(i);
if(mInputNodes.count(node.name()) > 0) {
}
if(node.op() == "Placeholder") {
if(node.name().find("keras_learning_phase") != std::string::npos) {
//mLearningPhaseTensors.insert(node.name());
mLearningPhaseTensors.push_back(node.name());
} else {
// Input node found:
// Get its shape
// Input nodes use the Op Placeholder
reportInfo() << "Found input node: " << i << " with name " << node.name() << reportEnd();
auto shape = getShape(node);
reportInfo() << "Node has shape " << shape.toString() << reportEnd();
if(mInputNodes.count(node.name()) == 0) {
if(nodesSpecified) {
throw Exception("Encountered unknown node " + node.name());
}
reportInfo() << "Node was not specified by user" << reportEnd();
// If node has not been specified by user, we need to add it
// and thus know its type (fast image or tensor)
// It is assumed to be an image if input shape has at least 4 dimensions
NodeType type = NodeType::TENSOR;
if(shape.getKnownDimensions() >= 2) {
reportInfo() << "Assuming node is an image" << reportEnd();
type = NodeType::IMAGE;
} else {
reportInfo() << "Assuming node is a tensor" << reportEnd();
}
addInputNode(inputCounter, tensorflow_graph.node(0).name(), type, shape);
++inputCounter;
}
// Set its shape
mInputNodes[node.name()].shape = shape;
}
}
}
reportInfo() << "Creating session." << reportEnd();
tensorflow::Status s = mSession->Create(tensorflow_graph);
if (!s.ok()) {
throw Exception("Could not create TensorFlow Graph");
}
//tensorflow::graph::SetDefaultDevice("/gpu:0", &tensorflow_graph);
// Clear the proto to save memory space.
tensorflow_graph.Clear();
reportInfo() << "TensorFlow graph loaded from: " << networkFilename << reportEnd();
setIsLoaded(true);
}