TensorFlow C++ API：如何用FileWriter查看TensorBoard及获取对应张量名

Hey there! Let's break down your two TensorFlow C++ API questions clearly, with code examples to make things concrete.

To log summaries to TensorBoard using the C++ API, you'll work with FileWriter alongside Summary and Event protobuf objects. Here's a straightforward walkthrough:

• Include necessary headers:

  #include "tensorflow/core/framework/summary.pb.h"
  #include "tensorflow/core/util/event.pb.h"
  #include "tensorflow/core/platform/env.h"
  #include "tensorflow/core/graph/graph_def.h"
  

• Initialize the FileWriter:
  Specify the directory where TensorBoard will pull logs from—this writer will create the directory automatically if it doesn't exist.

  std::string log_dir = "./tensorboard_logs";
  std::unique_ptr<tensorflow::FileWriter> writer(
      tensorflow::FileWriter::New(log_dir, tensorflow::Env::Default()).release());
  

• Create and write a Summary entry:
  Let's use a scalar value (like training loss) as an example. Build a Summary::Value object, attach it to a Summary proto, wrap that in an Event, and write it to disk:

  int step = 0;
  float loss = 0.45f;
  
  // Build scalar summary
  tensorflow::Summary summary;
  auto* value = summary.add_value();
  value->set_tag("training/loss"); // This is the name TensorBoard will display
  value->set_simple_value(loss);
  
  // Wrap in Event and write
  tensorflow::Event event;
  event.set_step(step);
  event.set_wall_time(tensorflow::Env::Default()->NowSeconds());
  *event.mutable_summary() = summary;
  
  writer->WriteEvent(event);
  

• Flush and clean up:
  Make sure to flush the writer to ensure all events are saved to disk, then close it when you're done:

  writer->Flush();
  writer->Close();
  

After running this code, start TensorBoard with tensorboard --logdir=./tensorboard_logs and you'll see the "training/loss" scalar in the Scalars tab.

The FileWriter itself doesn't have tied tensor names—you define the display names (via the tag in Summary::Value) when creating your summary entries. That said, if you need to pull names of existing tensors in your graph (for consistency with your model's structure), here's how:

• If you have direct access to the tensor object:
  Every tensorflow::Tensor has a name() method that returns its full identifier (e.g., "dense_1/BiasAdd:0"). You can use this as the tag in your summary:

  tensorflow::Tensor my_tensor(/* initialize your tensor here */);
  std::string tensor_name = my_tensor.name();
  
  // Use this name as the summary tag
  auto* value = summary.add_value();
  value->set_tag("model_tensors/" + tensor_name);
  // Populate the appropriate field (e.g., histogram_value for histograms)
  

• If you need to extract tensor names from a GraphDef:
  If you're working with a pre-built graph (like a saved model), iterate over the graph's nodes to get their output tensor names:

  tensorflow::GraphDef graph_def;
  // Load your graph into graph_def (e.g., from a .pb file)
  
  for (const auto& node : graph_def.node()) {
      // Each node's outputs follow the format "<node_name>:<output_index>"
      for (int i = 0; i < node.output_size(); ++i) {
          std::string tensor_name = node.name() + ":" + std::to_string(i);
          std::cout << "Tensor name: " << tensor_name << std::endl;
          // Use this name to create summaries for these tensors
      }
  }
  

When logging non-scalar tensor data (like histograms or images), populate the corresponding field in Summary::Value (e.g., histogram_value() or image_value()) instead of simple_value().

内容的提问来源于stack exchange，提问作者Jianfeng Luo