UnicodeEncodeError when deploying Nemo Text Normalization on Triton Inference Server

Describe the bug

I have a problem when deloy Text Nemo Text Normarlzation on Triton Inference Server. The Normalizer can not load the default “whitelist.tsv” file. Whenever I run tritonserver --model-repository models/, the bug will appear.

from nemo_text_processing.text_normalization.normalize import Normalizer
normalizer = Normalizer(input_case='cased', lang='en')

Results:

Internal: UnicodeDecodeError: 'ascii' codec can't decode byte 0xce in position 150: ordinal not in range(128)

At:
  /usr/lib/python3.8/encodings/ascii.py(26): decode
  /usr/local/lib/python3.8/dist-packages/nemo_text_processing/text_normalization/en/utils.py(41): load_labels
  /usr/local/lib/python3.8/dist-packages/nemo_text_processing/text_normalization/en/taggers/whitelist.py(49): _get_whitelist_graph
  /usr/local/lib/python3.8/dist-packages/nemo_text_processing/text_normalization/en/taggers/whitelist.py(64): __init__
  /usr/local/lib/python3.8/dist-packages/nemo_text_processing/text_normalization/en/taggers/tokenize_and_classify.py(107): __init__
  /usr/local/lib/python3.8/dist-packages/nemo_text_processing/text_normalization/normalize.py(76): __init__

Environment overview (please complete the following information)

• Environment: Triton Inference Server 21.10
• Method of NeMo install: pip install nemo_toolkit[‘all’]

Additional context

There is no problem when I test this code without Triton Inference Server!.

[VahidooX]

Would you please try running your code with “export PYTHONIOENCODING=utf8”?

[xuanbinh-nguyen96]

I tried fixing the “load labels(abs_path)” function in the “nemo_text_processing/text normalization/en/utils.py” file at the line 40.

label_tsv = open(abs_path) >>> label_tsv = open(abs_path, encoding='utf-8')

And the problem was solved. I know it’s not a good idea to do this with libraries that have been installed directly through pip. I hope that this can be fixed in future updates.