sentiment_onnx_classify.py

import onnxruntime as ort
import torch
from transformers import AutoTokenizer
import numpy as np

tokenizer=AutoTokenizer.from_pretrained("sentiment_classifier/")

#create onnx & onnx_int_8 sessions
session=ort.InferenceSession("sent_clf_onnx/sentiment_classifier_onnx.onnx")
session_int8=ort.InferenceSession("sent_clf_onnx/sentiment_classifier_onnx_int8.onnx")

options=ort.SessionOptions()
options.inter_op_num_threads=1
options.intra_op_num_threads=1

def classify_sentiment_onnx(texts,_model=session,_tokenizer=tokenizer):
    """
        user will pass texts separated by comma
    """
    try:
        texts=texts.split(',')
    except:
        pass

    _inputs = _tokenizer(texts, padding=True, truncation=True,
                      return_tensors="np")

    input_feed={
        "input_ids":np.array(_inputs['input_ids']),
        "attention_mask":np.array((_inputs['attention_mask']))
    }

    output = _model.run(input_feed=input_feed, output_names=['output_0'])[0]

    output=np.argmax(output,axis=1)
    output = ['Positive' if i == 1 else 'Negative' for i in output]
    return output

def classify_sentiment_onnx_quant(texts, _model=session_int8, _tokenizer=tokenizer):
    """
        user will pass texts separated by comma
    """
    try:
        texts=texts.split(',')
    except:
        pass

    _inputs = _tokenizer(texts, padding=True, truncation=True,
                      return_tensors="np")


    input_feed={
        "input_ids":np.array(_inputs['input_ids']),
        "attention_mask":np.array((_inputs['attention_mask']))
    }

    output = _model.run(input_feed=input_feed, output_names=['output_0'])[0]

    output=np.argmax(output,axis=1)
    output = ['Positive' if i == 1 else 'Negative' for i in output]

    return output