app.py

# Explore 2019-2020 NHTSA Traffic Fatality data
# A. Harper / DATA 495 / February 2026

# VERY IMPORTANT
# Load into Hugging Face Space (using the Gradio Framework)
# Include requirements.txt file (list: torch==2.1.2, transformers==4.37.2, gradio==4.19.2, pandas)
# Also open the README file and add a new line: python_version: 3.11
# App does not run properly on default Python version.
# END

# To run, navigate to the App tab.  
# The app will prompt the user to upload a csv file, 
# then ask a question about the data. Click 'Submit' to run.

# Import tools and libraries
import multiprocessing
multiprocessing.set_start_method("spawn", force=True)
import torch
import pandas as pd
import gradio as gr
from transformers import AutoTokenizer, AutoModelForTableQuestionAnswering

# Load TAPAS model
model_name = "Meena/table-question-answering-tapas"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTableQuestionAnswering.from_pretrained(model_name).to("cpu")

# Global dataframe storage
df = None

def dataset_change(file):
    global df
    df = pd.read_csv(file.name)
    # Limit size for TAPAS - changed from 20 to 100 to show all rows.
    # Recommended limit for TAPAS is 200; 100 is well within the limit.
    df = df.head(100)
    # Convert everything to string - required for TAPAS
    df = df.astype(str)
    return df.head(100)

def answer_question(question):
    global df
    
    if df is None:
        return "Please upload a dataset first."

    inputs = tokenizer(
        table=df,
        queries=[question],
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    )

    outputs = model(**inputs)

    predicted_answer_coordinates, predicted_aggregation_indices = (
        tokenizer.convert_logits_to_predictions(
            inputs,
            outputs.logits.detach(),
            outputs.logits_aggregation.detach()
        )
    )

    answers = []
    for coordinates in predicted_answer_coordinates:
        if len(coordinates) == 0:
            answers.append("No answer found.")
        else:
            cell_values = [df.iat[row, col] for row, col in coordinates]
            answers.append(", ".join(map(str, cell_values)))

    return answers[0]

# Gradio UI
with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            file_input = gr.File(label="Upload CSV File")
            question_input = gr.Textbox(label="Input Question")
            submit_btn = gr.Button("Submit")
        with gr.Column():
            answer_output = gr.Textbox(label="Answer")

    with gr.Row():
        dataframe_preview = gr.Dataframe()

    file_input.change(fn=dataset_change, inputs=file_input, outputs=dataframe_preview)
    submit_btn.click(fn=answer_question, inputs=question_input, outputs=answer_output)

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