I've noticed that the examples provided in tensorflow/models repo don't work out of the box with the newest DRAGNN version and provided pre-trained models.
Would be helpful to add an example DRAGNN wrapper that works with CoNLL17 models. It should be implemented as an easy to use abstraction to get new users up and running with real applications fast.
I've noticed that the examples provided in tensorflow/models repo don't work out of the box with the newest DRAGNN version and provided pre-trained models.
Would be helpful to add an example DRAGNN wrapper that works with CoNLL17 models. It should be implemented as an easy to use abstraction to get new users up and running with real applications fast.