Initialize variables in Tensorflow C++ API

Question

I am having trouble finding a working way to initialize variables in the C++ API for Tensorflow. I am trying to make a basic fully connected single layer neural network for the MNIST dataset.

Currently, what I've found is that the proper way to do it is like this:

TF_CHECK_OK(session->Run({}, {}, {"init_all_vars_op"}, nullptr));

However, using this returns the following error:

tensorflow/cc/example/example.cc:178:71: error: no matching function for call to 'tensorflow::ClientSession::Run(<brace-enclosed initializer list>, <brace-enclosed initializer list>, <brace-enclosed initializer list>, std::nullptr_t)' .

Has anyone encountered this before? If you did, how did you fix it?

EDIT: I should add that if I remove any attempt at initializing the variables I use and just run the session it will compile, but it returns:

2017-06-27 11:31:12.861244: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.

2017-06-27 11:31:12.862007: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.

2017-06-27 11:31:12.862433: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.

2017-06-27 11:31:12.866282: F tensorflow/cc/example/example.cc:180] Non-OK-status: session->Run(feed ,{y_pred} ,&outputs) status: Failed precondition: Attempting to use uninitialized value weights

[[Node: MatMul = MatMul[T=DT_FLOAT, transpose_a=false, transpose_b=false, _device="/job:localhost/replica:0/task:0/cpu:0"](_arg_x_0_0, weights)]]

/usr/bin/bazel exited with code 8

Answer 1

I've managed to find a solution that only works in the context where the initial state of the variables is unimportant. You can use the Assign method to assign a value of 0 initially to the variable, as Assign doesn't require that the variable is initialized before-hand. This initialization should be done before the train loop:

auto init_value = Const(root, 0);

auto assign = Assign(root.WithOpName("assign"), var, init_value);

I've yet to run a test on this code snippet but it compiles in my current source code.

Answer 2

It looks like the code is not compiling with that call to session.run. It could be due to problematic type inference as tensorflow::ClientSession::Run is overloaded with many different signatures. Try replacing the {} initializers with explicit type names.