如何使用tf.train.Saver（）方法recover_last_checkpoints？

Question

The documentation writes that a list of checkpoint paths should be passed to it, but how to get the list? By hard coding? No, it's a silly practice. By parsing the protocol buffer file (a file named as checkpoint in your model directory)? But tensorflow does not implement a parser, does it? So do I have to implement one by myself? Do you have a good practice to get the checkpoint paths list?

I raise this question because these days I am troubled by one thing. As you know, a days-long training may crash for some reason, and I have to recover it from the latest checkpoint. Recovering training is easy, since I just need to write the following code:

restorer = tf.train.Saver()
restorer.restore(sess, latest_checkpoint)

I can hard code latest_checkpoint , or somewhat wiser, use tf.train.latest_checkpoint() .

However, a problem arises after I recover the training. Those old checkpoints files that are created before crash are left there. The Saver only manages the checkpoint files created in one run. I hope it could also manage the previously created checkpoints files so they would be automatically deleted, and I don't have to manually delete them every time. I think such repeating work is really silly.

Then I find the recover_last_checkpoints method in class tf.train.Saver() , which allows Saver to manage old checkpoints. But it's not handy to use. So is there any good solution?

Answer 1

As mentioned by @isarandi in a comment, the easiest way is to first recover all checkpoint paths using get_checkpoint_state‌ followed by all_model_checkpoi‌​nt_paths , which is basically an undocumented feature. You can then restore your latest state as such:

states = tf.train.get_checkpoint_state‌​(your_checkpoint_dir‌​)
checkpoint_paths = states.all_model_checkpoi‌​nt_paths
saver.recover_last_checkpoints(checkpoint_paths)