AllenNLP: Run on Production with Django

This post describes howto run a custom AllenNLP model within Django and Django Rest Framework.

The View code:

import json

from rest_framework import status, viewsets
from rest_framework.parsers import JSONParser
from rest_framework.response import Response

from .utils import predict

class PredictViewSet(viewsets.ViewSet):
    parser_classes = (JSONParser,)

    def create(self, request):
        example call:
        curl -X POST "http://localhost:8000/" -H "content-type: application/json" --data '{"source": "1/5/2003"}'
        if isinstance(, str):
            data = json.loads(
            data =

        if not "source" in data:
            return Response(
                'Format: {"source": string}', status=status.HTTP_400_BAD_REQUEST
        return Response(predict(data.get("source")))

And the with the predict method.

from allennlp.common.util import import_submodules
from allennlp.models.archival import load_archive
from allennlp.predictors import Predictor

def predict(source):
    # register custom model code
    # the library code is in the same Django app (main) in the folder "library"

    # load model without using cuda
    archive = load_archive("../models/model.tar.gz", cuda_device=-1)
    # select predictor needed for this model
    predictor = Predictor.from_archive(archive, "seq2seq")

    result = predictor.predict_json({"source": source})
    # add a confidence bool to help ignoring bad results
    confident = True if abs(result["class_log_probabilities"][0]) < 0.05 else False

    return {
        "predicted": "".join(result["predicted_tokens"]),
        "probability": result["class_log_probabilities"][0],
        "confident": confident,