TF 2.1 Interoperability

[pertschuk]

Is it possible #4 is affecting the accuracy of results in tf 2.1.

I ported the code over with the convert to tf2 script and ran on provided dev_dataset.tf and the model did not accurately rank results.

....
'probs': array([0.9881346 , 0.01186538], dtype=float32), 'label_ids': 0}
{'probs': array([0.9881408 , 0.01185923], dtype=float32), 'label_ids': 0}
{'probs': array([0.9881171 , 0.01188295], dtype=float32), 'label_ids': 0}
{'probs': array([0.9413739 , 0.05862608], dtype=float32), 'label_ids': 0}
{'probs': array([0.46223718, 0.5377629 ], dtype=float32), 'label_ids': 0}
{'probs': array([0.98812693, 0.01187303], dtype=float32), 'label_ids': 0}
{'probs': array([0.9881612 , 0.01183886], dtype=float32), 'label_ids': 0}
{'probs': array([0.98815   , 0.01184994], dtype=float32), 'label_ids': 0}
{'probs': array([0.985965  , 0.01403502], dtype=float32), 'label_ids': 0}
{'probs': array([0.9881386 , 0.01186145], dtype=float32), 'label_ids': 0}
{'probs': array([0.98811185, 0.01188817], dtype=float32), 'label_ids': 0}
{'probs': array([0.988456  , 0.01154405], dtype=float32), 'label_ids': 0}
{'probs': array([0.9881373, 0.0118627], dtype=float32), 'label_ids': 0}
Ranks from duobert: [11 29 18 20  1  9 17 28 26 21  6 13 16 10  7 19 14 27  4  3  0 25 12  8
 15 24  5 23  2 22]
MonoBERT MRR: 1.0 
DuoBERT MRR: 0.047619047619047616