app.py

import streamlit as st
import pandas as pd
import numpy as np
from bokeh.plotting import figure
from bokeh.models import ColumnDataSource, DataTable, TableColumn, CustomJS, Select, Button, HoverTool, LinearColorMapper, ColorBar, FuncTickFormatter, FixedTicker
from bokeh.layouts import column
from bokeh.palettes import Reds9, Blues9, Oranges9, Purples9, Greys9, BuGn9, Greens9, RdYlGn11, linear_palette
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE, trustworthiness
from sklearn.metrics import pairwise_distances
from sklearn.preprocessing import MinMaxScaler
from sklearn.pipeline import Pipeline
from sklearn.base import BaseEstimator, TransformerMixin
import io
import ot
from sklearn.linear_model import LinearRegression
from scipy.stats import binned_statistic_2d
import json
import itertools
import matplotlib.pyplot as plt
import matplotlib.colors as mcolors
import zipfile
import tempfile


class RelativeScaler(BaseEstimator, TransformerMixin):
    """
    Escala la primera columna al rango [-1, 1] y ajusta las siguientes columnas
    manteniendo la proporción relativa respecto a la primera.
    """
    
    def fit(self, X, y=None):
        X = np.asarray(X)
        self.pc1_min_ = X[:, 0].min()
        self.pc1_max_ = X[:, 0].max()
        self.scaling_factor_ = (self.pc1_max_ - self.pc1_min_) / 2
        return self

    def transform(self, X):
        X = np.asarray(X)
        pc1 = X[:, 0]
        pc1_scaled = 2 * (pc1 - self.pc1_min_) / (self.pc1_max_ - self.pc1_min_) - 1
        
        transformed = [pc1_scaled.reshape(-1, 1)]
        
        for i in range(1, X.shape[1]):
            scaled_i = X[:, i] / self.scaling_factor_
            transformed.append(scaled_i.reshape(-1, 1))
        
        return np.hstack(transformed)

N_COMPONENTS = 3
TSNE_NEIGHBOURS = 150
# WEIGHT_FACTOR = 0.05

TOOLTIPS = """
<div>
    <div>
        <img src="@img{safe}" style="width:128px; height:auto; float: left; margin: 0px 15px 15px 0px;" alt="@img" border="2"></img>
    </div>
    <div>
        <span style="font-size: 17px; font-weight: bold;">@label</span>
    </div>
    <div>
        <span style="font-size: 14px;">X: @x, Y: @y</span>
    </div>
</div>
"""

def config_style():
    # st.set_page_config(layout="wide")
    st.markdown("""
        <style>
        .main-title { font-size: 50px; color: #4CAF50; text-align: center; }
        .sub-title { font-size: 30px; color: #555; }
        .custom-text { font-size: 18px; line-height: 1.5; }
        .bk-legend {
            max-height: 200px;
            overflow-y: auto;
        }
        </style>
    """, unsafe_allow_html=True)
    st.markdown('<h1 class="main-title">Merit Embeddings 🎒📃🏆</h1>', unsafe_allow_html=True)

def load_embeddings(model, version, embedding_prefix, weight_factor):
    if model == "Donut":
        df_real = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_secret_all_{weight_factor}embeddings.csv")
        df_par = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-paragraph-degradation-seq_{weight_factor}embeddings.csv")
        df_line = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-line-degradation-seq_{weight_factor}embeddings.csv")
        df_seq  = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-seq_{weight_factor}embeddings.csv")
        df_rot  = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-rotation-degradation-seq_{weight_factor}embeddings.csv")
        df_zoom = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-zoom-degradation-seq_{weight_factor}embeddings.csv")
        df_render = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_es-render-seq_{weight_factor}embeddings.csv")
        df_pretratrained = pd.read_csv(f"data/donut/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_IIT-CDIP_{weight_factor}embeddings.csv")
        
        # Asignar etiquetas de versión
        df_real["version"] = "real"
        df_par["version"] = "synthetic"
        df_line["version"] = "synthetic"
        df_seq["version"] = "synthetic"
        df_rot["version"] = "synthetic"
        df_zoom["version"] = "synthetic"
        df_render["version"] = "synthetic"
        df_pretratrained["version"] = "pretrained"

        # Asignar fuente (source)
        df_par["source"] = "es-digital-paragraph-degradation-seq"
        df_line["source"] = "es-digital-line-degradation-seq"
        df_seq["source"] = "es-digital-seq"
        df_rot["source"] = "es-digital-rotation-degradation-seq"
        df_zoom["source"] = "es-digital-zoom-degradation-seq"
        df_render["source"] = "es-render-seq"
        df_pretratrained["source"] = "pretrained"

        return {"real": df_real, 
                "synthetic": pd.concat([df_seq, df_line, df_par, df_rot, df_zoom, df_render], ignore_index=True),
                "pretrained": df_pretratrained}
    
    elif model == "Idefics2":
        df_real = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_secret_britanico_{weight_factor}embeddings.csv")
        df_par = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-paragraph-degradation-seq_{weight_factor}embeddings.csv")
        df_line = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-line-degradation-seq_{weight_factor}embeddings.csv")
        df_seq  = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-seq_{weight_factor}embeddings.csv")
        df_rot  = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-rotation-degradation-seq_{weight_factor}embeddings.csv")
        df_zoom = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-zoom-degradation-seq_{weight_factor}embeddings.csv")
        df_render = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-render-seq_{weight_factor}embeddings.csv")
        
        # Cargar ambos subconjuntos pretrained y combinarlos
        df_pretratrained_PDFA = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_PDFA_{weight_factor}embeddings.csv")
        df_pretratrained_IDL = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_IDL_{weight_factor}embeddings.csv")
        df_pretratrained = pd.concat([df_pretratrained_PDFA, df_pretratrained_IDL], ignore_index=True)
        
        # Asignar etiquetas de versión
        df_real["version"] = "real"
        df_par["version"] = "synthetic"
        df_line["version"] = "synthetic"
        df_seq["version"] = "synthetic"
        df_rot["version"] = "synthetic"
        df_zoom["version"] = "synthetic"
        df_render["version"] = "synthetic"
        df_pretratrained["version"] = "pretrained"

        # Asignar fuente (source)
        df_par["source"] = "es-digital-paragraph-degradation-seq"
        df_line["source"] = "es-digital-line-degradation-seq"
        df_seq["source"] = "es-digital-seq"
        df_rot["source"] = "es-digital-rotation-degradation-seq"
        df_zoom["source"] = "es-digital-zoom-degradation-seq"
        df_render["source"] = "es-render-seq"
        df_pretratrained["source"] = "pretrained"

        return {"real": df_real, 
                "synthetic": pd.concat([df_seq, df_line, df_par, df_rot, df_zoom, df_render], ignore_index=True),
                "pretrained": df_pretratrained}
    
    elif model == "Idefics2-patient":
        df_real = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_secret_britanico_{weight_factor}embeddings.csv")
        df_par = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-paragraph-degradation-seq_{weight_factor}embeddings.csv")
        df_line = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-line-degradation-seq_{weight_factor}embeddings.csv")
        df_seq  = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-seq_{weight_factor}embeddings.csv")
        df_rot  = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-rotation-degradation-seq_{weight_factor}embeddings.csv")
        df_zoom = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-zoom-degradation-seq_{weight_factor}embeddings.csv")
        df_render = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_es-render-seq_{weight_factor}embeddings.csv")
        
        # Cargar ambos subconjuntos pretrained y combinarlos
        df_pretratrained_PDFA = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_PDFA_{weight_factor}embeddings.csv")
        df_pretratrained_IDL = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_IDL_{weight_factor}embeddings.csv")
        df_pretratrained = pd.concat([df_pretratrained_PDFA, df_pretratrained_IDL], ignore_index=True)
        
        # Asignar etiquetas de versión
        df_real["version"] = "real"
        df_par["version"] = "synthetic"
        df_line["version"] = "synthetic"
        df_seq["version"] = "synthetic"
        df_rot["version"] = "synthetic"
        df_zoom["version"] = "synthetic"
        df_render["version"] = "synthetic"
        df_pretratrained["version"] = "pretrained"

        # Asignar fuente (source)
        df_par["source"] = "es-digital-paragraph-degradation-seq"
        df_line["source"] = "es-digital-line-degradation-seq"
        df_seq["source"] = "es-digital-seq"
        df_rot["source"] = "es-digital-rotation-degradation-seq"
        df_zoom["source"] = "es-digital-zoom-degradation-seq"
        df_render["source"] = "es-render-seq"
        df_pretratrained["source"] = "pretrained"

        return {"real": df_real, 
                "synthetic": pd.concat([df_seq, df_line, df_par, df_rot, df_zoom, df_render], ignore_index=True),
                "pretrained": df_pretratrained}
    
    elif model == "Paligemma":
        df_real = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_secret_all_{weight_factor}embeddings.csv")
        df_par = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-paragraph-degradation-seq_{weight_factor}embeddings.csv")
        df_line = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-line-degradation-seq_{weight_factor}embeddings.csv")
        df_seq  = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-seq_{weight_factor}embeddings.csv")
        df_rot  = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-rotation-degradation-seq_{weight_factor}embeddings.csv")
        df_zoom = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-zoom-degradation-seq_{weight_factor}embeddings.csv")
        df_render = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-render-seq_{weight_factor}embeddings.csv")
        
        # Cargar ambos subconjuntos pretrained y combinarlos
        # TODO Pretrained de idefics2, se mantienen para evitar error, pero se debe meter los de paligemma
        df_pretratrained_PDFA = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_PDFA_{weight_factor}embeddings.csv")
        df_pretratrained_IDL = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_IDL_{weight_factor}embeddings.csv")
        df_pretratrained = pd.concat([df_pretratrained_PDFA, df_pretratrained_IDL], ignore_index=True)
        
        # Asignar etiquetas de versión
        df_real["version"] = "real"
        df_par["version"] = "synthetic"
        df_line["version"] = "synthetic"
        df_seq["version"] = "synthetic"
        df_rot["version"] = "synthetic"
        df_zoom["version"] = "synthetic"
        df_render["version"] = "synthetic"
        df_pretratrained["version"] = "pretrained"

