RNN in TensorFlow realtime implementation

Question

I want to implement a RNN in TensorFlow to track the position of a few objects. I want a realtime implementation so I would give the network a datapoint at the time. The examples I have seen in TensorFlow the input is always a sequence of datapoints at different times.. How can I input only one datapoint at the time?

Answer 1

As the Tensorflow documentation, https://www.tensorflow.org/api_docs/python/tf/nn/dynamic_rnn

• You can specify the initial state as an input for the RNN network,
• Also, Tensorflow RNN networks output the final state.

Therefore, you can give the input as the shape of [batchsize, 1, inputsize] (don't forget the initial state ;) and will receive an output with the following shape [batchsize, 1, outputsize] and the final state.

Then you can input the final state of the previous execution as the initial state to the next execution.

This will allow you to create a RNN that can run in real-time ;)