interface.py

# interface.py

# Importar 'spaces' y decoradores antes que cualquier biblioteca que pueda inicializar CUDA
from decorators import gpu_decorator

# Luego importar cualquier cosa relacionada con PyTorch o el modelo que va a usar la GPU
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
import io
from sympy import symbols, lambdify, sympify

# Importar otras partes necesarias del código (config, etc.)
from config import DEVICE, MODEL_PATH, MAX_LENGTH, TEMPERATURE

# Cargar el modelo fuera de la función para evitar la inicialización innecesaria cada vez que se llame a la función
model_path = MODEL_PATH
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

###############################


# bioprocess_model.py

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from scipy.integrate import odeint
from scipy.optimize import curve_fit
from sklearn.metrics import mean_squared_error
import seaborn as sns

class BioprocessModel:
    def __init__(self):
        self.params = {}
        self.r2 = {}
        self.rmse = {}
        self.datax = []
        self.datas = []
        self.datap = []
        self.dataxp = []
        self.datasp = []
        self.datapp = []
        self.datax_std = []
        self.datas_std = []
        self.datap_std = []

    @staticmethod
    def logistic(time, xo, xm, um):
        return (xo * np.exp(um * time)) / (1 - (xo / xm) * (1 - np.exp(um * time)))

    @staticmethod
    def substrate(time, so, p, q, xo, xm, um):
        return so - (p * xo * ((np.exp(um * time)) / (1 - (xo / xm) * (1 - np.exp(um * time))) - 1)) - \
               (q * (xm / um) * np.log(1 - (xo / xm) * (1 - np.exp(um * time))))

    @staticmethod
    def product(time, po, alpha, beta, xo, xm, um):
        return po + (alpha * xo * ((np.exp(um * time) / (1 - (xo / xm) * (1 - np.exp(um * time)))) - 1)) + \
               (beta * (xm / um) * np.log(1 - (xo / xm) * (1 - np.exp(um * time))))

    @staticmethod
    def logistic_diff(X, t, params):
        xo, xm, um = params
        dXdt = um * X * (1 - X / xm)
        return dXdt

    def substrate_diff(self, S, t, params, biomass_params, X_func):
        so, p, q = params
        xo, xm, um = biomass_params
        X_t = X_func(t)
        dSdt = -p * (um * X_t * (1 - X_t / xm)) - q * X_t
        return dSdt

    def product_diff(self, P, t, params, biomass_params, X_func):
        po, alpha, beta = params
        xo, xm, um = biomass_params
        X_t = X_func(t)
        dPdt = alpha * (um * X_t * (1 - X_t / xm)) + beta * X_t
        return dPdt

    def process_data(self, df):
        biomass_cols = [col for col in df.columns if 'Biomasa' in col]
        substrate_cols = [col for col in df.columns if 'Sustrato' in col]
        product_cols = [col for col in df.columns if 'Producto' in col]

        time_col = [col for col in df.columns if 'Tiempo' in col][0]
        time = df[time_col].values

        data_biomass = np.array([df[col].values for col in biomass_cols])
        self.datax.append(data_biomass)
        self.dataxp.append(np.mean(data_biomass, axis=0))
        self.datax_std.append(np.std(data_biomass, axis=0, ddof=1))

        data_substrate = np.array([df[col].values for col in substrate_cols])
        self.datas.append(data_substrate)
        self.datasp.append(np.mean(data_substrate, axis=0))
        self.datas_std.append(np.std(data_substrate, axis=0, ddof=1))

        data_product = np.array([df[col].values for col in product_cols])
        self.datap.append(data_product)
        self.datapp.append(np.mean(data_product, axis=0))
        self.datap_std.append(np.std(data_product, axis=0, ddof=1))

        self.time = time

    def fit_model(self, model_type='logistic'):
        if model_type == 'logistic':
            self.fit_biomass = self.fit_biomass_logistic
            self.fit_substrate = self.fit_substrate_logistic
            self.fit_product = self.fit_product_logistic

