Op_kernel.cc:1192 Out of range: Attempted to repeat an empty dataset infinitely

tensorflow

(Ironflood) #1

Hi,

I’m trying out the latest tensorflow version (until now was running an older TF one, from a month ago) and I keep on having the same issues with a larger dataset than toy-ende:

2018-02-04 13:11:24.924583: W tensorflow/core/framework/op_kernel.cc:1192] Out of range: Attempted to repeat an empty dataset infinitely.

When I switch back to the older version (a month ago) I don’t get the same error. The dataset I feed is maybe beyond input limits? It’s a 2GB file with 1m samples of max length 512 per sample.
Running the toy-ende works as expected, I only face the issue with other dataset.

Thank you!


(Guillaume Klein) #2

Hello,

We had success running trainings on datasets with up to 22M examples so dataset sizes should not be an issue.

What values did you assign to:

  • maximum_features_length
  • maximum_labels_length

in your run configuration?

It seems like all your training examples were filtered out. Also, what TensorFlow version are you using?


(Ironflood) #3

Thanks for the quick reply!

In my dataset input length is max 512 and output max is 10.
I tried several config params for length:

maximum_features_length = 512
maximum_labels_length = 10

and
maximum_features_length = 60
maximum_labels_length = 10

Getting same results, with the warning with current version of Open NMT, no issue with 1 month old version.
The TF version is 1.4.1 and I didn’t change it.


(Guillaume Klein) #4

Can you share the complete configuration if that’s possible?

EDIT: I get this message when the training files are actually empty.