Variables not initialising

[joeyearsley]

System:

TF 1.2 Python 3.5 Ubuntu 16.04

Problem:

We create the preprocessing and model graphs in a script (we'll call it build_graph.py) then go to call the actual tf.run in another script (we'll call train func) at the end of the script.

The Hyperdash monitor is a decorator for the train func, but when we call this script it complains the initial Variables aren't initialised, however, when we run the scripts without hyperdash there are no such complaints.

TL;DR

Variable initialisation problem

[richardartoul]

Hey Joe,

Could you test something for me?

In the decorated train func, could you add the following lines of code:

graph = tf.get_default_graph()
with graph.as_default():
  // Rest of your code

Taken from here: https://github.com/fchollet/keras/issues/2397

I know you're not using Keras, but I wonder if this is just a general issue with Tensorflow and multi-threading

[richardartoul]

@joeyearsley did that work for you?