        # Asignar fuente (source)
        df_par["source"] = "es-digital-paragraph-degradation-seq"
        df_line["source"] = "es-digital-line-degradation-seq"
        df_seq["source"] = "es-digital-seq"
        df_rot["source"] = "es-digital-rotation-degradation-seq"
        df_zoom["source"] = "es-digital-zoom-degradation-seq"
        df_render["source"] = "es-render-seq"
        df_pretratrained["source"] = "pretrained"

        return {"real": df_real, 
                "synthetic": pd.concat([df_seq, df_line, df_par, df_rot, df_zoom, df_render], ignore_index=True),
                "pretrained": df_pretratrained}
    
    elif model == "Llava":
        df_real = pd.read_csv(f"data/llava/{version}/{embedding_prefix}/de_Rodrigo_merit_secret_all_{weight_factor}embeddings.csv")
        #TODO Embeddings de Paligemma se mantienen para evitar error
        df_par = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-paragraph-degradation-seq_{weight_factor}embeddings.csv")
        df_line = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-line-degradation-seq_{weight_factor}embeddings.csv")
        df_seq  = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-seq_{weight_factor}embeddings.csv")
        df_rot  = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-rotation-degradation-seq_{weight_factor}embeddings.csv")
        df_zoom = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-digital-zoom-degradation-seq_{weight_factor}embeddings.csv")
        df_render = pd.read_csv(f"data/paligemma/{version}/{embedding_prefix}/de_Rodrigo_merit_es-render-seq_{weight_factor}embeddings.csv")
        
        # Cargar ambos subconjuntos pretrained y combinarlos
        # TODO Pretrained de idefics2, se mantienen para evitar error, pero se debe meter los de paligemma
        df_pretratrained_PDFA = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_PDFA_{weight_factor}embeddings.csv")
        df_pretratrained_IDL = pd.read_csv(f"data/idefics2/{version}/{embedding_prefix}/de_Rodrigo_merit_aux_IDL_{weight_factor}embeddings.csv")
        df_pretratrained = pd.concat([df_pretratrained_PDFA, df_pretratrained_IDL], ignore_index=True)
        
        # Asignar etiquetas de versión
        df_real["version"] = "real"
        df_par["version"] = "synthetic"
        df_line["version"] = "synthetic"
        df_seq["version"] = "synthetic"
        df_rot["version"] = "synthetic"
        df_zoom["version"] = "synthetic"
        df_render["version"] = "synthetic"
        df_pretratrained["version"] = "pretrained"

        # Asignar fuente (source)
        df_par["source"] = "es-digital-paragraph-degradation-seq"
        df_line["source"] = "es-digital-line-degradation-seq"
        df_seq["source"] = "es-digital-seq"
        df_rot["source"] = "es-digital-rotation-degradation-seq"
        df_zoom["source"] = "es-digital-zoom-degradation-seq"
        df_render["source"] = "es-render-seq"
        df_pretratrained["source"] = "pretrained"

        return {"real": df_real, 
                "synthetic": pd.concat([df_seq, df_line, df_par, df_rot, df_zoom, df_render], ignore_index=True),
                "pretrained": df_pretratrained}
    
    
    else:
        st.error("Modelo no reconocido")
        return None

def split_versions(df_combined, reduced):
    # Asignar las coordenadas si la reducción es 2D
    if reduced.shape[1] == 2:
        df_combined['x'] = reduced[:, 0]
        df_combined['y'] = reduced[:, 1]
    df_real = df_combined[df_combined["version"] == "real"].copy()
    df_synth = df_combined[df_combined["version"] == "synthetic"].copy()
    df_pretrained = df_combined[df_combined["version"] == "pretrained"].copy()
    
    unique_real = sorted(df_real['label'].unique().tolist())
    unique_synth = {}
    for source in df_synth["source"].unique():
        unique_synth[source] = sorted(df_synth[df_synth["source"] == source]['label'].unique().tolist())
    unique_pretrained = sorted(df_pretrained['label'].unique().tolist())
    
    df_dict = {"real": df_real, "synthetic": df_synth, "pretrained": df_pretrained}
    unique_subsets = {"real": unique_real, "synthetic": unique_synth, "pretrained": unique_pretrained}
    return df_dict, unique_subsets

def get_embedding_from_df(df):
    # Retorna el embedding completo (4 dimensiones en este caso) guardado en la columna 'embedding'
    if 'embedding' in df.columns:
        return np.stack(df['embedding'].to_numpy())
    elif 'x' in df.columns and 'y' in df.columns:
        return df[['x', 'y']].values
    else:
        raise ValueError("No se encontró embedding o coordenadas x,y en el DataFrame.")

def compute_cluster_distance(synthetic_points, real_points, metric="wasserstein", bins=20):
    if metric.lower() == "wasserstein":
        n = synthetic_points.shape[0]
        m = real_points.shape[0]
        weights = np.ones(n) / n
        weights_real = np.ones(m) / m
        M = ot.dist(synthetic_points, real_points, metric='euclidean')
        return ot.emd2(weights, weights_real, M)
    elif metric.lower() == "euclidean":
        center_syn = np.mean(synthetic_points, axis=0)
        center_real = np.mean(real_points, axis=0)
        return np.linalg.norm(center_syn - center_real)
    elif metric.lower() == "kl":
        # Para KL usamos histogramas multidimensionales con límites globales en cada dimensión
        all_points = np.vstack([synthetic_points, real_points])
        edges = [
            np.linspace(np.min(all_points[:, i]), np.max(all_points[:, i]), bins+1)
            for i in range(all_points.shape[1])
        ]
        H_syn, _ = np.histogramdd(synthetic_points, bins=edges)
        H_real, _ = np.histogramdd(real_points, bins=edges)
        eps = 1e-10
        P = H_syn + eps
        Q = H_real + eps
        P = P / P.sum()
        Q = Q / Q.sum()
        kl = np.sum(P * np.log(P / Q))
        return kl
    else:
        raise ValueError("Métrica desconocida. Usa 'wasserstein', 'euclidean' o 'kl'.")

def compute_cluster_distances_synthetic_individual(synthetic_df: pd.DataFrame, df_real: pd.DataFrame, real_labels: list, metric="wasserstein", bins=20) -> pd.DataFrame:
    distances = {}
    groups = synthetic_df.groupby(['source', 'label'])
    for (source, label), group in groups:
        key = f"{label} ({source})"
        data = get_embedding_from_df(group)
        distances[key] = {}
        for real_label in real_labels:
            real_data = get_embedding_from_df(df_real[df_real['label'] == real_label])
            d = compute_cluster_distance(data, real_data, metric=metric, bins=bins)
            distances[key][real_label] = d
    for source, group in synthetic_df.groupby('source'):
        key = f"Global ({source})"
        data = get_embedding_from_df(group)
        distances[key] = {}
        for real_label in real_labels:
            real_data = get_embedding_from_df(df_real[df_real['label'] == real_label])
            d = compute_cluster_distance(data, real_data, metric=metric, bins=bins)
            distances[key][real_label] = d
    return pd.DataFrame(distances).T

def compute_continuity(X, X_embedded, n_neighbors=5):
    n = X.shape[0]
    D_high = pairwise_distances(X, metric='euclidean')
    D_low = pairwise_distances(X_embedded, metric='euclidean')
    indices_high = np.argsort(D_high, axis=1)
    indices_low = np.argsort(D_low, axis=1)
    k_high = indices_high[:, 1:n_neighbors+1]
    k_low = indices_low[:, 1:n_neighbors+1]
    total = 0.0
    for i in range(n):
        set_high = set(k_high[i])
        set_low = set(k_low[i])
        missing = set_high - set_low
        for j in missing:
            rank = np.where(indices_low[i] == j)[0][0]
            total += (rank - n_neighbors)
    norm = 2.0 / (n * n_neighbors * (2*n - 3*n_neighbors - 1))
    continuity_value = 1 - norm * total
    return continuity_value

def create_table(df_distances):
    df_table = df_distances.copy()
    df_table.reset_index(inplace=True)
    df_table.rename(columns={'index': 'Synthetic'}, inplace=True)
    min_row = {"Synthetic": "Min."}
    mean_row = {"Synthetic": "Mean"}
    max_row = {"Synthetic": "Max."}
    for col in df_table.columns:
        if col != "Synthetic":
            min_row[col] = df_table[col].min()
            mean_row[col] = df_table[col].mean()
            max_row[col] = df_table[col].max()
    df_table = pd.concat([df_table, pd.DataFrame([min_row, mean_row, max_row])], ignore_index=True)
    source_table = ColumnDataSource(df_table)
    columns = [TableColumn(field='Synthetic', title='Synthetic')]
    for col in df_table.columns:
        if col != 'Synthetic':
            columns.append(TableColumn(field=col, title=col))
    total_height = 30 + len(df_table)*28
    data_table = DataTable(source=source_table, columns=columns, sizing_mode='stretch_width', height=total_height)
    return data_table, df_table, source_table

def create_figure(dfs, unique_subsets, color_maps, model_name):
    # Se crea el plot para el embedding reducido (asumiendo que es 2D)
    fig = figure(width=600, height=600, tools="wheel_zoom,pan,reset,save", active_scroll="wheel_zoom", tooltips=TOOLTIPS, title="")
    fig.match_aspect = True
    
    # Renderizar datos reales
    real_renderers = add_dataset_to_fig(fig, dfs["real"], unique_subsets["real"],
                                        marker="circle", color_mapping=color_maps["real"],
                                        group_label="Real")
    
    # Renderizar datos sintéticos (por fuente)
    marker_mapping = {
        "es-digital-paragraph-degradation-seq": "x",
        "es-digital-line-degradation-seq": "cross",
        "es-digital-seq": "triangle",
        "es-digital-rotation-degradation-seq": "diamond",
        "es-digital-zoom-degradation-seq": "asterisk",
        "es-render-seq": "inverted_triangle"
    }
    synthetic_renderers = {}
    synth_df = dfs["synthetic"]
    for source in unique_subsets["synthetic"]:
        df_source = synth_df[synth_df["source"] == source]
        marker = marker_mapping.get(source, "square")
        renderers = add_synthetic_dataset_to_fig(fig, df_source, unique_subsets["synthetic"][source],
                                                  marker=marker,
                                                  color_mapping=color_maps["synthetic"][source],
                                                  group_label=source)
        synthetic_renderers.update(renderers)
    
