app.py

import streamlit as st
import numpy as np
import torch
import random
from transformers import (
    GPT2LMHeadModel, GPT2Tokenizer, Trainer, TrainingArguments, DataCollatorForLanguageModeling
)
from datasets import Dataset
from huggingface_hub import HfApi
import plotly.graph_objects as go
import time
from datetime import datetime
from typing import Dict, List, Any
import pandas as pd  # Added pandas import

# Cyberpunk and Loading Animation Styling
def setup_cyberpunk_style():
    st.markdown("""
        <style>
        /* [Your existing CSS styles here] */
        </style>
    """, unsafe_allow_html=True)

# Prepare Dataset Function with Padding Token Fix
def prepare_dataset(data, tokenizer, block_size=128):
    tokenizer.pad_token = tokenizer.eos_token
    def tokenize_function(examples):
        return tokenizer(examples['text'], truncation=True, max_length=block_size, padding='max_length')

    raw_dataset = Dataset.from_dict({'text': data})
    tokenized_dataset = raw_dataset.map(tokenize_function, batched=True, remove_columns=['text'])
    tokenized_dataset = tokenized_dataset.map(lambda examples: {'labels': examples['input_ids']}, batched=True)
    tokenized_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'labels'])
    return tokenized_dataset

# Define Model Initialization
def initialize_model(model_name="gpt2"):
    model = GPT2LMHeadModel.from_pretrained(model_name)
    tokenizer = GPT2Tokenizer.from_pretrained(model_name)
    tokenizer.pad_token = tokenizer.eos_token
    return model, tokenizer

# Load Dataset Function with Uploaded File Option
def load_dataset(data_source="demo", tokenizer=None, uploaded_file=None):
    if data_source == "demo":
        data = [
            "In the neon-lit streets of Neo-Tokyo, a lone hacker fights against the oppressive megacorporations.",
            "The rain falls in sheets, washing away the bloodstains from the alleyways.",
            "She plugs into the matrix, seeking answers to questions that have haunted her for years."
        ]
    elif data_source == "uploaded file" and uploaded_file is not None:
        if uploaded_file.name.endswith(".txt"):
            data = [uploaded_file.read().decode("utf-8")]
        elif uploaded_file.name.endswith(".csv"):
            df = pd.read_csv(uploaded_file)
            data = df[df.columns[0]].astype(str).tolist()  # Ensure all data is string
        else:
            data = ["Unsupported file format."]
    else:
        data = ["No file uploaded. Please upload a dataset."]
    
    dataset = prepare_dataset(data, tokenizer)
    return dataset

# Train Model Function
def train_model(model, train_dataset, tokenizer, epochs=3, batch_size=4, use_ga=False, ga_params=None):
    if not use_ga:
        training_args = TrainingArguments(
            output_dir="./results",
            overwrite_output_dir=True,
            num_train_epochs=epochs,
            per_device_train_batch_size=batch_size,
            save_steps=10_000,
            save_total_limit=2,
            logging_dir="./logs",
            logging_steps=1,
            logging_strategy='steps',
            report_to=None,  # Disable default logging to WandB or other services
        )
        
        data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
        
        trainer = Trainer(
            model=model,
            args=training_args,
            data_collator=data_collator,
            train_dataset=train_dataset,
        )
        trainer.train()
        return trainer.state.log_history
    else:
        # GA training logic
        param_bounds = {
            'learning_rate': (1e-5, 5e-5),
            'epochs': (1, ga_params['max_epochs']),
            'batch_size': [2, 4, 8, 16]
        }

        population = create_ga_population(ga_params['population_size'], param_bounds)
        best_individual = None
        best_fitness = float('inf')
        all_losses = []

        for generation in range(ga_params['num_generations']):
            fitnesses = []
            for idx, individual in enumerate(population):
                model_copy = GPT2LMHeadModel.from_pretrained('gpt2')
                training_args = TrainingArguments(
                    output_dir=f"./results/ga_{generation}_{idx}",
                    num_train_epochs=individual['epochs'],
                    per_device_train_batch_size=individual['batch_size'],
                    learning_rate=individual['learning_rate'],
                    logging_steps=1,
                    logging_strategy='steps',
                    report_to=None,  # Disable default logging to WandB or other services
                )

                trainer = Trainer(
                    model=model_copy,
                    args=training_args,
                    train_dataset=train_dataset,
                )

