.Save, .Restore for contrib.learn.DNNClassifier #3306

Closed
jah456 opened this Issue Jul 14, 2016 · 4 comments

Projects

None yet

5 participants

@jah456
jah456 commented Jul 14, 2016

The older functions like skflow.TensorFlowDNNClassifier had methods .save and .restore. These were supposedly migrated over to the contrib.learn functions, but there are no longer save and restore methods that I can find. Is there a new protocol for saving the graph and variables with the new contrib.learn package?

@jmchen-g
Contributor

@ispirmustafa Could you take a look at this please?

@llealgt
llealgt commented Jul 15, 2016 edited

Hi, any news on this? , im having hard times trying to save a DNNClassifier , followed some examples using train.Saver() but it says (No variables to save)

@michaelisard
Member

@ispirmustafa any thoughts? Assigning @martinwicke since @ispirmustafa is not in the TensorFlow org.

@martinwicke
Member

Checkpoints are saved automatically, and if you specify the same directory when you create another Estimator (of the same kind/same parameters), it will load the checkpoints and you can continue training.

There is currently a bug which prevents you from using the restored Estimator for inference without running at least one step of training.

I will close this issue -- for detailed questions, please ask on StackOverflow, we can track questions and answers better over there.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment