Allennlp: Prediction output changes with same input as while calling multiple times.

[2509dk]

I am using ner-model-2020.02.10.tar.gz model and i would like to predict.
but when i use the allen nlp object in loop then its behaviour is changing. below is the sample code

import re

from allennlp.predictors.predictor import Predictor
ner_var = Predictor.from_path("/home/dheeraj/Desktop/allenNlp/ner-model-2020.02.10.tar.gz")

org_obj = ner_var

text = [' Client: TechMahindra/UPC, India ',
        ' Client: TechMahindra/UPC, India ']

for i in text:
    org = org_obj.predict(i)
    print(org['tags'])
    print(org['words'])
`

Output:-
['O', 'O', 'U-LOC', 'O', 'O', 'O', 'U-LOC']
['Client', ':', 'TechMahindra', '/', 'UPC', ',', 'India']
['O', 'O', 'O', 'O', 'O', 'O', 'U-LOC']
['Client', ':', 'TechMahindra', '/', 'UPC', ',', 'India']

above you can see , for same input , for first line(TechMahindra) tagged as U-LOC but in second line (same line) you can see different output.

So why the behaviour is not consistent. as the output should be same.
please help on this.

[epwalsh]

This model uses ELMO, which is non-deterministic across runs.

[2509dk]

@epwalsh do we have any alternative solution for same , as we need to use the predictor object multiple times in loop and we need consistent prediction.

[epwalsh]

Then I would recommend either using a different model or using a cache on top of this model.