    def fit_biomass_logistic(self, time, biomass, bounds):
        popt, _ = curve_fit(self.logistic, time, biomass, bounds=bounds, maxfev=10000)
        self.params['biomass'] = {'xo': popt[0], 'xm': popt[1], 'um': popt[2]}
        y_pred = self.logistic(time, *popt)
        self.r2['biomass'] = 1 - (np.sum((biomass - y_pred) ** 2) / np.sum((biomass - np.mean(biomass)) ** 2))
        self.rmse['biomass'] = np.sqrt(mean_squared_error(biomass, y_pred))
        return y_pred

    def fit_substrate_logistic(self, time, substrate, biomass_params, bounds):
        popt, _ = curve_fit(lambda t, so, p, q: self.substrate(t, so, p, q, *biomass_params.values()),
                            time, substrate, bounds=bounds)
        self.params['substrate'] = {'so': popt[0], 'p': popt[1], 'q': popt[2]}
        y_pred = self.substrate(time, *popt, *biomass_params.values())
        self.r2['substrate'] = 1 - (np.sum((substrate - y_pred) ** 2) / np.sum((substrate - np.mean(substrate)) ** 2))
        self.rmse['substrate'] = np.sqrt(mean_squared_error(substrate, y_pred))
        return y_pred

    def fit_product_logistic(self, time, product, biomass_params, bounds):
        popt, _ = curve_fit(lambda t, po, alpha, beta: self.product(t, po, alpha, beta, *biomass_params.values()),
                            time, product, bounds=bounds)
        self.params['product'] = {'po': popt[0], 'alpha': popt[1], 'beta': popt[2]}
        y_pred = self.product(time, *popt, *biomass_params.values())
        self.r2['product'] = 1 - (np.sum((product - y_pred) ** 2) / np.sum((product - np.mean(product)) ** 2))
        self.rmse['product'] = np.sqrt(mean_squared_error(product, y_pred))
        return y_pred

    def plot_combined_results(self, time, biomass, substrate, product,
                              y_pred_biomass, y_pred_substrate, y_pred_product,
                              biomass_std=None, substrate_std=None, product_std=None,
                              experiment_name='', legend_position='best', params_position='upper right',
                              show_legend=True, show_params=True,
                              style='whitegrid', line_color='#0000FF', point_color='#000000',
                              line_style='-', marker_style='o'):
        sns.set_style(style)

        fig, ax1 = plt.subplots(figsize=(10, 7))
        ax1.set_xlabel('Tiempo')
        ax1.set_ylabel('Biomasa', color=line_color)

        ax1.plot(time, biomass, marker=marker_style, linestyle='', color=point_color, label='Biomasa (Datos)')
        ax1.plot(time, y_pred_biomass, linestyle=line_style, color=line_color, label='Biomasa (Modelo)')
        ax1.tick_params(axis='y', labelcolor=line_color)

        ax2 = ax1.twinx()
        ax2.set_ylabel('Sustrato', color='green')
        ax2.plot(time, substrate, marker=marker_style, linestyle='', color='green', label='Sustrato (Datos)')
        ax2.plot(time, y_pred_substrate, linestyle=line_style, color='green', label='Sustrato (Modelo)')
        ax2.tick_params(axis='y', labelcolor='green')

        ax3 = ax1.twinx()
        ax3.spines["right"].set_position(("axes", 1.1))
        ax3.set_ylabel('Producto', color='red')
        ax3.plot(time, product, marker=marker_style, linestyle='', color='red', label='Producto (Datos)')
        ax3.plot(time, y_pred_product, linestyle=line_style, color='red', label='Producto (Modelo)')
        ax3.tick_params(axis='y', labelcolor='red')

        fig.tight_layout()
        return fig



###############################

# Decorador GPU aplicado para manejar la ejecución en GPU si está disponible
@gpu_decorator(duration=300)
def generate_analysis(prompt, max_length=1024, device=None):
    try:
        # Si el dispositivo no se especifica, usa CPU por defecto
        if device is None:
            device = torch.device('cpu')
        
        # Mover el modelo al dispositivo adecuado (GPU o CPU) si es necesario
        if next(model.parameters()).device != device:
            model.to(device)
        
        # Preparar los datos de entrada en el dispositivo correcto
        input_ids = tokenizer.encode(prompt, return_tensors='pt').to(device)
        max_gen_length = min(max_length + input_ids.size(1), model.config.max_position_embeddings)

