app.py

import streamlit as st
import pandas as pd
from streamlit import cli as stcli
from transformers import pipeline
from sentence_transformers import SentenceTransformer, util
import sys

HISTORY_WEIGHT = 100 # set history weight (if found any keyword from history, it will priorities based on its weight)

@st.cache(allow_output_mutation=True)
def get_model(model):
	return pipeline("fill-mask", model=model, top_k=20)#set the maximum of tokens to be retrieved after each inference to model

def main(nlp, semantic_model):
    data_load_state = st.text('Inference to model...')
    result = nlp(text+' '+nlp.tokenizer.mask_token)
    data_load_state.text('')
    sem_list=[semantic_text.strip()]
    if len(semantic_text):
        predicted_seq=[rec['sequence'] for rec in result]
        predicted_embeddings = semantic_model.encode(predicted_seq, convert_to_tensor=True)
        semantic_history_embeddings = semantic_model.encode(sem_list, convert_to_tensor=True)
        cosine_scores = util.cos_sim(predicted_embeddings, semantic_history_embeddings)
    
    for index, r in enumerate(result):
        if len(semantic_text):
                if len(r['token_str'])>2: #skip spcial chars such as "?"
                    result[index]['score']+=float(sum(cosine_scores[index]))*HISTORY_WEIGHT
        if r['token_str'].lower().strip() in history_keyword_text.lower().strip() and len(r['token_str'].lower().strip())>1:
            #found from history, then increase the score of tokens
            result[index]['score']*=HISTORY_WEIGHT
            
    #sort the results        
    df=pd.DataFrame(result).sort_values(by='score', ascending=False)
    
#    show the results as a table
    st.table(df)
#    print(df)
    
    
if __name__ == '__main__':
    if st._is_running_with_streamlit:
        st.markdown("""
# Introduction
This is an example of an auto-complete approach where the next token suggested based on users's history Keyword match & Semantic similarity of users's history (log).
The next token is predicted per probability and a weight if it is appeared in keyword user's history or there is a similarity to semantic user's history

""")
        history_keyword_text = st.text_input("Enter users's history <keywords match> (optional, i.e., 'Gates')", value="")
        
        text = st.text_input("Enter a text for auto completion...", value='Where is Bill')
        
        semantic_text = st.text_input("Enter users's history <semantic> (optional, i.e., 'Microsoft or President')", value="Microsoft")
        
        model = st.selectbox("Choose a model", ["roberta-base", "bert-base-uncased"])
        
        data_load_state = st.text('Loading model...')

        semantic_model = SentenceTransformer('all-MiniLM-L6-v2')
        nlp = get_model(model)
        
        main(nlp, semantic_model)
    else:
        sys.argv = ['streamlit', 'run', sys.argv[0]]
        sys.exit(stcli.main())