--decode_to_file does not create output file

[mehmedes]

During inference, I'm not able to create a file containing the inference output.
I've tried --decode_to_file, but no output file is being created...

[lukaszkaiser]

It should be sufficient to just use --decode_from_file=path, it creates a file names path.decodes... where in ... it puts the model name and so on. Did you try that?

[mehmedes]

Actually, I did.
This is my decoding command:

PROBLEM=wmt_ende_bpe32k
MODEL=transformer
HPARAMS=transformer_base

DATA_DIR=$HOME/t2t_data
TMP_DIR=/tmp/t2t_datagen
TRAIN_DIR=$HOME/t2t_train/$PROBLEM/$MODEL-$HPARAMS


BEAM_SIZE=4
ALPHA=0.6

t2t-trainer \
  --data_dir=$DATA_DIR \
  --problems=$PROBLEM \
  --model=$MODEL \
  --hparams_set=$HPARAMS \
  --output_dir=$TRAIN_DIR \
  --train_steps=0 \
  --eval_steps=0 \
  --decode_beam_size=$BEAM_SIZE \
  --decode_alpha=$ALPHA \
  --decode_from_file /.../newsdev2016.tok.bpe.32000.en

But I couldn't find any output file anywhere.

[mehmedes]

Now, I see. It's in the tmp folder. Thanks!

[mehmedes]

While translating a file, I got the following error. What could this be connected with:

File "/usr/local/bin/t2t-trainer", line 83, in <module>
    tf.app.run()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "/usr/local/bin/t2t-trainer", line 79, in main
    schedule=FLAGS.schedule)
  File "/usr/local/lib/python2.7/dist-packages/tensor2tensor/utils/trainer_utils.py", line 240, in run
    run_locally(exp_fn(output_dir))
  File "/usr/local/lib/python2.7/dist-packages/tensor2tensor/utils/trainer_utils.py", line 543, in run_locally
    decode_from_file(estimator, FLAGS.decode_from_file)
  File "/usr/local/lib/python2.7/dist-packages/tensor2tensor/utils/trainer_utils.py", line 645, in decode_from_file
    result_iter = estimator.predict(input_fn=input_fn.next, as_iterable=True)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/util/deprecation.py", line 289, in new_func
    return func(*args, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 590, in predict
    as_iterable=as_iterable)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 883, in _infer_model
    features = self._get_features_from_input_fn(input_fn)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.py", line 863, in _get_features_from_input_fn
    result = input_fn()
  File "/usr/local/lib/python2.7/dist-packages/tensor2tensor/utils/trainer_utils.py", line 725, in _decode_batch_input_fn
    input_ids = vocabulary.encode(inputs)
  File "/usr/local/lib/python2.7/dist-packages/tensor2tensor/data_generators/text_encoder.py", line 132, in encode
    ret = [self._token_to_id[tok] for tok in sentence.strip().split()]
KeyError: '\xc3\x88'

[mehmedes]

O, that's only a character issue while translating a word containing the character "È"!

[cshanbo]

Hi all,
I met the same issues.

1. Indeed, the output will be in a file in TMP directory. But the --decode_to_file doesn't work.
2. the KeyError: '\xc3\x88' exception raised here.

I think this is caused by the OOVs like @mehmedes said. I'm not sure whether the vocab.bpe.32000 in this data can cover the pre-processed training data in English to German translation task.

I used my data, where the training data contains some unknown words out of vocabulary. I instead rewrote the code to something like:

def encode(self, sentence):
    """Converts a space-separated string of tokens to a list of ids."""
    ret = [self._token_to_id[tok] if tok in self._token_to_id \ 
           else self._token_to_id['UNK'] for tok in sentence.strip().split()]
    return ret[::-1] if self._reverse else ret

Where I can ensure the UNK symbol in my vocabulary.
I'm testing this setting to see if it works.
Any advice will be helpful

Thank you

[lukaszkaiser]

Does this only happen with BPE, or with the standard "tokens_32k" too? We don't have a built-in tokenizer for BPE, it was used only for papers to have perplexities comparable with other papers. It cannot be detokenized, so I believe it's better to use our own tokenizer. Or is the problem the same?