        # Generar el texto
        generated_ids = model.generate(
            input_ids=input_ids,
            max_length=max_gen_length,
            temperature=0.7,
            num_return_sequences=1,
            no_repeat_ngram_size=2,
            early_stopping=True
        )

        # Decodificar la respuesta generada
        output_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        analysis = output_text[len(prompt):].strip()
        return analysis
    except RuntimeError as e:
        return f"Error durante la ejecución: {str(e)}"
    except Exception as e:
        return f"Ocurrió un error durante el análisis: {e}"

def parse_bounds(bounds_str, num_params):
    try:
        bounds = eval(f"[{bounds_str}]")
        if len(bounds) != num_params:
            raise ValueError
        lower_bounds = [b[0] for b in bounds]
        upper_bounds = [b[1] for b in bounds]
        return lower_bounds, upper_bounds
    except:
        lower_bounds = [-np.inf] * num_params
        upper_bounds = [np.inf] * num_params
        return lower_bounds, upper_bounds

def process_and_plot(
    file,
    biomass_eq1, biomass_eq2, biomass_eq3,
    biomass_param1, biomass_param2, biomass_param3,
    biomass_bound1, biomass_bound2, biomass_bound3,
    substrate_eq1, substrate_eq2, substrate_eq3,
    substrate_param1, substrate_param2, substrate_param3,
    substrate_bound1, substrate_bound2, substrate_bound3,
    product_eq1, product_eq2, product_eq3,
    product_param1, product_param2, product_param3,
    product_bound1, product_bound2, product_bound3,
    legend_position,
    show_legend,
    show_params,
    biomass_eq_count,
    substrate_eq_count,
    product_eq_count,
    device=None
):
    # Leer el archivo Excel
    df = pd.read_excel(file.name)
    
    # Verificar que las columnas necesarias estén presentes
    expected_columns = ['Tiempo', 'Biomasa', 'Sustrato', 'Producto']
    for col in expected_columns:
        if col not in df.columns:
            raise KeyError(f"La columna esperada '{col}' no se encuentra en el archivo Excel.")

    # Asignar los datos desde las columnas
    time = df['Tiempo'].values
    biomass_data = df['Biomasa'].values
    substrate_data = df['Sustrato'].values
    product_data = df['Producto'].values

    # Convierte los contadores a enteros
    biomass_eq_count = int(biomass_eq_count)
    substrate_eq_count = int(substrate_eq_count)
    product_eq_count = int(product_eq_count)

    # Recolecta las ecuaciones, parámetros y límites según los contadores
    biomass_eqs = [biomass_eq1, biomass_eq2, biomass_eq3][:biomass_eq_count]
    biomass_params = [biomass_param1, biomass_param2, biomass_param3][:biomass_eq_count]
    biomass_bounds = [biomass_bound1, biomass_bound2, biomass_bound3][:biomass_eq_count]

    substrate_eqs = [substrate_eq1, substrate_eq2, substrate_eq3][:substrate_eq_count]
    substrate_params = [substrate_param1, substrate_param2, substrate_param3][:substrate_eq_count]
    substrate_bounds = [substrate_bound1, substrate_bound2, substrate_bound3][:substrate_eq_count]

    product_eqs = [product_eq1, product_eq2, product_eq3][:product_eq_count]
    product_params = [product_param1, product_param2, product_param3][:product_eq_count]
    product_bounds = [product_bound1, product_bound2, product_bound3][:product_eq_count]

    biomass_results = []
    substrate_results = []
    product_results = []

    # Ajusta los modelos de Biomasa
    for i in range(len(biomass_eqs)):
        equation = biomass_eqs[i]
        params_str = biomass_params[i]
        bounds_str = biomass_bounds[i]

        model = BioprocessModel()
        model.set_model('biomass', equation, params_str)

        params = [param.strip() for param in params_str.split(',')]
        lower_bounds, upper_bounds = parse_bounds(bounds_str, len(params))

        y_pred = model.fit_model(
            'biomass', time, biomass_data,
            bounds=(lower_bounds, upper_bounds)
        )
        biomass_results.append({
            'model': model,
            'y_pred': y_pred,
            'equation': equation
        })

    # Usa el primer modelo de biomasa para X(t)
    biomass_model = biomass_results[0]['model']
    X_t_func = biomass_model.models['biomass']['function']
    biomass_params_values = list(biomass_model.params['biomass'].values())

