app.py

import re
import requests
import gradio as gr
import pandas as pd
from transformers import pipeline
from transformers import AutoTokenizer
from transformers import AutoModelForSequenceClassification

def process_tweet(tweet):
    # remove links
    tweet = re.sub('((www\.[\s]+)|(https?://[^\s]+))', '', tweet)
    # remove usernames
    tweet = re.sub('@[^\s]+', '', tweet)
    # remove additional white spaces
    tweet = re.sub('[\s]+', ' ', tweet)
    # replace hashtags with words
    tweet = re.sub(r'#([^\s]+)', r'\1', tweet)
    # trim
    tweet = tweet.strip('\'"')
    return tweet

tokenizer = AutoTokenizer.from_pretrained(
    "azamat/geocoder_coordinates_model"
)

relevancy_pipeline = pipeline("sentiment-analysis", model="azamat/geocoder_relevancy_model")

coordinates_model = AutoModelForSequenceClassification.from_pretrained(
    "azamat/geocoder_coordinates_model",
)

def predict_relevancy(text):
    outputs = relevancy_pipeline(text)
    return outputs[0]['label'], outputs[0]['score']

def predict_coordinates(text):
    encoding = tokenizer(text, padding="max_length", truncation=True, \
        max_length=128, return_tensors='pt')
    outputs = coordinates_model(**encoding)
    return round(outputs[0][0][0].item(), 3), round(outputs[0][0][1].item(), 3)

def reverse_geocode(lat, lon):
    payload = {
        'lat'             : lat, 
        'lon'             : lon, 
        'zoom'            : 12, 
        'format'          : 'jsonv2',
        'accept-language' : 'en'
    }
    try:
        r = requests.get('https://geocode.maps.co/reverse', params=payload)
        return r.json()['display_name']
    except:
        return "No data"

def predict(text):
    text = process_tweet(text)
    data = {
        "relevancy_score"  : 0,
        "lat"              : 0,
        "lon"              : 0,
        "reversed lat/lon" : ""
    }
    relevancy_label, relevancy_score = predict_relevancy(text)
    if relevancy_label == 'relevant':
        data['relevancy_score'] = round(relevancy_score * 100, 2)
        
        lat, lon = predict_coordinates(text)
        data['lat'] = lat
        data['lon'] = lon
        
        reverse_geocoded = reverse_geocode(lat, lon)
        data['reversed lat/lon'] = reverse_geocoded
        
    return pd.DataFrame([data])

with gr.Blocks() as demo:

    gr.Markdown("# **<p align='center'>Twitter geocoding with 🤗 Transformers</p>**")
    gr.Markdown("### <div align='left'>Pipeline consists of:</div>")
    gr.Markdown("### <div align='left'>1) Relevancy scoring model - predicts whether a tweet has geocoding related information</div>")
    gr.Markdown("### <div align='left'>2) Coordinate predicting model - predicts exact latitude and longitude of user by tweet</div>")
    gr.Markdown("### <div align='left'>3) Nominatim API for reverse geocoding lat/lon - uses open street map to reverse geocode lat and lon</div>")

    inputs = gr.Textbox(placeholder="Enter the tweet")
    outputs = [gr.Dataframe(label="Geocoded data")]
    inputs.submit(predict, inputs=inputs, outputs=outputs)

if __name__ == "__main__":
    demo.launch()