    # Agregar el subset pretrained (se puede usar un marcador distinto, por ejemplo, "triangle")
    pretrained_renderers = add_dataset_to_fig(fig, dfs["pretrained"], unique_subsets["pretrained"],
                                               marker="triangle", color_mapping=color_maps["pretrained"],
                                               group_label="Pretrained")
    
    fig.legend.location = "top_right"
    fig.legend.click_policy = "hide"
    show_legend = st.checkbox("Show Legend", value=False, key=f"legend_{model_name}")
    fig.legend.visible = show_legend
    return fig, real_renderers, synthetic_renderers, pretrained_renderers

def add_dataset_to_fig(fig, df, selected_labels, marker, color_mapping, group_label):
    renderers = {}
    for label in selected_labels:
        subset = df[df['label'] == label]
        if subset.empty:
            continue
        source = ColumnDataSource(data=dict(
            x=subset['x'],
            y=subset['y'],
            label=subset['label'],
            img=subset.get('img', "")
        ))
        color = color_mapping[label]
        legend_label = f"{label} ({group_label})"
        if marker == "circle":
            r = fig.circle('x', 'y', size=10, source=source,
                           fill_color=color, line_color=color,
                           legend_label=legend_label)
        elif marker == "square":
            r = fig.square('x', 'y', size=10, source=source,
                           fill_color=color, line_color=color,
                           legend_label=legend_label)
        elif marker == "triangle":
            r = fig.triangle('x', 'y', size=12, source=source,
                             fill_color=color, line_color=color,
                             legend_label=legend_label)
        renderers[label + f" ({group_label})"] = r
    return renderers

def add_synthetic_dataset_to_fig(fig, df, labels, marker, color_mapping, group_label):
    renderers = {}
    for label in labels:
        subset = df[df['label'] == label]
        if subset.empty:
            continue
        source_obj = ColumnDataSource(data=dict(
            x=subset['x'],
            y=subset['y'],
            label=subset['label'],
            img=subset.get('img', "")
        ))
        color = color_mapping[label]
        legend_label = group_label
        if marker == "square":
            r = fig.square('x', 'y', size=10, source=source_obj,
                           fill_color=color, line_color=color,
                           legend_label=legend_label)
        elif marker == "triangle":
            r = fig.triangle('x', 'y', size=12, source=source_obj,
                             fill_color=color, line_color=color,
                             legend_label=legend_label)
        elif marker == "inverted_triangle":
            r = fig.inverted_triangle('x', 'y', size=12, source=source_obj,
                                      fill_color=color, line_color=color,
                                      legend_label=legend_label)
        elif marker == "diamond":
            r = fig.diamond('x', 'y', size=10, source=source_obj,
                            fill_color=color, line_color=color,
                            legend_label=legend_label)
        elif marker == "cross":
            r = fig.cross('x', 'y', size=12, source=source_obj,
                          fill_color=color, line_color=color,
                          legend_label=legend_label)
        elif marker == "x":
            r = fig.x('x', 'y', size=12, source=source_obj,
                      fill_color=color, line_color=color,
                      legend_label=legend_label)
        elif marker == "asterisk":
            r = fig.asterisk('x', 'y', size=12, source=source_obj,
                             fill_color=color, line_color=color,
                             legend_label=legend_label)
        else:
            r = fig.circle('x', 'y', size=10, source=source_obj,
                           fill_color=color, line_color=color,
                           legend_label=legend_label)
        renderers[label + f" ({group_label})"] = r
    return renderers

def get_color_maps(unique_subsets):
    color_map = {}
    num_real = len(unique_subsets["real"])
    red_palette = Reds9[:num_real] if num_real <= 9 else (Reds9 * ((num_real // 9) + 1))[:num_real]
    color_map["real"] = {label: red_palette[i] for i, label in enumerate(sorted(unique_subsets["real"]))}
    
    color_map["synthetic"] = {}
    for source, labels in unique_subsets["synthetic"].items():
        if source == "es-digital-seq":
            palette = Blues9[:len(labels)] if len(labels) <= 9 else (Blues9 * ((len(labels)//9)+1))[:len(labels)]
        elif source == "es-digital-line-degradation-seq":
            palette = Purples9[:len(labels)] if len(labels) <= 9 else (Purples9 * ((len(labels)//9)+1))[:len(labels)]
        elif source == "es-digital-paragraph-degradation-seq":
            palette = BuGn9[:len(labels)] if len(labels) <= 9 else (BuGn9 * ((len(labels)//9)+1))[:len(labels)]
        elif source == "es-digital-rotation-degradation-seq":
            palette = Greys9[:len(labels)] if len(labels) <= 9 else (Greys9 * ((len(labels)//9)+1))[:len(labels)]
        elif source == "es-digital-zoom-degradation-seq":
            palette = Oranges9[:len(labels)] if len(labels) <= 9 else (Oranges9 * ((len(labels)//9)+1))[:len(labels)]
        elif source == "es-render-seq":
            palette = Greens9[:len(labels)] if len(labels) <= 9 else (Greens9 * ((len(labels)//9)+1))[:len(labels)]
        else:
            palette = Blues9[:len(labels)] if len(labels) <= 9 else (Blues9 * ((len(labels)//9)+1))[:len(labels)]
        color_map["synthetic"][source] = {label: palette[i] for i, label in enumerate(sorted(labels))}
    
    # Asignar colores al subset pretrained usando, por ejemplo, la paleta Purples9
    num_pretrained = len(unique_subsets["pretrained"])
    purple_palette = Purples9[:num_pretrained] if num_pretrained <= 9 else (Purples9 * ((num_pretrained // 9) + 1))[:num_pretrained]
    color_map["pretrained"] = {label: purple_palette[i] for i, label in enumerate(sorted(unique_subsets["pretrained"]))}
    
    return color_map

def calculate_cluster_centers(df, labels):
    centers = {}
    for label in labels:
        subset = df[df['label'] == label]
        if not subset.empty and 'x' in subset.columns and 'y' in subset.columns:
            centers[label] = (subset['x'].mean(), subset['y'].mean())
    return centers

def compute_global_regression(df_combined, embedding_cols, tsne_params, df_f1, reduction_method="t-SNE", distance_metric="wasserstein"):
    if reduction_method == "PCA":
        reducer = Pipeline([
            ("pca", PCA(n_components=N_COMPONENTS)),
            ("rel_scaler", RelativeScaler())
        ])
    else:
        reducer = TSNE(n_components=2, random_state=42, 
                         perplexity=tsne_params["perplexity"], 
                         learning_rate=tsne_params["learning_rate"])
    
    reduced = reducer.fit_transform(df_combined[embedding_cols].values)
    # Guardamos el embedding completo (por ejemplo, 4 dimensiones en PCA)
    df_combined['embedding'] = list(reduced)
    # Si el embedding es 2D, asignamos x e y para visualización
    if reduced.shape[1] == 2:
        df_combined['x'] = reduced[:, 0]
        df_combined['y'] = reduced[:, 1]
    
    explained_variance = None
    if reduction_method == "PCA":
        explained_variance = reducer.named_steps["pca"].explained_variance_ratio_
    
    trust = None
    cont = None
    if reduction_method == "t-SNE":
        X = df_combined[embedding_cols].values
        trust = trustworthiness(X, reduced, n_neighbors=TSNE_NEIGHBOURS)
        cont = compute_continuity(X, reduced, n_neighbors=TSNE_NEIGHBOURS)
    
    dfs_reduced, unique_subsets = split_versions(df_combined, reduced)
    
    df_distances = compute_cluster_distances_synthetic_individual(
        dfs_reduced["synthetic"],
        dfs_reduced["real"],
        unique_subsets["real"],
        metric=distance_metric
    )
    
    global_distances = {}
    for idx in df_distances.index:
        if idx.startswith("Global"):
            source = idx.split("(")[1].rstrip(")")
            global_distances[source] = df_distances.loc[idx].values
            
    all_x = []
    all_y = []
    for source in df_f1.columns:
        if source in global_distances:
            x_vals = global_distances[source]
            y_vals = df_f1[source].values
            all_x.extend(x_vals)
            all_y.extend(y_vals)
    all_x_arr = np.array(all_x).reshape(-1, 1)
    all_y_arr = np.array(all_y)
    
    model_global = LinearRegression().fit(all_x_arr, all_y_arr)
    r2 = model_global.score(all_x_arr, all_y_arr)
    slope = model_global.coef_[0]
    intercept = model_global.intercept_
    
    scatter_fig = figure(width=600, height=600, tools="pan,wheel_zoom,reset,save", y_range=(0, 1), 
                         title="Scatter Plot: Distance vs F1")
    source_colors = {
        "es-digital-paragraph-degradation-seq": "blue",
        "es-digital-line-degradation-seq": "green",
        "es-digital-seq": "red",
        "es-digital-zoom-degradation-seq": "orange",
        "es-digital-rotation-degradation-seq": "purple",
        "es-digital-rotation-zoom-degradation-seq": "brown",
        "es-render-seq": "cyan"
    }
    for source in df_f1.columns:
        if source in global_distances:
            x_vals = global_distances[source]
            y_vals = df_f1[source].values
            data = {"x": x_vals, "y": y_vals, "Fuente": [source]*len(x_vals)}
            cds = ColumnDataSource(data=data)
            scatter_fig.circle('x', 'y', size=8, alpha=0.7, source=cds,
                               fill_color=source_colors.get(source, "gray"),
                               line_color=source_colors.get(source, "gray"),
                               legend_label=source)
    scatter_fig.xaxis.axis_label = "Distance (Global, por Colegio)"
    scatter_fig.yaxis.axis_label = "F1 Score"
    scatter_fig.legend.location = "top_right"
    hover_tool = HoverTool(tooltips=[("Distance", "@x"), ("F1", "@y"), ("Subset", "@Fuente")])
    scatter_fig.add_tools(hover_tool)
    # scatter_fig.match_aspect = True
    
    x_line = np.linspace(all_x_arr.min(), all_x_arr.max(), 100)
    y_line = model_global.predict(x_line.reshape(-1, 1))
    scatter_fig.line(x_line, y_line, line_width=2, line_color="black", legend_label="Global Regression")
    
    results = {
        "R2": r2,
        "slope": slope,
        "intercept": intercept,
        "scatter_fig": scatter_fig,
        "dfs_reduced": dfs_reduced,
        "unique_subsets": unique_subsets,
        "df_distances": df_distances,
        "explained_variance": explained_variance,
        "trustworthiness": trust,
        "continuity": cont
    }
    
