Time taken for multi_round_infer.sh to run?

[pidugusundeep]

Iam running the scripts on gitpod and it's taking a long time? May I know how much time it usually takes?

[alexrus]

How big in your test dataset?

[melisa-writer]

Hi, I am facing the same problem. Here are my (unsuccessful) attempts:

• Set batch size to 1 and tried decoding on CPU. For sentences ~10 tokens I am getting ~6 sec decoding time.

• Exported the estimator using tf.estimator.export.ServingInputReceiver and estimator.export_saved_model. The decoding time stayed almost the same.

• Tried decoding on TPU in Colab. I am getting the following error:

  
  INFO:tensorflow:Restoring parameters from PIE_ckpt/pie_model.ckpt
  INFO:tensorflow:Error recorded from prediction_loop: Restoring from checkpoint failed. This is most likely due to a Variable name or other graph key that is missing from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Op type not registered 'MapAndBatchDatasetV2' in binary running on n-dfbb99ed-w-0. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib, accessing (e.g.) tf.contrib.resampler should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed. INFO:tensorflow:prediction_loop marked as finished WARNING:tensorflow:Reraising captured error Traceback (most recent call last): File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py", line 1334, in _do_call return fn(*args) File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py", line 1317, in _run_fn self._extend_graph() File "/usr/local/lib/python3.6/dist-packages/tensorflow/python/client/session.py", line 1352, in _extend_graph tf_session.ExtendSession(self._session) tensorflow.python.framework.errors_impl.NotFoundError: Op type not registered 'MapAndBatchDatasetV2' in binary running on n-dfbb99ed-w-0. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib, accessing (e.g.) tf.contrib.resampler should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed.

Would you mind dropping some hints how to achieve the decoding speed reported in the paper?

[alexrus]

@melisa-qordoba please give more details on how you exported and then imported the estimator.

What was the batch size when you used the exported estimator, still 1? if yes, try with more and see if there is any improvement.

As the paper noted, this was made for accuracy rather than speed.