import streamlit as st
import numpy as np
import tensorflow as tf
from tensorflow import keras
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
import random

# Define a function to generate a dataset
def generate_dataset(task_id):
    X, y = make_classification(n_samples=100, n_features=10, n_informative=5, n_redundant=3, n_repeated=2, random_state=task_id)
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=task_id)
    return X_train, X_test, y_train, y_test

# Define a neural network class
class Net(keras.Model):
    def __init__(self):
        super(Net, self).__init__()
        self.fc1 = keras.layers.Dense(20, activation='relu', input_shape=(10,))
        self.fc2 = keras.layers.Dense(10, activation='relu')
        self.fc3 = keras.layers.Dense(2)

    def call(self, x):
        x = self.fc1(x)
        x = self.fc2(x)
        x = self.fc3(x)
        return x

# Define a genetic algorithm class
class GeneticAlgorithm:
    def __init__(self, population_size):
        self.population_size = population_size
        self.population = [Net() for _ in range(population_size)]

    def selection(self, task_id):
        X_train, X_test, y_train, y_test = generate_dataset(task_id)
        fitness = []
        for net in self.population:
            net.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
            net.fit(X_train, y_train, epochs=10, verbose=0)
            loss, accuracy = net.evaluate(X_test, y_test, verbose=0)
            fitness.append(accuracy)
        self.population = [self.population[i] for i in np.argsort(fitness)[-self.population_size//2:]]

    def crossover(self):
        offspring = []
        for _ in range(self.population_size//2):
            parent1, parent2 = random.sample(self.population, 2)
            child = Net()
            child.fc1.set_weights((np.array(parent1.fc1.get_weights()) + np.array(parent2.fc1.get_weights())) / 2)
            child.fc2.set_weights((np.array(parent1.fc2.get_weights()) + np.array(parent2.fc2.get_weights())) / 2)
            child.fc3.set_weights((np.array(parent1.fc3.get_weights()) + np.array(parent2.fc3.get_weights())) / 2)
            offspring.append(child)
        self.population += offspring

    def mutation(self):
        for net in self.population:
            if random.random() < 0.1:
                weights = net.fc1.get_weights()
                weights[0] += np.random.randn(*weights[0].shape) * 0.1
                weights[1] += np.random.randn(*weights[1].shape) * 0.1
                net.fc1.set_weights(weights)
                weights = net.fc2.get_weights()
                weights[0] += np.random.randn(*weights[0].shape) * 0.1
                weights[1] += np.random.randn(*weights[1].shape) * 0.1
                net.fc2.set_weights(weights)
                weights = net.fc3.get_weights()
                weights[0] += np.random.randn(*weights[0].shape) * 0.1
                weights[1] += np.random.randn(*weights[1].shape) * 0.1
                net.fc3.set_weights(weights)

# Streamlit app
st.title("Evolution of Sub-Models")

# Parameters
st.sidebar.header("Parameters")
population_size = st.sidebar.slider("Population size", 10, 100, 50)
num_tasks = st.sidebar.slider("Number of tasks", 1, 10, 5)
num_generations = st.sidebar.slider("Number of generations", 1, 100, 10)

# Run the evolution
if st.button("Run evolution"):
    ga = GeneticAlgorithm(population_size)
    for generation in range(num_generations):
        for task_id in range(num_tasks):
            ga.selection(task_id)
            ga.crossover()
            ga.mutation()
        st.write(f"Generation {generation+1} complete")

    # Evaluate the final population
    final_accuracy = []
    for task_id in range(num_tasks):
        X_train, X_test, y_train, y_test = generate_dataset(task_id)
        accuracy = []
        for net in ga.population:
            net.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
            net.fit(X_train, y_train, epochs=10, verbose=0)
            loss, acc = net.evaluate(X_test, y_test, verbose=0)
            accuracy.append(acc)
        final_accuracy.append(np.mean(accuracy))
    st.write(f"Final accuracy: {np.mean(final_accuracy)}")