app.py

import streamlit as st
import pandas as pd
import os
import torch
import transformers
from transformers import GPT2LMHeadModel, GPT2Tokenizer, AutoTokenizer, AutoModelForCausalLM, pipeline
from huggingface_hub import HfFolder
from io import StringIO
from tqdm import tqdm
import accelerate
from accelerate import init_empty_weights, infer_auto_device_map, dispatch_model, disk_offload

# Access the Hugging Face API token from environment variables
hf_token = os.getenv('HF_API_TOKEN')

if not hf_token:
    raise ValueError("Hugging Face API token is not set. Please set the HF_API_TOKEN environment variable.")
HfFolder.save_token(hf_token)

# Set environment variable to avoid floating-point errors
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'

# Load the GPT-2 tokenizer and model
tokenizer_gpt2 = GPT2Tokenizer.from_pretrained('gpt2')
model_gpt2 = GPT2LMHeadModel.from_pretrained('gpt2')

# Create a pipeline for text generation using GPT-2
text_generator = pipeline("text-generation", model=model_gpt2, tokenizer=tokenizer_gpt2)

# Load the Llama-3 model and tokenizer once during startup
tokenizer_llama = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B", token=hf_token)
# Define the offload directory
offload_dir = "./offload"  
os.makedirs(offload_dir, exist_ok=True)  # Create the directory if it doesn't exist

# Load the Llama model with disk offloading
with init_empty_weights():
    model_llama = AutoModelForCausalLM.from_pretrained(
        "meta-llama/Meta-Llama-3-8B",
        torch_dtype='auto',
        device_map='auto',
        token=hf_token
    )

# Offload the model to the specified directory
disk_offload(model_llama, offload_dir)  # Pass the offload directory

# Define your prompt template
prompt_template = """\
You are an expert in generating synthetic data for machine learning models.
Your task is to generate a synthetic tabular dataset based on the description provided below.
Description: {description}
The dataset should include the following columns: {columns}
Please provide the data in CSV format with a minimum of 100 rows per generation.
Ensure that the data is realistic, does not contain any duplicate rows, and follows any specific conditions mentioned.
Example Description:
Generate a dataset for predicting house prices with columns: 'Size', 'Location', 'Number of Bedrooms', 'Price'
Example Output:
Size,Location,Number of Bedrooms,Price
1200,Suburban,3,250000
900,Urban,2,200000
1500,Rural,4,300000
...
Description:
{description}
Columns:
{columns}
Output: """

def preprocess_user_prompt(user_prompt):
    generated_text = text_generator(user_prompt, max_length=60, num_return_sequences=1, truncation=True)[0]["generated_text"]
    return generated_text

def format_prompt(description, columns):
    processed_description = preprocess_user_prompt(description)
    prompt = prompt_template.format(description=processed_description, columns=",".join(columns))
    return prompt

generation_params = {
    "top_p": 0.90,
    "temperature": 0.8,
    "max_new_tokens": 512,
    "return_full_text": False,
    "use_cache": False
}

def generate_synthetic_data(description, columns):
    try:
        # Prepare the input for the Llama model
        formatted_prompt = format_prompt(description, columns)
        
        # Tokenize the prompt with truncation enabled
        inputs = tokenizer_llama(formatted_prompt, return_tensors="pt", truncation=True, max_length=512).to(model_llama.device)
        
        # Generate synthetic data
        with torch.no_grad():
            outputs = model_llama.generate(
                **inputs,
                max_length=256,
                top_p=generation_params["top_p"],
                temperature=generation_params["temperature"],
                num_return_sequences=1,
                pad_token_id=tokenizer_llama.eos_token_id
            )
        
        # Decode the generated output
        generated_text = tokenizer_llama.decode(outputs[0], skip_special_tokens=True)
        
        # Return the generated synthetic data
        return generated_text
    except Exception as e:
        st.error(f"Error in generate_synthetic_data: {e}")
        return f"Error: {e}"

def generate_large_synthetic_data(description, columns, num_rows=1000, rows_per_generation=100):
    data_frames = []
    num_iterations = num_rows // rows_per_generation

    # Create a progress bar
    progress_bar = st.progress(0)
    
    for i in tqdm(range(num_iterations)):
        generated_data = generate_synthetic_data(description, columns)
        if "Error" in generated_data:
            return generated_data
        df_synthetic = process_generated_data(generated_data)
        data_frames.append(df_synthetic)
        
        # Update the progress bar
        progress_bar.progress((i + 1) / num_iterations)
    
    return pd.concat(data_frames, ignore_index=True)

def process_generated_data(csv_data):
    data = StringIO(csv_data)
    df = pd.read_csv(data)
    return df

# Streamlit app interface
st.title("Synthetic Data Generator")
description = st.text_input("Description", "e.g., Generate a dataset for predicting students' grades")
columns = st.text_input("Columns (comma-separated)", "e.g., name, age, course, grade")

if st.button("Generate"):
    description = description.strip()
    columns = [col.strip() for col in columns.split(',')]
    df_synthetic = generate_large_synthetic_data(description, columns)
    
    if isinstance(df_synthetic, str) and "Error" in df_synthetic:
        st.error(df_synthetic)  # Display error message if any
    else:
        st.success("Synthetic Data Generated!")
        st.dataframe(df_synthetic)  # Display the generated DataFrame
        st.download_button(
            label="Download CSV",
            data=df_synthetic.to_csv(index=False),
            file_name="synthetic_data.csv",
            mime="text/csv"
        )