app.py

import gradio as gr
import pandas as pd
from transformers import TapexTokenizer, BartForConditionalGeneration, pipeline

# Initialize TAPEX (Microsoft) model and tokenizer
tokenizer_tapex = TapexTokenizer.from_pretrained("microsoft/tapex-large-finetuned-wtq")
model_tapex = BartForConditionalGeneration.from_pretrained("microsoft/tapex-large-finetuned-wtq")

# Initialize TAPAS (Google) models and pipelines
pipe_tapas = pipeline(task="table-question-answering", model="google/tapas-large-finetuned-wtq")
pipe_tapas2 = pipeline(task="table-question-answering", model="google/tapas-large-finetuned-wikisql-supervised")

def chunk_dataframe(df, max_tokens=1024):
    """
    Chunk a large dataframe into smaller pieces that fit within the token limit.
    For simplicity, we're assuming the number of rows determines the token count. 
    """
    chunk_size = max_tokens // len(df.columns)  # Approximate number of rows that fit
    return [df[i:i+chunk_size] for i in range(0, len(df), chunk_size)]

def process_table_query(query, table_data):
    """
    Process a query and CSV data using TAPEX.
    """
    # Convert all columns in the table to strings for TAPEX compatibility
    table_data = table_data.astype(str)

    # Chunk the table if it's too large
    chunks = chunk_dataframe(table_data)

    results = []
    for chunk in chunks:
        # Microsoft TAPEX model (using TAPEX tokenizer and model)
        encoding = tokenizer_tapex(table=chunk, query=query, return_tensors="pt", max_length=1024, truncation=True)
        outputs = model_tapex.generate(**encoding)
        result_tapex = tokenizer_tapex.batch_decode(outputs, skip_special_tokens=True)[0]
        results.append(result_tapex)

    # Aggregate results (example: summing numerical values if the query is sum-related)
    try:
        # Convert the list of results to floats and sum them
        numerical_results = [float(r) for r in results if r.strip().isdigit()]
        total_sum = sum(numerical_results)
        return str(total_sum)
    except ValueError:
        # If the results are not numerical, return the joined string
        return ' '.join(results)


def answer_query_from_csv(query, file):
    """
    Function to handle file input and return model results.
    """
    # Read the file into a DataFrame
    table_data = pd.read_csv(file)

    # Convert object-type columns to lowercase, ensuring only valid strings are affected
    for column in table_data.columns:
        if table_data[column].dtype == 'object':
            table_data[column] = table_data[column].apply(lambda x: x.lower() if isinstance(x, str) else x)

    
    
    # Extract year, month, day, and time components for datetime columns
    for column in table_data.columns:
        if pd.api.types.is_datetime64_any_dtype(table_data[column]):
            table_data[f'{column}_year'] = table_data[column].dt.year
            table_data[f'{column}_month'] = table_data[column].dt.month
            table_data[f'{column}_day'] = table_data[column].dt.day
            table_data[f'{column}_time'] = table_data[column].dt.strftime('%H:%M:%S')
     # Ensure all data in the table is converted to string
    table_data = table_data.astype(str)
    # Process the CSV file and query using TAPEX
    result_tapex = process_table_query(query, table_data)

    # Process the query using TAPAS pipelines (ensure all cells are strings)
    result_tapas = pipe_tapas(table=table_data, query=query)['cells'][0]
    result_tapas2 = pipe_tapas2(table=table_data, query=query)['cells'][0]
    
    return result_tapex, result_tapas, result_tapas2



# Create Gradio interface
with gr.Blocks() as interface:
    gr.Markdown("# Table Question Answering with TAPEX and TAPAS Models")
    
    # Add a notice about the token limit
    gr.Markdown("### Note: Only the first 1024 tokens (query + table data) will be considered per chunk. If your table is too large, it will be chunked and processed separately.")
    
    # Two-column layout (input on the left, output on the right)
    with gr.Row():
        with gr.Column():
            # Input fields for the query and file
            query_input = gr.Textbox(label="Enter your query:")
            csv_input = gr.File(label="Upload your CSV file")

        with gr.Column():
            # Output textboxes for the answers
            result_tapex = gr.Textbox(label="TAPEX Answer")
            result_tapas = gr.Textbox(label="TAPAS (WikiTableQuestions) Answer")
            result_tapas2 = gr.Textbox(label="TAPAS (WikiSQL) Answer")
    
    # Submit button
    submit_btn = gr.Button("Submit")
    
    # Action when submit button is clicked
    submit_btn.click(
        fn=answer_query_from_csv, 
        inputs=[query_input, csv_input], 
        outputs=[result_tapex, result_tapas, result_tapas2]
    )

# Launch the Gradio interface
interface.launch(share=True)