                # Capture the training result
                train_result = trainer.train()
                
                # Safely retrieve the training loss
                fitness = train_result.metrics.get('train_loss', None)
                if fitness is None:
                    # If 'train_loss' is not available, try to compute it from log history
                    if 'loss' in trainer.state.log_history[-1]:
                        fitness = trainer.state.log_history[-1]['loss']
                    else:
                        fitness = float('inf')  # Assign a large number if loss is not available

                fitnesses.append(fitness)
                all_losses.extend(trainer.state.log_history)

                if fitness < best_fitness:
                    best_fitness = fitness
                    best_individual = individual
                    model.load_state_dict(model_copy.state_dict())

                del model_copy
                torch.cuda.empty_cache()

            # GA operations
            parents = select_ga_parents(population, fitnesses, ga_params['num_parents'])
            offspring_size = ga_params['population_size'] - ga_params['num_parents']
            offspring = ga_crossover(parents, offspring_size)
            offspring = ga_mutation(offspring, param_bounds, ga_params['mutation_rate'])
            population = parents + offspring

        return all_losses

# GA-related functions
def create_ga_population(size: int, param_bounds: Dict[str, Any]) -> List[Dict[str, Any]]:
    """Create initial population for genetic algorithm"""
    population = []
    for _ in range(size):
        individual = {
            'learning_rate': random.uniform(*param_bounds['learning_rate']),
            'epochs': random.randint(*param_bounds['epochs']),
            'batch_size': random.choice(param_bounds['batch_size']),
        }
        population.append(individual)
    return population

def select_ga_parents(population: List[Dict[str, Any]], fitnesses: List[float], num_parents: int) -> List[Dict[str, Any]]:
    """Select best performing individuals as parents"""
    parents = [population[i] for i in np.argsort(fitnesses)[:num_parents]]
    return parents

def ga_crossover(parents: List[Dict[str, Any]], offspring_size: int) -> List[Dict[str, Any]]:
    """Create offspring through crossover of parents"""
    offspring = []
    for _ in range(offspring_size):
        parent1 = random.choice(parents)
        parent2 = random.choice(parents)
        child = {
            'learning_rate': random.choice([parent1['learning_rate'], parent2['learning_rate']]),
            'epochs': random.choice([parent1['epochs'], parent2['epochs']]),
            'batch_size': random.choice([parent1['batch_size'], parent2['batch_size']]),
        }
        offspring.append(child)
    return offspring

def ga_mutation(offspring: List[Dict[str, Any]], param_bounds: Dict[str, Any], mutation_rate: float = 0.1) -> List[Dict[str, Any]]:
    """Apply random mutations to offspring"""
    for individual in offspring:
        if random.random() < mutation_rate:
            individual['learning_rate'] = random.uniform(*param_bounds['learning_rate'])
        if random.random() < mutation_rate:
            individual['epochs'] = random.randint(*param_bounds['epochs'])
        if random.random() < mutation_rate:
            individual['batch_size'] = random.choice(param_bounds['batch_size'])
    return offspring

# Main App Logic
def main():
    setup_cyberpunk_style()
    st.markdown('<h1 class="main-title">Neural Training Hub</h1>', unsafe_allow_html=True)
    
    # Sidebar Configuration with Additional Options
    with st.sidebar:
        st.markdown("### Configuration Panel")
    
        # Hugging Face API Token Input
        hf_token = st.text_input("Enter your Hugging Face Token", type="password")
        if hf_token:
            api = HfApi()
            api.set_access_token(hf_token)
            st.success("Hugging Face token added successfully!")
    
