Define a feed_dict in c++ for Tensorflow models

Question

This question is related to this one: Export Tensorflow graphs from Python for use in C++

I'm trying to export a Tensorflow model from Python to C++. The problem is, my neural net starts with a placeholder to receive input, which requires a feed_dict. I cannot find any c++ API to supply a feed_dict for my model. What can I do?

If there's no API for supplying feed_dicts, how should I change my model so that it can be trained and exported for c++ purposes without placeholders?

Answer 1

The tensorflow::Session::Run() method is the C++ equivalent of the Python tf.Session.run() method, and it supports feeding tensors using the inputs argument. Like so many things in C++ versus Python, it's just a little more tricky to use (and in this case it looks like the documentation is a bit poorer...).

The inputs argument has type const std::vector<std::pair<string, Tensor>>&. Let's break this down:

• Each element of inputs corresponds to a single tensor (such as a placeholder) that you want to feed in the Run() call. An element has type std::pair<string, Tensor>.

• The first element of the std::pair<string, Tensor> is the name of the tensor in the graph that you want to feed. For example, let's say in Python you had:

  p = tf.placeholder(..., name="placeholder")
  # ...
  sess.run(..., feed_dict={p: ...})
  

  ...then in C++ the first element of the pair would be the value of p.name, which in this case would be "placeholder:0"

• The second element of the std::pair<string, Tensor> is the value that you want to feed, as a tensorflow::Tensor object. You have to build this yourself in C++, and it's a bit more complicated that defining a Numpy array or a Python object, but here's an example of how to specify a 2 x 2 matrix:

  using tensorflow::Tensor;
  using tensorflow::TensorShape;
  
  Tensor t(DT_FLOAT, TensorShape({2, 2}));
  auto t_matrix = t.matrix<float>();
  t_matrix(0, 0) = 1.0;
  t_matrix(0, 1) = 0.0;
  t_matrix(1, 0) = 0.0;
  t_matrix(1, 1) = 1.0;
  

  ...and you can then pass t as the second element of the pair.