About the evaluation method for the selector.

[ZhichaoOuyang]

Why is the model restored in each evaluation on the selector the model saved in the last iteration? So how to choose the best model in training?

[ZhichaoOuyang]

in util.py:
def load_ckpt(saver, sess, ckpt_dir='train', ckpt_path=None):
"""Load checkpoint from the train directory and restore it to saver and sess, waiting 10 secs in the case of failure. Also returns checkpoint name."""
ckpt_dir = os.path.join(FLAGS.log_root, ckpt_dir)
while True:
try:
if not ckpt_path:
latest_filename = "checkpoint_best" if "eval" in ckpt_dir else None
ckpt_state = tf.train.get_checkpoint_state(ckpt_dir, latest_filename=latest_filename)
ckpt_path = ckpt_state.model_checkpoint_path
tf.logging.info('Loading checkpoint %s', ckpt_path)

The last two line means always load the last iteration model?

[HsuWanTing]

Yes, if you don't give a pretrained model path, it will use the latest checkpoint as default.