This is default in newer Tensorflow versions. Sess can now be initialized before the graph is build and the graph can be obtained using sess.graph. Hence the changes offers better workflow.

If an init function is defined in `hypes['init_function']` the corresponding function will be called instead of the default global initializer. This is important for finetuning.

I am not to happy with the current implementation. I will think of a better way to generalize this.