* Add name to logger.autodir for different runs and simplify
similarity learning (incl. removing tf.FLAGS in favor of argparse)

* remove 'required' from mnist-embedding.py

* next fix

* c'mon travis

* Update logger.py

* fix linting..

It was always confusing, that most functions use `name` and
PrintData expected `label`.

* Add ignore to SaverRestore

When fine-tuning a model, one might want to skip some tensors from
restoring.

The following example will not load the learning rate from the
checkpoint file:

  config.session_init = SaverRestore(args.load,
      ignore=['learning_rate'])

* Update sessinit.py

A binary representation `"\n" * 5 + "hh".encode('utf-8')` reproduces
the error from travis-ci.