    if reduction_method == "PCA":
        results["pca_model"] = reducer  # Agregamos el objeto PCA para usarlo luego en los plots
    
    return results

# def get_color(color_entry):
#     if isinstance(color_entry, dict):
#         # Extrae el primer valor (o ajusta según convenga)
#         return list(color_entry.values())[0]
#     return color_entry

def optimize_tsne_params(df_combined, embedding_cols, df_f1, distance_metric):
    perplexity_range = np.linspace(30, 50, 10)
    learning_rate_range = np.linspace(200, 1000, 20)
    
    best_R2 = -np.inf
    best_params = None
    total_steps = len(perplexity_range) * len(learning_rate_range)
    step = 0

    progress_text = st.empty()
    
    for p in perplexity_range:
        for lr in learning_rate_range:
            step += 1
            progress_text.text(f"Evaluating: Perplexity={p:.2f}, Learning Rate={lr:.2f} (Step {step}/{total_steps})")
            
            tsne_params = {"perplexity": p, "learning_rate": lr}
            result = compute_global_regression(df_combined, embedding_cols, tsne_params, df_f1, reduction_method="t-SNE", distance_metric=distance_metric)
            r2_temp = result["R2"]
            st.write(f"Parameters: Perplexity={p:.2f}, Learning Rate={lr:.2f} -> R²={r2_temp:.4f}")
            
            if r2_temp > best_R2:
                best_R2 = r2_temp
                best_params = (p, lr)
    
    progress_text.text("Optimization completed!")
    return best_params, best_R2

def run_model(model_name):
    version = st.selectbox("Select Model Version:", options=["vanilla", "finetuned_real"], key=f"version_{model_name}")
    # Selector para el método de cómputo del embedding
    embedding_computation = st.selectbox("¿Cómo se computa el embedding?", options=["averaged", "weighted"], key=f"embedding_method_{model_name}")
    # Se asigna el prefijo correspondiente

    if embedding_computation == "weighted":
         selected_weight_factor = st.selectbox(
             "Seleccione el Weight Factor", 
             options=[0.05, 0.1, 0.25, 0.5], 
             index=0,  # índice 1 para que por defecto sea 0.05
             key=f"weight_factor_{model_name}"
         )
         weight_factor = f"{selected_weight_factor}_"
    else:
         weight_factor = ""
    
    embeddings = load_embeddings(model_name, version, embedding_computation, weight_factor)
    if embeddings is None:
        return

    # Nuevo selector para incluir o excluir el dataset pretrained
    include_pretrained = st.checkbox("Incluir dataset pretrained", value=False, key=f"legend_{model_name}_pretrained")
    if not include_pretrained:
        # Removemos la entrada pretrained del diccionario, si existe.
        embeddings.pop("pretrained", None)
    
    # Extraer columnas de embedding de los datos "real"
    embedding_cols = [col for col in embeddings["real"].columns if col.startswith("dim_")]
    # Concatenamos los datasets disponibles (ahora, sin pretrained si se deseleccionó)
    df_combined = pd.concat(list(embeddings.values()), ignore_index=True)
    
    try:
        df_f1 = pd.read_csv("data/f1-donut.csv", sep=';', index_col=0)
    except Exception as e:
        st.error(f"Error loading f1-donut.csv: {e}")
        return

    st.markdown('<h6 class="sub-title">Select Dimensionality Reduction Method</h6>', unsafe_allow_html=True)
    reduction_method = st.selectbox("", options=["PCA", "t-SNE"], key=f"reduction_{model_name}")
    
    distance_metric = st.selectbox("Select Distance Metric:", 
                                   options=["Euclidean", "Wasserstein", "KL"], 
                                   key=f"distance_metric_{model_name}")
    
    tsne_params = {}
    if reduction_method == "t-SNE":
        if st.button("Optimize TSNE parameters", key=f"optimize_tsne_{model_name}"):
            st.info("Running optimization, this can take a while...")
            best_params, best_R2 = optimize_tsne_params(df_combined, embedding_cols, df_f1, distance_metric.lower())
            st.success(f"Best parameters: Perplexity = {best_params[0]:.2f}, Learning Rate = {best_params[1]:.2f} with R² = {best_R2:.4f}")
            tsne_params = {"perplexity": best_params[0], "learning_rate": best_params[1]}
        else:
            perplexity_val = st.number_input(
                "Perplexity", 
                min_value=5.0, 
                max_value=50.0, 
                value=30.0, 
                step=1.0, 
                format="%.2f",
                key=f"perplexity_{model_name}"
            )
            learning_rate_val = st.number_input(
                "Learning Rate", 
                min_value=10.0, 
                max_value=1000.0, 
                value=200.0, 
                step=10.0, 
                format="%.2f",
                key=f"learning_rate_{model_name}"
            )
            tsne_params = {"perplexity": perplexity_val, "learning_rate": learning_rate_val}
    
    result = compute_global_regression(df_combined, embedding_cols, tsne_params, df_f1, reduction_method=reduction_method, distance_metric=distance_metric.lower())
    
    reg_metrics = pd.DataFrame({
        "Slope": [result["slope"]],
        "Intercept": [result["intercept"]],
        "R2": [result["R2"]]
    })
    st.table(reg_metrics)
    
    if reduction_method == "PCA" and result["explained_variance"] is not None:
        st.subheader("Explained Variance Ratio")
        component_names = [f"PC{i+1}" for i in range(len(result["explained_variance"]))]
        variance_df = pd.DataFrame({
            "Component": component_names,
            "Explained Variance": result["explained_variance"]
        })
        st.table(variance_df)
    elif reduction_method == "t-SNE":
        st.subheader("t-SNE Quality Metrics")
        st.write(f"Trustworthiness: {result['trustworthiness']:.4f}")
        st.write(f"Continuity: {result['continuity']:.4f}")

    # # Mostrar los plots de loadings si se usó PCA (para el conjunto combinado)
    # if reduction_method == "PCA" and result.get("pca_model") is not None:
    #     # pca_model = result["pca_model"]
    #     pca_model = result["pca_model"].named_steps["pca"]
    #     components = pca_model.components_  # Shape: (n_components, n_features)
        
    #     st.subheader("Pesos de las Componentes Principales (Loadings) - Conjunto Combinado")
    #     for i, comp in enumerate(components):
    #         source = ColumnDataSource(data=dict(
    #             dimensions=embedding_cols,
    #             weight=comp
    #         ))
    #         p = figure(x_range=embedding_cols, title=f"Componente Principal {i+1}",
    #                    plot_height=400, plot_width=600,
    #                    toolbar_location="above",
    #                    tools="pan,wheel_zoom,reset,save,hover",
    #                    active_scroll="wheel_zoom")

    #         # Establecer fondo blanco
    #         p.background_fill_color = "white"
    #         # Mostrar solo grilla horizontal
    #         p.xgrid.grid_line_color = None
    #         p.ygrid.grid_line_color = "gray"
    #         p.vbar(x='dimensions', top='weight', width=0.8, source=source)
    #         p.xaxis.major_label_text_font_size = '0pt'
    #         hover = HoverTool(tooltips=[("Dimensión", "@dimensions"), ("Peso", "@weight")])
    #         p.add_tools(hover)
    #         p.xaxis.axis_label = "Dimensiones originales"
    #         p.yaxis.axis_label = "Peso"
    #         st.bokeh_chart(p)
    
    data_table, df_table, source_table = create_table(result["df_distances"])
    real_subset_names = list(df_table.columns[1:])
    real_select = Select(title="", value=real_subset_names[0], options=real_subset_names)
    reset_button = Button(label="Reset Colors", button_type="primary")
    line_source = ColumnDataSource(data={'x': [], 'y': []})
    
    # if (reduction_method == "t-SNE" and N_COMPONENTS == 2) or (reduction_method == "PCA" and N_COMPONENTS == 2):
    #     fig, real_renderers, synthetic_renderers, pretrained_renderers = create_figure(
    #         result["dfs_reduced"],
    #         result["unique_subsets"],
    #         get_color_maps(result["unique_subsets"]),
    #         model_name
    #     )
    #     fig.line('x', 'y', source=line_source, line_width=2, line_color='black')
    #     centers_real = calculate_cluster_centers(result["dfs_reduced"]["real"], result["unique_subsets"]["real"])
    #     real_centers_js = {k: [v[0], v[1]] for k, v in centers_real.items()}
    #     synthetic_centers = {}
    #     synth_labels = sorted(result["dfs_reduced"]["synthetic"]['label'].unique().tolist())
    #     for label in synth_labels:
    #         subset = result["dfs_reduced"]["synthetic"][result["dfs_reduced"]["synthetic"]['label'] == label]
    #         if 'x' in subset.columns and 'y' in subset.columns:
    #             synthetic_centers[label] = [subset['x'].mean(), subset['y'].mean()]
    #     callback = CustomJS(args=dict(source=source_table, line_source=line_source,
    #                               synthetic_centers=synthetic_centers,
    #                               real_centers=real_centers_js,
    #                               real_select=real_select),
    #     code="""
    #         var selected = source.selected.indices;
    #         if (selected.length > 0) {
    #             var idx = selected[0];
    #             var data = source.data;
    #             var synth_label = data['Synthetic'][idx];
    #             var real_label = real_select.value;
    #             var syn_coords = synthetic_centers[synth_label];
    #             var real_coords = real_centers[real_label];
    #             line_source.data = {'x': [syn_coords[0], real_coords[0]], 'y': [syn_coords[1], real_coords[1]]};
    #             line_source.change.emit();
    #         } else {
    #             line_source.data = {'x': [], 'y': []};
    #             line_source.change.emit();
    #         }
    #     """)
    #     source_table.selected.js_on_change('indices', callback)
    #     real_select.js_on_change('value', callback)
    
    #     reset_callback = CustomJS(args=dict(line_source=line_source),
    #     code="""
    #         line_source.data = {'x': [], 'y': []};
    #         line_source.change.emit();
    #     """)
    #     reset_button.js_on_event("button_click", reset_callback)
    #     layout = column(fig, result["scatter_fig"], column(real_select, reset_button, data_table))
    # else:
    #     layout = column(result["scatter_fig"], column(real_select, reset_button, data_table))
    
