TensorFlow C++, runtime issue

Question

I have created a model using TensorFlow python. Now I have loaded the model in C++ and run it using session->Run (feeding input tensors). The code is complied and linked well, however as soon as it reaches the session->Run during the run time, It doesn't proceed further, and it doesn't create any error message either! It looks like it runs forever! I also checked the CPU usage, but it does not also indicate any intensive computation!

Answer 1

When a TensorFlow program blocks forever, one common issue is that the step is blocked on an empty queue in a q.dequeue() or q.dequeue_many() op.

One possibility is that your Python model depends on prefetching threads (derived from tf.train.QueueRunner objects). Many of the input reading (e.g. using tf.TFRecordReader), and batching (e.g. using tf.train.batch()) pipelines implicitly create queues and queue runners.

If you had to run tf.train.start_queue_runners() in your Python program, then you will need to do the equivalent thing in your C++ code, by forking threads to run the appropriate q.enqueue() ops. Alternatively, you can prepare the input in your C++ program and feed the graph so that the operations you are trying to run do not depend on dequeuing elements for a queue.