app.py

import streamlit as st


# Library for Entailment
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model

tokenizer = AutoTokenizer.from_pretrained("roberta-large-mnli")

text_classification_model = AutoModelForSequenceClassification.from_pretrained("roberta-large-mnli")



### Streamlit interface ###
      
st.title("Text Classification")

st.subheader("Entailment, neutral or contradiction?")

with st.form("submission_form", clear_on_submit=False):

       threshold = st.slider("Threshold", min_value=0.0, max_value=1.0, step=0.1, value=0.7)

       sentence_1 = st.text_input("Sentence 1 input")
       
       sentence_2 = st.text_input("Sentence 2 input")
       
       submit_button_compare = st.form_submit_button("Compare Sentences")
       
# If submit_button_compare clicked
if submit_button_compare:

       print("Comparing sentences...")

       ### Text classification - entailment, neutral or contradiction ###

       raw_inputs = [f"{sentence_1}</s></s>{sentence_2}"]

       inputs = tokenizer(raw_inputs, padding=True, truncation=True, return_tensors="pt")

       # print(inputs)

       outputs = text_classification_model(**inputs)

       outputs = torch.nn.functional.softmax(outputs.logits, dim = -1)
       # print(outputs)

       # argmax_index = torch.argmax(outputs).item()

       print(text_classification_model.config.id2label[0], ":", round(outputs[0][0].item()*100,2),"%")
       print(text_classification_model.config.id2label[1], ":", round(outputs[0][1].item()*100,2),"%")
       print(text_classification_model.config.id2label[2], ":", round(outputs[0][2].item()*100,2),"%")

       st.subheader("Text classification for both sentences:")

       st.write(text_classification_model.config.id2label[1], ":", round(outputs[0][1].item()*100,2),"%")
       st.write(text_classification_model.config.id2label[0], ":", round(outputs[0][0].item()*100,2),"%")
       st.write(text_classification_model.config.id2label[2], ":", round(outputs[0][2].item()*100,2),"%")

       entailment_score = round(outputs[0][2].item(),2)

       if entailment_score >= threshold:
              st.subheader("The statements are very similar!")
       else:
              st.subheader("The statements are not close enough")