    # st.bokeh_chart(layout, use_container_width=True)

    buffer = io.BytesIO()
    df_table.to_excel(buffer, index=False)
    buffer.seek(0)
    
    st.download_button(
        label="Export Table",
        data=buffer,
        file_name=f"cluster_distances_{model_name}.xlsx",
        mime="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
        key=f"download_button_excel_{model_name}"
    )

    if reduction_method == "PCA":
        st.markdown("## PCA - Solo Muestras Reales")
        # -------------------------------------------------------------------------
        # 1. PCA sobre las muestras reales
        df_real_only = embeddings["real"].copy()
        reducer_real = Pipeline([
            ("pca", PCA(n_components=N_COMPONENTS)),
            ("rel_scaler", RelativeScaler())
        ])

        reduced_real = reducer_real.fit_transform(df_real_only[embedding_cols].values)

        
        # Agregar columnas PC1, PC2, … a df_real_only
        for i in range(reduced_real.shape[1]):
            df_real_only[f'PC{i+1}'] = reduced_real[:, i]
        
        explained_variance_real = reducer_real.named_steps["pca"].explained_variance_ratio_
        unique_labels_real = sorted(df_real_only['label'].unique().tolist())
        
        # Mapeo de colores para las muestras reales usando la paleta Reds9
        num_labels = len(unique_labels_real)
        if num_labels <= 9:
            red_palette = Reds9[:num_labels]
        else:
            red_palette = (Reds9 * ((num_labels // 9) + 1))[:num_labels]
        real_color_mapping = {label: red_palette[i] for i, label in enumerate(unique_labels_real)}
        
        # Mostrar tabla de Explained Variance Ratio
        st.subheader("PCA - Real: Explained Variance Ratio")
        component_names_real = [f"PC{i+1}" for i in range(len(explained_variance_real))]
        variance_df_real = pd.DataFrame({
            "Component": component_names_real,
            "Explained Variance": explained_variance_real
        })
        st.table(variance_df_real)
        
        # Mostrar los plots de loadings para cada componente
        # st.subheader("PCA - Real: Component Loadings")
        # st.markdown("### Pesos de las Componentes Principales (Loadings) - Conjunto Combinado")
        # for i, comp in enumerate(reducer_real.named_steps["pca"].components_):
        #     source = ColumnDataSource(data=dict(
        #         dimensions=embedding_cols,
        #         weight=comp
        #     ))
        #     p = figure(
        #         x_range=embedding_cols,
        #         title=f"Componente Principal {i+1}",
        #         plot_height=400,
        #         plot_width=600,
        #         toolbar_location="above",
        #         tools="pan,wheel_zoom,reset,save,hover",
        #         active_scroll="wheel_zoom"
        #     )
        #     p.background_fill_color = "white"
        #     p.xgrid.grid_line_color = None
        #     p.ygrid.grid_line_color = "gray"
        #     p.vbar(x='dimensions', top='weight', width=0.8, source=source,
        #         fill_color="#2b83ba", line_color="#2b83ba")
        #     p.xaxis.axis_label = "Dimensiones Originales"
        #     p.xaxis.major_label_text_font_size = '0pt'
        #     hover = p.select_one(HoverTool)
        #     hover.tooltips = [("Dimensión", "@dimensions"), ("Peso", "@weight")]
        #     st.bokeh_chart(p)
        
        # -------------------------------------------------------------------------
        # 2. Proyección de todos los subconjuntos usando los loadings de df_real (para PC completos)
        # Se proyectan real, synthetic y pretrained (si existen) y se agregan todas las PC's.
        df_all = {}
        # Real
        df_real_proj = embeddings["real"].copy()
        # proj_real = reducer_real.named_steps["pca"].transform(df_real_proj[embedding_cols].values)
        proj_real = reducer_real.transform(df_real_proj[embedding_cols].values)
        for i in range(proj_real.shape[1]):
            df_real_proj[f'PC{i+1}'] = proj_real[:, i]
        df_all["real"] = df_real_proj

        # Synthetic
        if "synthetic" in embeddings:
            df_synth_proj = embeddings["synthetic"].copy()
            proj_synth = reducer_real.transform(df_synth_proj[embedding_cols].values)
            for i in range(proj_synth.shape[1]):
                df_synth_proj[f'PC{i+1}'] = proj_synth[:, i]
            df_all["synthetic"] = df_synth_proj

        # Pretrained
        if "pretrained" in embeddings:
            df_pretr_proj = embeddings["pretrained"].copy()
            proj_pretr = reducer_real.transform(df_pretr_proj[embedding_cols].values)
            for i in range(proj_pretr.shape[1]):
                df_pretr_proj[f'PC{i+1}'] = proj_pretr[:, i]
            df_all["pretrained"] = df_pretr_proj

        # Para el plot global usaremos PC1 y PC2 (se asignan a 'x' y 'y')
        for key in df_all:
            df_all[key]["x"] = df_all[key]["PC1"]
            df_all[key]["y"] = df_all[key]["PC2"]

        # Construir los subconjuntos únicos para agrupar:
        unique_subsets = {}
        unique_subsets["real"] = sorted(df_all["real"]['label'].unique().tolist())
        if "synthetic" in df_all:
            unique_synth = {}
            for source in df_all["synthetic"]["source"].unique():
                unique_synth[source] = sorted(df_all["synthetic"][df_all["synthetic"]["source"] == source]['label'].unique().tolist())
            unique_subsets["synthetic"] = unique_synth
        else:
            unique_subsets["synthetic"] = {}
        if "pretrained" in df_all:
            unique_subsets["pretrained"] = sorted(df_all["pretrained"]['label'].unique().tolist())
        else:
            unique_subsets["pretrained"] = []
        
        # Obtener mapeo de colores para cada subconjunto (función definida externamente)
        color_maps = get_color_maps(unique_subsets)
        
        # Mapeo de marcadores para synthetic (por source)
        marker_mapping = {
            "es-digital-paragraph-degradation-seq": "x",
            "es-digital-line-degradation-seq": "cross",
            "es-digital-seq": "triangle",
            "es-digital-rotation-degradation-seq": "diamond",
            "es-digital-zoom-degradation-seq": "asterisk",
            "es-render-seq": "inverted_triangle"
        }
        
        # Plot global: se muestran real, synthetic y pretrained (según checkbox)
        st.subheader("PCA - Todos los subconjuntos proyectados (PC1 vs PC2)")
        fig_all = figure(
            title="PCA - Todos los subconjuntos proyectados",
            plot_width=600,
            plot_height=600,
            tools="pan,wheel_zoom,reset,save",
            active_scroll="wheel_zoom",
            background_fill_color="white",
            tooltips=TOOLTIPS
        )
        fig_all.xgrid.grid_line_color = None
        fig_all.ygrid.grid_line_color = "gray"
        
        # Plotear las muestras reales, agrupadas por label
        for label in unique_subsets["real"]:
            subset = df_all["real"][df_all["real"]['label'] == label]
            source = ColumnDataSource(data={
                'x': subset['x'],
                'y': subset['y'],
                'label': subset['label'],
                'img': subset['img']
            })
            fig_all.circle('x', 'y', size=10,
                        fill_color=color_maps["real"][label],
                        line_color=color_maps["real"][label],
                        legend_label=f"Real: {label}",
                        source=source)
        
        show_real_only = st.checkbox("Show only real samples", value=True, key=f"show_real_only_{model_name}")
        if not show_real_only:
            # Agregar synthetic
            if unique_subsets["synthetic"]:
                for source_name, labels in unique_subsets["synthetic"].items():
                    df_source = df_all["synthetic"][df_all["synthetic"]["source"] == source_name]
                    marker = marker_mapping.get(source_name, "square")
                    # Se usa el mapeo de colores para synthetic
                    color_val = color_maps["synthetic"][source_name]
                    renderers = add_synthetic_dataset_to_fig(
                        fig_all, df_source, labels,
                        marker=marker,
                        color_mapping=color_val,
                        group_label=source_name
                    )
            # Agregar pretrained
            if unique_subsets["pretrained"]:
                for label in unique_subsets["pretrained"]:
                    subset = df_all["pretrained"][df_all["pretrained"]['label'] == label]
                    source = ColumnDataSource(data={
                        'x': subset['x'],
                        'y': subset['y'],
                        'label': subset['label'],
                        'img': subset['img']
                    })
                    fig_all.triangle('x', 'y', size=10,
                                    fill_color=color_maps["pretrained"][label],
                                    line_color=color_maps["pretrained"][label],
                                    legend_label=f"Pretrained: {label}",
                                    source=source)
        
        show_legend_global = st.checkbox("Show Legend", value=False, key=f"legend_global_{model_name}")
        fig_all.legend.visible = show_legend_global
        fig_all.legend.location = "top_right"
        fig_all.match_aspect = True
        st.bokeh_chart(fig_all)
        
        # Calcular centroide y radio (usando solo las muestras reales)
        center_x = df_all["real"]['x'].mean()
        center_y = df_all["real"]['y'].mean()
        distances = np.sqrt((df_all["real"]['x'] - center_x)**2 + (df_all["real"]['y'] - center_y)**2)
        radius = distances.max()
        
        # Dibujar el centroide y la circunferencia
        centroid_glyph = fig_all.circle(
            x=center_x, y=center_y, size=15,
            fill_color="white", line_color="black",
            legend_label="Centroide",
            name="centroid"
        )
        circumference_glyph = fig_all.circle(
            x=center_x, y=center_y, radius=radius,
            fill_color=None, line_color="black",
            line_dash="dashed",
            legend_label="Circunferencia",
            name="circumference"
        )
        
        # Ajustar ejes y tooltips
        fig_all.xaxis.axis_label = "PC1"
        fig_all.yaxis.axis_label = "PC2"
        hover_all = fig_all.select_one(HoverTool)
        hover_all.renderers = [r for r in fig_all.renderers if r.name not in ["centroid", "circumference"]]
        
        st.write(f"El radio de la circunferencia (calculado a partir de las muestras reales) es: {radius:.4f}")
        
        # -------------------------------------------------------------------------
        # Calcular el rango global: recorrer todas las proyecciones de todos los subconjuntos
        all_vals = []
        for key in df_all:
            for comp in [f'PC{i+1}' for i in range(N_COMPONENTS)]:
                all_vals.append(df_all[key][comp])
        all_vals = pd.concat(all_vals)
        # Tomar el máximo valor absoluto de todas las proyecciones
        max_val = all_vals.abs().max()
        global_range = (-max_val, max_val)