        # Training Parameters
        training_epochs = st.slider("Training Epochs", min_value=1, max_value=5, value=3)
        batch_size = st.slider("Batch Size", min_value=2, max_value=8, value=4)
        model_choice = st.selectbox("Model Selection", ("gpt2", "distilgpt2", "gpt2-medium"))
        
        # Dataset Source Selection
        data_source = st.selectbox("Data Source", ("demo", "uploaded file"))
        uploaded_file = st.file_uploader("Upload a text file", type=["txt", "csv"]) if data_source == "uploaded file" else None
        
        custom_learning_rate = st.number_input("Learning Rate", min_value=1e-6, max_value=5e-4, value=3e-5, step=1e-6, format="%.6f")
    
        # Advanced Settings Toggle
        advanced_toggle = st.checkbox("Advanced Training Settings")
        if advanced_toggle:
            warmup_steps = st.slider("Warmup Steps", min_value=0, max_value=500, value=100)
            weight_decay = st.slider("Weight Decay", min_value=0.0, max_value=0.1, step=0.01, value=0.01)
        else:
            warmup_steps = 100
            weight_decay = 0.01
    
        # Add training method selection
        training_method = st.selectbox("Training Method", ("Standard", "Genetic Algorithm"))
        
        if training_method == "Genetic Algorithm":
            st.markdown("### GA Parameters")
            ga_params = {
                'population_size': st.slider("Population Size", min_value=4, max_value=10, value=6),
                'num_generations': st.slider("Number of Generations", min_value=1, max_value=5, value=3),
                'num_parents': st.slider("Number of Parents", min_value=2, max_value=4, value=2),
                'mutation_rate': st.slider("Mutation Rate", min_value=0.0, max_value=1.0, value=0.1),
                'max_epochs': training_epochs
            }
        else:
            ga_params = None

    # Initialize model and tokenizer
    if 'model' not in st.session_state:
        model, tokenizer = initialize_model(model_name=model_choice)
        st.session_state['model'] = model
        st.session_state['tokenizer'] = tokenizer
        st.session_state['model_name'] = model_choice
    else:
        if st.session_state.get('model_name') != model_choice:
            model, tokenizer = initialize_model(model_name=model_choice)
            st.session_state['model'] = model
            st.session_state['tokenizer'] = tokenizer
            st.session_state['model_name'] = model_choice
        else:
            model = st.session_state['model']
            tokenizer = st.session_state['tokenizer']

    # Load Dataset
    train_dataset = load_dataset(data_source, tokenizer, uploaded_file=uploaded_file)

    # Go Button to Start Training
    if st.button("Go"):
        st.markdown("### Model Training Progress")
        progress_bar = st.progress(0)
        status_text = st.empty()
        status_text.text("Training in progress...")

        # Train the model
        if training_method == "Standard":
            logs = train_model(model, train_dataset, tokenizer, training_epochs, batch_size)
        else:
            logs = train_model(model, train_dataset, tokenizer, training_epochs, batch_size, use_ga=True, ga_params=ga_params)

        # Update progress bar to 100%
        progress_bar.progress(100)
        status_text.text("Training complete!")

        # Store the model and logs in st.session_state
        st.session_state['model'] = model
        st.session_state['logs'] = logs

    # Plot the losses if available
    if 'logs' in st.session_state:
        logs = st.session_state['logs']
        losses = [log['loss'] for log in logs if 'loss' in log]
        steps = list(range(len(losses)))
        if losses:
            # Plot the losses
            fig = go.Figure()
            fig.add_trace(go.Scatter(x=steps, y=losses, mode='lines+markers', name='Training Loss', line=dict(color='#00ff9d')))
            fig.update_layout(
                title="Training Progress",
                xaxis_title="Training Steps",
                yaxis_title="Loss",
                template="plotly_dark",
                plot_bgcolor='rgba(0,0,0,0)',
                paper_bgcolor='rgba(0,0,0,0)',
                font=dict(color='#00ff9d')
            )
            st.plotly_chart(fig, use_container_width=True)
        else:
            st.write("No loss data available to plot.")
    else:
        st.write("Train the model to see the loss plot.")

    # After training, you can use the model for inference
    st.markdown("### Model Inference")
    with st.form("inference_form"):
        user_input = st.text_input("Enter prompt for the model:")
        submitted = st.form_submit_button("Generate")
        if submitted:
            if 'model' in st.session_state:
                model = st.session_state['model']
                tokenizer = st.session_state['tokenizer']
                inputs = tokenizer(user_input, return_tensors="pt")
                outputs = model.generate(inputs['input_ids'], max_length=100, num_return_sequences=1)
                response = tokenizer.decode(outputs[0], skip_special_tokens=True)
                st.write("Model output:", response)
            else:
                st.write("Please train the model first.")

if __name__ == "__main__":
    main()