    # Ajusta los modelos de Sustrato
    for i in range(len(substrate_eqs)):
        equation = substrate_eqs[i]
        params_str = substrate_params[i]
        bounds_str = substrate_bounds[i]

        model = BioprocessModel()

        t_symbol = symbols('t')
        expr_substrate = sympify(equation)
        substrate_params_symbols = symbols([param.strip() for param in params_str.split(',')])
        substrate_func_expr = expr_substrate.subs('X(t)', X_t_func(t_symbol, *biomass_params_values))
        substrate_func = lambdify(
            (t_symbol, *substrate_params_symbols),
            substrate_func_expr,
            'numpy'
        )
        model.models['substrate'] = {
            'function': substrate_func,
            'params': [param.strip() for param in params_str.split(',')]
        }

        params = model.models['substrate']['params']
        lower_bounds, upper_bounds = parse_bounds(bounds_str, len(params))

        y_pred = model.fit_model(
            'substrate', time, substrate_data,
            bounds=(lower_bounds, upper_bounds)
        )
        substrate_results.append({
            'model': model,
            'y_pred': y_pred,
            'equation': equation
        })

    # Ajusta los modelos de Producto
    for i in range(len(product_eqs)):
        equation = product_eqs[i]
        params_str = product_params[i]
        bounds_str = product_bounds[i]

        model = BioprocessModel()

        t_symbol = symbols('t')
        expr_product = sympify(equation)
        product_params_symbols = symbols([param.strip() for param in params_str.split(',')])
        product_func_expr = expr_product.subs('X(t)', X_t_func(t_symbol, *biomass_params_values))
        product_func = lambdify(
            (t_symbol, *product_params_symbols),
            product_func_expr,
            'numpy'
        )
        model.models['product'] = {
            'function': product_func,
            'params': [param.strip() for param in params_str.split(',')]
        }

        params = model.models['product']['params']
        lower_bounds, upper_bounds = parse_bounds(bounds_str, len(params))

        y_pred = model.fit_model(
            'product', time, product_data,
            bounds=(lower_bounds, upper_bounds)
        )
        product_results.append({
            'model': model,
            'y_pred': y_pred,
            'equation': equation
        })

    # Genera las gráficas
    fig, axs = plt.subplots(3, 1, figsize=(10, 15))

    # Gráfica de Biomasa
    axs[0].plot(time, biomass_data, 'o', label='Datos de Biomasa')
    for i, result in enumerate(biomass_results):
        axs[0].plot(time, result['y_pred'], '-', label=f'Modelo de Biomasa {i+1}')
    axs[0].set_xlabel('Tiempo')
    axs[0].set_ylabel('Biomasa')
    if show_legend:
        axs[0].legend(loc=legend_position)

    # Gráfica de Sustrato
    axs[1].plot(time, substrate_data, 'o', label='Datos de Sustrato')
    for i, result in enumerate(substrate_results):
        axs[1].plot(time, result['y_pred'], '-', label=f'Modelo de Sustrato {i+1}')
    axs[1].set_xlabel('Tiempo')
    axs[1].set_ylabel('Sustrato')
    if show_legend:
        axs[1].legend(loc=legend_position)

    # Gráfica de Producto
    axs[2].plot(time, product_data, 'o', label='Datos de Producto')
    for i, result in enumerate(product_results):
        axs[2].plot(time, result['y_pred'], '-', label=f'Modelo de Producto {i+1}')
    axs[2].set_xlabel('Tiempo')
    axs[2].set_ylabel('Producto')
    if show_legend:
        axs[2].legend(loc=legend_position)

    plt.tight_layout()
    buf = io.BytesIO()
    plt.savefig(buf, format='png')
    buf.seek(0)
    image = Image.open(buf)

    prompt = f"""
Eres un experto en modelado de bioprocesos.

Analiza los siguientes resultados experimentales y proporciona un veredicto sobre la calidad de los modelos, sugiriendo mejoras si es necesario.

Biomasa:
{biomass_results}

Sustrato:
{substrate_results}

Producto:
{product_results}
"""
    analysis = generate_analysis(prompt, device=device)

    return [image], analysis