        # 3. Scatter plots para cada combinación (vistas planta, alzado y perfil)
        st.subheader("Scatter Plots: Vistas de Componentes (Combinaciones)")
        pairs = list(itertools.combinations(range(N_COMPONENTS), 2))
        for (i, j) in pairs:
            x_comp = f'PC{i+1}'
            y_comp = f'PC{j+1}'
            
            st.markdown(f"### Scatter Plot: {x_comp} vs {y_comp}")
            # Usar el rango global para ambos ejes
            p = figure(
                title=f"{x_comp} vs {y_comp}",
                plot_width=700,
                plot_height=700,
                x_range=global_range,
                y_range=global_range,
                tools="pan,wheel_zoom,reset,save,hover",
                active_scroll="wheel_zoom",
                background_fill_color="white",
                tooltips=TOOLTIPS
            )
            # Etiquetas de ejes
            p.xaxis.axis_label = x_comp
            p.yaxis.axis_label = y_comp

            # Muestras reales: se usan directamente los valores de PC{i+1} y PC{j+1}
            for label in unique_subsets["real"]:
                subset = df_all["real"][df_all["real"]['label'] == label]
                source = ColumnDataSource(data={
                    'x': subset[x_comp],
                    'y': subset[y_comp],
                    'label': subset['label'],
                    'img': subset['img']
                })
                p.circle('x', 'y', size=10,
                        fill_color=color_maps["real"][label],
                        line_color=color_maps["real"][label],
                        legend_label=f"Real: {label}",
                        source=source)
            
            # Selector para incluir o no synthetic y pretrained en este gráfico
            show_pair_only_real = st.checkbox("Show only real samples", value=True, key=f"pair_show_real_{i}_{j}_{model_name}")
            if not show_pair_only_real:
                # Synthetic
                if "synthetic" in df_all:
                    for source_name, labels in unique_subsets["synthetic"].items():
                        # Obtener las filas de synthetic para ese source y asignar el rango adecuado
                        df_source = df_all["synthetic"][df_all["synthetic"]["source"] == source_name].copy()
                        df_source["x"] = df_source[x_comp]
                        df_source["y"] = df_source[y_comp]
                        marker = marker_mapping.get(source_name, "square")
                        renderers = add_synthetic_dataset_to_fig(
                            p, df_source, labels,
                            marker=marker,
                            color_mapping=color_maps["synthetic"][source_name],
                            group_label=source_name
                        )
                # Pretrained
                if "pretrained" in df_all:
                    for label in unique_subsets["pretrained"]:
                        subset = df_all["pretrained"][df_all["pretrained"]['label'] == label]
                        source = ColumnDataSource(data={
                            'x': subset[x_comp],
                            'y': subset[y_comp],
                            'label': subset['label'],
                            'img': subset['img']
                        })
                        p.triangle('x', 'y', size=10,
                                fill_color=color_maps["pretrained"][label],
                                line_color=color_maps["pretrained"][label],
                                legend_label=f"Pretrained: {label}",
                                source=source)
            show_legend_pair = st.checkbox("Show Legend", value=False, key=f"legend_pair_{i}_{j}_{model_name}")
            p.legend.visible = show_legend_pair
            st.bokeh_chart(p)

        
        # -------------------------------------------------------------------------
        # 4. Cálculo de distancias y scatter plot: Distance vs F1 (usando PC1 y PC2 globales)
        


        # Genera una paleta de 256 colores basada en RdYlGn11
        cmap = plt.get_cmap("RdYlGn")
        red_green_palette = [mcolors.rgb2hex(cmap(i)) for i in np.linspace(0, 1, 256)]



        # real_labels_new = sorted(df_all["real"]['label'].unique().tolist())
        # df_distances_new = compute_cluster_distances_synthetic_individual(
        #     df_all["synthetic"],
        #     df_all["real"],
        #     real_labels_new,
        #     metric="wasserstein",  # O la métrica que prefieras
        #     bins=20
        # )
        
        # global_distances_new = {}
        # for idx in df_distances_new.index:
        #     if idx.startswith("Global"):
        #         source_name = idx.split("(")[1].rstrip(")")
        #         global_distances_new[source_name] = df_distances_new.loc[idx].values
        
        # all_x_new = []
        # all_y_new = []
        # for source in df_f1.columns:
        #     if source in global_distances_new:
        #         x_vals = global_distances_new[source]
        #         y_vals = df_f1[source].values
        #         all_x_new.extend(x_vals)
        #         all_y_new.extend(y_vals)
        # all_x_arr_new = np.array(all_x_new).reshape(-1, 1)
        # all_y_arr_new = np.array(all_y_new)
        
        # model_global_new = LinearRegression().fit(all_x_arr_new, all_y_arr_new)
        # r2_new = model_global_new.score(all_x_arr_new, all_y_arr_new)
        # slope_new = model_global_new.coef_[0]
        # intercept_new = model_global_new.intercept_
        
        # scatter_fig_new = figure(
        #     width=600,
        #     height=600,
        #     tools="pan,wheel_zoom,reset,save,hover",
        #     active_scroll="wheel_zoom",
        #     title="Scatter Plot: Distance vs F1 (Nueva PCA)",
        #     background_fill_color="white",
        #     y_range=(0, 1)
        # )
        # scatter_fig_new.xgrid.grid_line_color = None
        # scatter_fig_new.ygrid.grid_line_color = "gray"
        # scatter_fig_new.match_aspect = True
        
        # source_colors = {
        #     "es-digital-paragraph-degradation-seq": "blue",
        #     "es-digital-line-degradation-seq": "green",
        #     "es-digital-seq": "red",
        #     "es-digital-zoom-degradation-seq": "orange",
        #     "es-digital-rotation-degradation-seq": "purple",
        #     "es-digital-rotation-zoom-degradation-seq": "brown",
        #     "es-render-seq": "cyan"
        # }
        
        # for source in df_f1.columns:
        #     if source in global_distances_new:
        #         x_vals = global_distances_new[source]
        #         y_vals = df_f1[source].values
        #         data = {"x": x_vals, "y": y_vals, "Fuente": [source]*len(x_vals)}
        #         cds = ColumnDataSource(data=data)
        #         scatter_fig_new.circle(
        #             'x', 'y', size=8, alpha=0.7, source=cds,
        #             fill_color=source_colors.get(source, "gray"),
        #             line_color=source_colors.get(source, "gray"),
        #             legend_label=source
        #         )
        
        # scatter_fig_new.xaxis.axis_label = "Distance (Global, por Colegio) - Nueva PCA"
        # scatter_fig_new.yaxis.axis_label = "F1 Score"
        # scatter_fig_new.legend.location = "top_right"
        # hover_tool_new = scatter_fig_new.select_one(HoverTool)
        # hover_tool_new.tooltips = [("Distance", "@x"), ("F1", "@y"), ("Subset", "@Fuente")]
        # x_line_new = np.linspace(all_x_arr_new.min(), all_x_arr_new.max(), 100)
        # y_line_new = model_global_new.predict(x_line_new.reshape(-1,1))
        # scatter_fig_new.line(x_line_new, y_line_new, line_width=2, line_color="black", legend_label="Global Regression")
        # st.bokeh_chart(scatter_fig_new)
        # st.write(f"Regresión global (Nueva PCA): R² = {r2_new:.4f}, Slope = {slope_new:.4f}, Intercept = {intercept_new:.4f}")
        
        # -------------------------------------------------------------------------
        # 5. BLOQUE: Heatmap de Características
        st.markdown("## Heatmap de Características")
        try:
            df_heat = pd.read_csv(f"data/heatmaps_{model_name.lower()}.csv")
        except Exception as e:
            st.error(f"Error al cargar heatmaps.csv: {e}")
            df_heat = None
        
        if df_heat is not None:
            if 'img' not in df_all["real"].columns:
                st.error("La columna 'img' no se encuentra en las muestras reales para hacer el merge con heatmaps.csv.")
            else:
                # Crear columna 'name' en las muestras reales (si aún no existe)
                df_all["real"]["name"] = df_all["real"]["img"].apply(
                    lambda x: x.split("/")[-1].replace(".png", "") if isinstance(x, str) else x
                )
                # Merge de las posiciones reales con el CSV de heatmaps (se usa el merge base)
                df_heatmap_base = pd.merge(df_all["real"], df_heat, on="name", how="inner")
                
                # Extraer opciones de feature (excluyendo 'name')
                feature_options = [col for col in df_heat.columns if col != "name"]
                selected_feature = st.selectbox("Select heatmap feature:", 
                                                options=feature_options, key=f"heatmap_{model_name}")
                select_extra_dataset_hm = st.selectbox("Select a dataset:", 
                                                options=model_options_with_default, key=f"heatmap_extra_dataset_{model_name}")
                
                # Definir un rango fijo para los ejes (por ejemplo, de -1 a 1) y rejilla
                x_min, x_max = -1, 1
                y_min, y_max = -1, 1
                grid_size = 50
                x_bins = np.linspace(x_min, x_max, grid_size + 1)
                y_bins = np.linspace(y_min, y_max, grid_size + 1)
                
                # Listas para almacenar las figuras de heatmap y sus nombres
                heatmap_figures = []
                heatmap_names = []

                # Generar heatmaps para cada combinación de componentes
                pairs = list(itertools.combinations(range(N_COMPONENTS), 2))
                for (i, j) in pairs:
                    x_comp = f'PC{i+1}'
                    y_comp = f'PC{j+1}'
                    st.markdown(f"### Heatmap: {x_comp} vs {y_comp}")
                    
                    # Crear un DataFrame de heatmap para la combinación actual a partir del merge base
                    df_heatmap = df_heatmap_base.copy()
                    df_heatmap["x"] = df_heatmap[x_comp]
                    df_heatmap["y"] = df_heatmap[y_comp]
                    
                    # Si la feature seleccionada no es numérica, convertir a códigos y guardar la correspondencia
                    cat_mapping = None
                    if df_heatmap[selected_feature].dtype == bool or not pd.api.types.is_numeric_dtype(df_heatmap[selected_feature]):
                        cat = df_heatmap[selected_feature].astype('category')
                        cat_mapping = list(cat.cat.categories)
                        df_heatmap[selected_feature] = cat.cat.codes
                    
