app.py

import gradio as gr
import os
import torch
import torchtext
from torchtext.data.utils import get_tokenizer

from model import create_model

tokenizer = get_tokenizer("basic_english")

model = create_model()

model.load_state_dict(torch.load(f="model.pt", map_location=torch.device("cpu")))

vocab = torch.load("vocab.pt", map_location=torch.device("cpu"))
text_pipeline = lambda x: vocab(tokenizer(x))


def predict(text):
    with torch.no_grad():
        text = torch.tensor(text_pipeline(text), dtype=torch.int64)
        text = torch.unsqueeze(text, 0)
        result = model(text).squeeze()
        result = (round(result.item(), 2)) * 10000
        return result


title = "Salary Predictor"
description = "This is a test project to see if I could built a Machine Learning model to predict salary offered based on a posted job description. To test, copy a whole job description from a linkedIn post. Results are in USD."
article = """ This project was built by Marie Pelletier <br/>
            <a href="https://www.linkedin.com/in/marie-pelletier-14837441/">LinkedIn</a><br/>
            <a href="https://github.com/Marie000">Github</a>
        </p>
        <p>This is a work in progress and is not meant to be used as an accurate predictor of 
        salary. It is limited by the <a
            href="https://www.kaggle.com/datasets/arshkon/linkedin-job-postings/data">
            data that was
            used</a>, 
            which is US-based, from 2023. It also does not take into account regions or 
            fluctuations in the market over time. The dataset included over 33,000 job postings,
              but only 13,000 had salary information.</p>
        <p>For all you statistic nerds, the r-squared score of the model was around 0.6. 
        For the rest of you, that level of accuracy can probably be described as "better than 
        nothing". It might have been better if I had more data to work with. Or maybe there is 
        just so much we can infer from a description alone.</p>

        <p>For this first test, only the description of the job is taken into account. 
        Adding other information, including the date of the posting and the location, 
        could improve the prediction.</p>
        <p>The google colab notebook used to generate this model can be found 
        <a href="https://github.com/Marie000/Linkedin-predictor-model">here</a>"""

demo = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(lines=20, placeholder="copy whole job description here"),
    outputs="number",
    title=title,
    description=description,
    article=article,
)

demo.launch(debug=False, share=True, show_error=True)