Currently, experiment configuration is hardcoded in (and duplicated across) training scripts. This is undesirable, as user-specific parameters can end up in them like so, which will generate errors when run by others.
It would be useful to separate these. I propose a .ini file plus a module that pushes the necessary configuration flags into tf.app.flags.FLAGS.
Currently, experiment configuration is hardcoded in (and duplicated across) training scripts. This is undesirable, as user-specific parameters can end up in them like so, which will generate errors when run by others.
It would be useful to separate these. I propose a .ini file plus a module that pushes the necessary configuration flags into tf.app.flags.FLAGS.