                    # Calcular la estadística binned (por ejemplo, la media) en la rejilla
                    try:
                        heat_stat, x_edges, y_edges, binnumber = binned_statistic_2d(
                            df_heatmap['x'], df_heatmap['y'], df_heatmap[selected_feature],
                            statistic='mean', bins=[x_bins, y_bins]
                        )
                    except TypeError:
                        cat = df_heatmap[selected_feature].astype('category')
                        cat_mapping = list(cat.cat.categories)
                        df_heatmap[selected_feature] = cat.cat.codes
                        heat_stat, x_edges, y_edges, binnumber = binned_statistic_2d(
                            df_heatmap['x'], df_heatmap['y'], df_heatmap[selected_feature],
                            statistic='mean', bins=[x_bins, y_bins]
                        )
                    # Transponer la matriz para alinear correctamente los ejes
                    # Transponer y limpiar valores inválidos (NaN, inf, -inf)
                    heatmap_data = heat_stat.T
                    # heatmap_data = np.nan_to_num(heat_stat.T, nan=0.0, posinf=0.0, neginf=0.0)
                    
                    # Definir el color mapper
                    if selected_feature in model_options:
                        color_mapper = LinearColorMapper(
                            palette=red_green_palette,
                            low=0,
                            high=1,
                            nan_color='rgba(0, 0, 0, 0)'
                        )
                    else:
                        color_mapper = LinearColorMapper(
                            palette="Viridis256",
                            low=np.nanmin(heatmap_data),
                            high=np.nanmax(heatmap_data),
                            nan_color='rgba(0, 0, 0, 0)'
                        )
                    
                    # Crear la figura para el heatmap con la misma escala para x e y
                    heatmap_fig = figure(title=f"Heatmap de '{selected_feature}' ({x_comp} vs {y_comp})",
                                        x_range=(x_min, x_max), y_range=(y_min, y_max),
                                        width=600, height=600,
                                        tools="pan,wheel_zoom,reset,save", active_scroll="wheel_zoom", tooltips=TOOLTIPS,
                                        sizing_mode="fixed")
                    heatmap_fig.match_aspect = True
                    
                    # Asignar etiquetas a los ejes
                    heatmap_fig.xaxis.axis_label = x_comp
                    heatmap_fig.yaxis.axis_label = y_comp

                    heatmap_fig.background_fill_color = "white"
                    heatmap_fig.border_fill_color = "white"

                    # Dibujar la imagen del heatmap
                    heatmap_fig.image(image=[heatmap_data], x=x_min, y=y_min,
                                    dw=x_max - x_min, dh=y_max - y_min,
                                    color_mapper=color_mapper)
                    
                    # Agregar la barra de color
                    color_bar = ColorBar(color_mapper=color_mapper, location=(0, 0))
                    if cat_mapping is not None:
                        ticks = list(range(len(cat_mapping)))
                        color_bar.ticker = FixedTicker(ticks=ticks)
                        categories_json = json.dumps(cat_mapping)
                        color_bar.formatter = FuncTickFormatter(code=f"""
                            var categories = {categories_json};
                            var index = Math.round(tick);
                            if(index >= 0 && index < categories.length) {{
                                return categories[index];
                            }} else {{
                                return "";
                            }}
                        """)
                    heatmap_fig.add_layout(color_bar, 'right')
                    
                    # Agregar renderer invisible para tooltips
                    source_points = ColumnDataSource(data={
                        'x': df_heatmap['x'],
                        'y': df_heatmap['y'],
                        'img': df_heatmap['img'],
                        'label': df_heatmap['name']
                    })
                    invisible_renderer = heatmap_fig.circle('x', 'y', size=10, source=source_points, fill_alpha=0, line_alpha=0.5)
                    # school = "patria"
                    if select_extra_dataset_hm != "-":
                        df_extra = df_all["synthetic"][df_all["synthetic"]["source"] == select_extra_dataset_hm].copy()
                        df_extra["x"] = df_extra[x_comp]
                        df_extra["y"] = df_extra[y_comp]
                        if 'name' not in df_extra.columns:
                            df_extra["name"] = df_extra["img"].apply(lambda x: x.split("/")[-1].replace(".png", "") if isinstance(x, str) else x)
                        # mask = df_extra["name"].str.contains(school, case=False, na=False)
                        # df_extra = df_extra[mask].copy()
                        source_extra_points = ColumnDataSource(data={
                            'x': df_extra['x'],
                            'y': df_extra['y'],
                            'img': df_extra['img'],
                            'label': df_extra['name']
                        })
                        extra_renderer = heatmap_fig.circle('x', 'y', size=5, source=source_extra_points, fill_alpha=0, line_alpha=0.5, color="purple")
                    
                    hover_tool_points = HoverTool(renderers=[invisible_renderer], tooltips=TOOLTIPS)
                    heatmap_fig.add_tools(hover_tool_points)
                    
                    # Mostrar el heatmap en la app
                    st.bokeh_chart(heatmap_fig)

        # Botón para descargar df_all (Embeddings in PCA Space)
        if st.button("Download Embeddings in PCA Space", key=f"click_download_pca_coordinates_{model_name}"):
            # Crear un nuevo diccionario para almacenar solo las columnas que comienzan con "PC" o "name"
            df_all_pca = {}
            for key, df in df_all.items():
                # Si es el conjunto sintético, separamos cada subset según la columna "source"
                if key == "synthetic":
                    for source in df["source"].unique():
                        df_subset = df[df["source"] == source].copy()
                        # Asegurarse de que exista la columna "name" (como se hace en el snippet de heatmaps)
                        if "img" in df_subset.columns and "name" not in df_subset.columns:
                            df_subset["name"] = df_subset["img"].apply(lambda x: x.split("/")[-1].replace(".png", "") if isinstance(x, str) else x)
                        pca_cols = [col for col in df_subset.columns if col.startswith("PC") or col == "name"]
                        # Usar un nombre de hoja que identifique que es sintético y el source correspondiente
                        sheet_name = f"synthetic_{source}"
                        df_all_pca[sheet_name] = df_subset[pca_cols].copy()
                else:
                    # Para "real" y otros (como "pretrained"), se guardan en una sola hoja
                    pca_cols = [col for col in df.columns if col.startswith("PC") or col == "name"]
                    df_all_pca[key] = df[pca_cols].copy()

            # Crear un buffer en memoria para el archivo Excel
            excel_buffer = io.BytesIO()
            # Escribir cada DataFrame en una hoja separada usando ExcelWriter
            with pd.ExcelWriter(excel_buffer, engine='openpyxl') as writer:
                for key_name, df in df_all_pca.items():
                    df.to_excel(writer, sheet_name=key_name, index=False)
            excel_buffer.seek(0)
            
            st.download_button(
                label="Download Embeddings in PCA Space",
                data=excel_buffer,
                file_name=f"df_all_pca_{model_name.lower()}.xlsx",
                mime="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
                key=f"download_pca_coordinates_{model_name}"
            )
            
    elif reduction_method == "t-SNE":
        st.markdown("## t-SNE - Solo Muestras Reales")
        # -------------------------------------------------------------------------
        # 1. t-SNE sobre las muestras reales
        df_real_only = embeddings["real"].copy()
        
        reducer_real = TSNE(n_components=2, perplexity=30, random_state=42)
        reduced_real = reducer_real.fit_transform(df_real_only[embedding_cols].values)
        
        # Agregar columnas TSNE1, TSNE2
        df_real_only['TSNE1'] = reduced_real[:, 0]
        df_real_only['TSNE2'] = reduced_real[:, 1]

        # Construir df_all de forma consistente con PCA
        df_all = {}
        df_all["real"] = df_real_only.copy()
        if "synthetic" in embeddings:
            df_all["synthetic"] = embeddings["synthetic"].copy()
        if "pretrained" in embeddings:
            df_all["pretrained"] = embeddings["pretrained"].copy()

        unique_labels_real = sorted(df_real_only['label'].unique().tolist())
        
        # Mapeo de colores para las muestras reales usando la paleta Reds9
        num_labels = len(unique_labels_real)
        if num_labels <= 9:
            red_palette = Reds9[:num_labels]
        else:
            red_palette = (Reds9 * ((num_labels // 9) + 1))[:num_labels]
        real_color_mapping = {label: red_palette[i] for i, label in enumerate(unique_labels_real)}

        # -------------------------------------------------------------------------
        # Crear plot interactivo con Bokeh
        st.subheader("t-SNE - Real: Visualización Interactiva")

        source = ColumnDataSource(df_real_only)

        hover = HoverTool(tooltips=[
            ("Index", "$index"),
            ("Label", "@label"),
            ("TSNE1", "@TSNE1"),
            ("TSNE2", "@TSNE2")
        ])

        p = figure(
            width=800,
            height=600,
            title="t-SNE sobre muestras reales",
            tools=["pan", "wheel_zoom", "box_zoom", "reset", hover]
        )

        for label in unique_labels_real:
            subset = df_real_only[df_real_only['label'] == label]
            p.scatter(
                x=subset["TSNE1"],
                y=subset["TSNE2"],
                size=8,
                color=real_color_mapping[label],
                alpha=0.7,
                legend_label=str(label)
            )

        p.legend.title = "Label"
        p.legend.location = "top_right"
        p.xaxis.axis_label = "t-SNE 1"
        p.yaxis.axis_label = "t-SNE 2"

        st.bokeh_chart(p, use_container_width=True)

        # -------------------------------------------------------------------------
        # BLOQUE: Heatmap de Características (adaptado a TSNE1 y TSNE2)
        st.markdown("## Heatmap de Características")
        try:
            df_heat = pd.read_csv(f"data/heatmaps_{model_name.lower()}.csv")
        except Exception as e:
            st.error(f"Error al cargar heatmaps.csv: {e}")
            df_heat = None
        
        if df_heat is not None:
            if 'img' not in df_all["real"].columns:
                st.error("La columna 'img' no se encuentra en las muestras reales para hacer el merge con heatmaps.csv.")
            else:
                # Crear columna 'name' en las muestras reales (si aún no existe)
                df_all["real"]["name"] = df_all["real"]["img"].apply(
                    lambda x: x.split("/")[-1].replace(".png", "") if isinstance(x, str) else x
                )
                # Merge con heatmaps.csv
                df_heatmap_base = pd.merge(df_all["real"], df_heat, on="name", how="inner")
                
