app.py

import os
import streamlit as st
import torch
import pandas as pd
import time
from tqdm import tqdm
from transformers import AutoModelForSequenceClassification, AutoTokenizer

# Set up Streamlit app
st.title("An App to Score Firm-Generated Text on Eight Risk Factors")
st.write("Note: You can either upload a CSV file or a single TXT file for scoring.")
st.write("If uploading a CSV file, ensure that it contains the following columns: cik, fyear, Item 1A (or Text). Item 1A should contain the respective risk factors section for each firm-year observation.")
st.write("If uploading a txt file, ensure it contains the respective risk factors section for each firm-year observation.")
# Hugging Face model directories
model_directories = {
    'finance': 'mgmtprofessor/finance_risk_factors',
    'accounting': 'mgmtprofessor/accounting_risk_factors',
    'technology': 'mgmtprofessor/technology_risk_factors',
    'international': 'mgmtprofessor/international_risk_factors',
    'operations': 'mgmtprofessor/operations_risk_factors',
    'marketing': 'mgmtprofessor/marketing_risk_factors',
    'management': 'mgmtprofessor/management_risk_factors',
    'legal': 'mgmtprofessor/legal_risk_factors'
}

# Check if CUDA is available
use_cuda = torch.cuda.is_available()

# Function to load a model from Hugging Face
def load_model(category):
    try:
        model_name = model_directories.get(category)
        if model_name:
            tokenizer = AutoTokenizer.from_pretrained(model_name)
            model = AutoModelForSequenceClassification.from_pretrained(model_name)
            return model, tokenizer
        else:
            st.error(f"No Hugging Face model found for {category}")
            return None, None
    except Exception as e:
        st.error(f"Failed to load model for {category}: {e}")
        return None, None

# Function to score a document and return the probability for class '1'
def score_document(model, tokenizer, text_data):
    if isinstance(text_data, str):
        text_data = [text_data]

    # Tokenize the input
    inputs = tokenizer(text_data, return_tensors="pt", padding=True, truncation=True)
    
    # Perform the prediction
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Get probabilities (softmax)
    probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
    
    # Get the probability associated with class '1'
    probability_class_1 = probabilities[:, 1].tolist()  # Return as list

    return probability_class_1

# Function to find the relevant text column
def get_text_column(df):
    possible_columns = ['Item 1A', 'Item 1A.', 'Item 1A. Risk Factors', 'text', 'Text']
    for col in possible_columns:
        if col in df.columns:
            return col
    return None  # Return None if no matching column is found

# Dropdown to select file type
file_type = st.selectbox("Select the file type to upload:", ["CSV", "TXT"])

# Track the start time
start_time = time.time()

# Handle CSV or TXT upload
if file_type == "CSV":
    csv_file = st.file_uploader("Upload a CSV file with text data", type=["csv"])
    
    if csv_file is not None:
        # Read the CSV file
        df = pd.read_csv(csv_file)

        # Find the relevant text column
        text_column = get_text_column(df)
        
        if text_column is None:
            st.error("No valid text column found. Please ensure your CSV contains 'Item 1A', 'Item 1A.', 'Item 1A. Risk Factors', 'Text', or 'text'.")
        else:
            # Extract text data from the identified column
            text_data = df[text_column].dropna().tolist()  # Extracts all non-empty rows

            # Initialize an empty DataFrame for results
            result_df = df.copy()
            
            # Progress bar
            progress_bar = st.progress(0)
            total_categories = len(model_directories)
            
            for i, category in enumerate(tqdm(model_directories.keys(), desc="Scoring documents")):
                # Load the pre-trained model for the current category
                model, tokenizer = load_model(category)
                
                # Skip the category if model loading fails
                if model is not None:
                    # Score the document for each row in the text data
                    category_scores = []
                    for text in text_data:
                        probability = score_document(model, tokenizer, text)
                        category_scores.append(probability[0])  # Extract the first (and only) value
                    
                    # Add the results to the DataFrame
                    result_df[category.capitalize()] = category_scores
                
                # Update the progress bar
                progress_bar.progress((i + 1) / total_categories)
                
                # Estimate remaining time
                elapsed_time = time.time() - start_time
                estimated_total_time = (elapsed_time / (i + 1)) * total_categories
                st.write(f"Elapsed time: {elapsed_time:.2f}s, Estimated time remaining: {estimated_total_time - elapsed_time:.2f}s")
            
            # Save results to CSV
            csv = result_df.to_csv(index=False).encode('utf-8')
            st.download_button(
                label="Download results as CSV",
                data=csv,
                file_name="document_scoring_results.csv",
                mime="text/csv",
            )

            # Display completion message
            st.success("Document scoring complete!")

elif file_type == "TXT":
    doc_file = st.file_uploader("Upload a TXT file", type=["txt"])
    
    if doc_file is not None:
        # Read the content of the uploaded .txt file
        text_data = doc_file.read().decode("utf-8")
        
        # Initialize an empty DataFrame for results
        result_df = pd.DataFrame(columns=["Category", "Probability"])
        
        # Progress bar
        progress_bar = st.progress(0)
        total_categories = len(model_directories)
        
        for i, category in enumerate(tqdm(model_directories.keys(), desc="Scoring documents")):
            # Load the pre-trained model for the current category
            model, tokenizer = load_model(category)
            
            # Skip the category if model loading fails
            if model is not None:
                # Score the document
                probability = score_document(model, tokenizer, text_data)
                
                # Create a DataFrame for the current result
                new_row = pd.DataFrame({
                    "Category": [category],
                    "Probability": [probability[0]]  # Extract the first value
                })
                
                # Use pd.concat to append the new row to the DataFrame
                result_df = pd.concat([result_df, new_row], ignore_index=True)
            
            # Update the progress bar
            progress_bar.progress((i + 1) / total_categories)
            
            # Estimate remaining time
            elapsed_time = time.time() - start_time
            estimated_total_time = (elapsed_time / (i + 1)) * total_categories
            st.write(f"Elapsed time: {elapsed_time:.2f}s, Estimated time remaining: {estimated_total_time - elapsed_time:.2f}s")

        # Save results to CSV
        csv = result_df.to_csv(index=False).encode('utf-8')
        st.download_button(
            label="Download results as CSV",
            data=csv,
            file_name="document_scoring_results.csv",
            mime="text/csv",
        )

        # Display completion message
        st.success("Document scoring complete!")

st.write("Note: Ensure the uploaded document is formatted correctly. The models are limited to 512 tokens and will be upgraded in a future version.")