                # Opciones de features
                feature_options = [col for col in df_heat.columns if col != "name"]
                selected_feature = st.selectbox("Select heatmap feature:", 
                                                options=feature_options, key=f"heatmap_{model_name}")
                select_extra_dataset_hm = st.selectbox("Select a dataset:", 
                                                options=model_options_with_default, key=f"heatmap_extra_dataset_{model_name}")        
                
                # Solo una combinación: TSNE1 vs TSNE2
                x_comp, y_comp = "TSNE1", "TSNE2"
                st.markdown(f"### Heatmap: {x_comp} vs {y_comp}")

                # Rango real de los datos
                df_heatmap = df_heatmap_base.copy()
                # Rango real de los datos
                x_min, x_max = df_heatmap[x_comp].min(), df_heatmap[x_comp].max()
                y_min, y_max = df_heatmap[y_comp].min(), df_heatmap[y_comp].max()

                # Opcional: ampliar un poco
                padding = 0.05
                x_min, x_max = x_min - padding, x_max + padding
                y_min, y_max = y_min - padding, y_max + padding

                # Definir rejilla
                grid_size = st.slider("Grid size (resolución del heatmap)", min_value=10, max_value=100, value=40, step=5, key=f"grid_size_{model_name}")
                x_bins = np.linspace(x_min, x_max, grid_size + 1)
                y_bins = np.linspace(y_min, y_max, grid_size + 1)

                # Usar los mismos valores para figure y image
                x_range, y_range = (x_min, x_max), (y_min, y_max)
                
                
                df_heatmap["x"] = df_heatmap[x_comp]
                df_heatmap["y"] = df_heatmap[y_comp]
                df_heatmap['x'] = np.nan_to_num(df_heatmap['x'], nan=0.0, posinf=0.0, neginf=0.0)
                df_heatmap['y'] = np.nan_to_num(df_heatmap['y'], nan=0.0, posinf=0.0, neginf=0.0)
                
                # Si la feature seleccionada no es numérica, convertir
                cat_mapping = None
                if df_heatmap[selected_feature].dtype == bool or not pd.api.types.is_numeric_dtype(df_heatmap[selected_feature]):
                    cat = df_heatmap[selected_feature].astype('category')
                    cat_mapping = list(cat.cat.categories)
                    df_heatmap[selected_feature] = cat.cat.codes
                
                # Calcular estadística binned
                heat_stat, x_edges, y_edges, binnumber = binned_statistic_2d(
                    df_heatmap['x'], df_heatmap['y'], df_heatmap[selected_feature],
                    statistic='mean', bins=[x_bins, y_bins]
                )

                # Mantener NaN para celdas vacías
                heatmap_data = heat_stat.T

                cmap = plt.get_cmap("RdYlGn")
                red_green_palette = [mcolors.rgb2hex(cmap(i)) for i in np.linspace(0, 1, 256)]

                # Color mapper
                if selected_feature in model_options:
                    color_mapper = LinearColorMapper(
                        palette=red_green_palette,
                        low=0,
                        high=1,
                        nan_color="white"
                    )
                else:
                    color_mapper = LinearColorMapper(
                        palette="Viridis256",
                        low=np.nanmin(heatmap_data),
                        high=np.nanmax(heatmap_data),
                        nan_color="white"
                    )
                

                # Figura heatmap
                heatmap_fig = figure(title=f"Heatmap de '{selected_feature}' ({x_comp} vs {y_comp})",
                                    x_range=x_range, y_range=y_range,
                                    width=600, height=600,
                                    tools="pan,wheel_zoom,reset,save", active_scroll="wheel_zoom", tooltips=TOOLTIPS,
                                    sizing_mode="fixed")
                heatmap_fig.match_aspect = True
                heatmap_fig.xaxis.axis_label = x_comp
                heatmap_fig.yaxis.axis_label = y_comp

                heatmap_fig.image(image=[heatmap_data], x=x_min, y=y_min,
                                dw=x_max - x_min, dh=y_max - y_min,
                                color_mapper=color_mapper)
                
                # Barra de color
                color_bar = ColorBar(color_mapper=color_mapper, location=(0, 0))
                if cat_mapping is not None:
                    ticks = list(range(len(cat_mapping)))
                    color_bar.ticker = FixedTicker(ticks=ticks)
                    categories_json = json.dumps(cat_mapping)
                    color_bar.formatter = FuncTickFormatter(code=f"""
                        var categories = {categories_json};
                        var index = Math.round(tick);
                        if(index >= 0 && index < categories.length) {{
                            return categories[index];
                        }} else {{
                            return "";
                        }}
                    """)
                heatmap_fig.add_layout(color_bar, 'right')

                heatmap_fig.background_fill_color = "white"
                heatmap_fig.border_fill_color = "white"
                
                # Tooltips
                source_points = ColumnDataSource(data={
                    'x': df_heatmap['x'],
                    'y': df_heatmap['y'],
                    'img': df_heatmap['img'],
                    'label': df_heatmap['name']
                })
                invisible_renderer = heatmap_fig.circle('x', 'y', size=10, source=source_points, fill_alpha=0, line_alpha=0.5)
                
                # if select_extra_dataset_hm != "-":
                #     df_extra = df_all["synthetic"][df_all["synthetic"]["source"] == select_extra_dataset_hm].copy()
                #     df_extra["x"] = df_extra[x_comp]
                #     df_extra["y"] = df_extra[y_comp]
                #     if 'name' not in df_extra.columns:
                #         df_extra["name"] = df_extra["img"].apply(lambda x: x.split("/")[-1].replace(".png", "") if isinstance(x, str) else x)
                #     source_extra_points = ColumnDataSource(data={
                #         'x': df_extra['x'],
                #         'y': df_extra['y'],
                #         'img': df_extra['img'],
                #         'label': df_extra['name']
                #     })
                #     heatmap_fig.circle('x', 'y', size=5, source=source_extra_points, fill_alpha=0, line_alpha=0.5, color="purple")
                
                hover_tool_points = HoverTool(renderers=[invisible_renderer], tooltips=TOOLTIPS)
                heatmap_fig.add_tools(hover_tool_points)
                
                st.bokeh_chart(heatmap_fig)

        # -------------------------------------------------------------------------
        # Botón para descargar df_all (Embeddings en t-SNE Space)
        if st.button("Download Embeddings in t-SNE Space", key=f"click_download_tsne_coordinates_{model_name}"):
            df_all_tsne = {}
            for key, df in df_all.items():
                if key == "synthetic":
                    for source in df["source"].unique():
                        df_subset = df[df["source"] == source].copy()
                        if "img" in df_subset.columns and "name" not in df_subset.columns:
                            df_subset["name"] = df_subset["img"].apply(lambda x: x.split("/")[-1].replace(".png", "") if isinstance(x, str) else x)
                        tsne_cols = [col for col in df_subset.columns if col.startswith("TSNE") or col == "name"]
                        sheet_name = f"synthetic_{source}"
                        df_all_tsne[sheet_name] = df_subset[tsne_cols].copy()
                else:
                    tsne_cols = [col for col in df.columns if col.startswith("TSNE") or col == "name"]
                    df_all_tsne[key] = df[tsne_cols].copy()

            excel_buffer = io.BytesIO()
            with pd.ExcelWriter(excel_buffer, engine='openpyxl') as writer:
                for key_name, df in df_all_tsne.items():
                    df.to_excel(writer, sheet_name=key_name, index=False)
            excel_buffer.seek(0)
            
            st.download_button(
                label="Download Embeddings in t-SNE Space",
                data=excel_buffer,
                file_name=f"df_all_tsne_{model_name.lower()}.xlsx",
                mime="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
                key=f"download_tsne_coordinates_{model_name}"
            )


def main():
    config_style()
    tabs = st.tabs(["Donut", "Idefics2", "Idefics2-patient", "Paligemma", "Llava"])
    with tabs[0]:
        st.markdown('<h2 class="sub-title">Donut 🤗</h2>', unsafe_allow_html=True)
        run_model("Donut")
    with tabs[1]:
        st.markdown('<h2 class="sub-title">Idefics2 🤗</h2>', unsafe_allow_html=True)
        run_model("Idefics2")
    with tabs[2]:
        st.markdown('<h2 class="sub-title">Idefics2-patient 🤗</h2>', unsafe_allow_html=True)
        run_model("Idefics2-patient")
    with tabs[3]:
        st.markdown('<h2 class="sub-title">Paligemma 🤗</h2>', unsafe_allow_html=True)
        run_model("Paligemma")
    with tabs[4]:
        st.markdown('<h2 class="sub-title">Llava 🤗</h2>', unsafe_allow_html=True)
        run_model("Llava")

if __name__ == "__main__":
    model_options = [
            "es-digital-paragraph-degradation-seq", 
            "es-digital-line-degradation-seq", 
            "es-digital-seq",
            "es-digital-noisy-degradation-seq",
            "es-digital-seq_es-digital-seq_britanico",
            "es-digital-seq_retamar_train-asc-synth",
            "es-digital-seq_britanico_es-digital-seq_aletamar",
            "es-digital-seq_retamar_train-asc-synth_es-digital-seq_britanico",
            "es-digital-seq_retamar_train-asc-synth_es-digital-seq_britanico_v2",
            "es-digital-seq_es-digital-seq_aletamar",
            "retamar_train-asc-synth_es-digital-seq_britanico",
            "es-digital-rotation-degradation-seq", 
            "es-digital-zoom-degradation-seq", 
            "es-render-seq",
            "es-digital-seq_filtered_deus",
            "es-digital-seq_filtered_liceo",
            "es-digital-seq_filtered_lusitano",
            "es-digital-seq_filtered_monterraso",
            "es-digital-seq_filtered_patria",
            "es-digital-zoom-0.25-degradation-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "es-digital-zoom-0.5-degradation-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "es-digital-zoom-0.625-degradation-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "es-digital-zoom-0.75-degradation-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "es-digital-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "es-digital-zoom-degradation-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "es-render-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder-017",
            "es-render-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder-016",
            "es-digital-rotation-degradation-seq_filtered_deus-liceo-lusitano-monterraso-patria-frozen-encoder",
            "retamar_train-asc-synth_filtered_retamar_train-frozen-encoder",
            "britanico-retamar_train-asc-synth_filtered_retamar_train-frozen-encoder",
            "combination-es-digital-seq",
            "combination-es-render-seq",
            "retamar_train-asc-synth_es-digital-seq_britanico",

    ]
    
    model_options_with_default = [""]
    model_options_with_default.extend(model_options)
    
    main()