streamlit.py

"""

t-SNE Explorer - Streamlit Application with MCP Server

A transparent t-SNE implementation with synthetic data generation and file upload support



This version provides:

1. Streamlit UI for web browser access

2. MCP server endpoints for android.py to connect remotely



Android clients can connect to this Hugging Face deployment as their MCP computation server.

"""

import os
import sys
import json
import warnings
import numpy as np
import pandas as pd
import streamlit as st
import plotly.graph_objects as go
from io import BytesIO
from PIL import Image
from pathlib import Path

# Suppress warnings
warnings.filterwarnings('ignore')
os.environ['PYTHONWARNINGS'] = 'ignore'

try:
    from sklearn.manifold import TSNE as SklearnTSNE
    from sklearn.datasets import fetch_openml
except Exception:
    SklearnTSNE = None
    fetch_openml = None


# ==================== Styling ====================

def inject_custom_css():
    """Inject custom CSS from web/style.css to match the original design"""
    st.markdown("""

    <style>

    /* Import styling from web folder */

    :root {

        --primary: #667eea;

        --primary-dark: #5568d3;

        --secondary: #764ba2;

        --success: #10b981;

        --warning: #f59e0b;

        --error: #ef4444;

    }



    /* Main app styling */

    .stApp {

        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);

    }



    /* Sidebar styling */

    .css-1d391kg {

        background-color: #f8f9fa;

    }



    /* Headers */

    h1 {

        color: white;

        font-weight: 800;

        text-shadow: 0 2px 10px rgba(0, 0, 0, 0.2);

    }



    h2, h3 {

        color: #667eea;

        font-weight: 700;

    }



    /* Buttons */

    .stButton > button {

        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);

        color: white;

        border: none;

        border-radius: 8px;

        padding: 12px 24px;

        font-weight: 600;

        transition: all 0.3s ease;

    }



    .stButton > button:hover {

        transform: translateY(-2px);

        box-shadow: 0 4px 12px rgba(102, 126, 234, 0.4);

    }



    /* Info boxes */

    .stAlert {

        border-radius: 8px;

        border-left: 4px solid #667eea;

    }



    /* Dataframes */

    .dataframe {

        border-radius: 8px;

        overflow: hidden;

    }



    /* Cards */

    .element-container {

        background: white;

        border-radius: 12px;

        padding: 10px;

        margin-bottom: 10px;

    }



    /* Progress bar */

    .stProgress > div > div {

        background: linear-gradient(90deg, #667eea 0%, #764ba2 100%);

    }

    </style>

    """, unsafe_allow_html=True)


# ==================== TSNEExplorer Backend Class (MCP-Ready) ====================

class TSNEExplorer:
    """

    Backend API for t-SNE computations.

    This class is MCP-ready - all methods return JSON-serializable data

    and can be called directly (Streamlit) or via API endpoints (future Android app).

    """

    def __init__(self):
        pass

    # ==================== Synthetic Data Generation ====================

    def generate_simplex_points(self, n, d, k, seed=42):
        """Generate n points in d dimensions with k distinct distance types"""
        np.random.seed(seed)

        # Validate inputs
        max_distances = (n * (n - 1)) // 2
        if k > max_distances:
            return {
                'success': False,
                'error': f'Cannot create {k} distinct distances with only {n} points. '
                        f'Maximum possible is {max_distances} distinct distances.'
            }

        if k < 1:
            return {
                'success': False,
                'error': f'k must be at least 1 (you specified k={k}).'
            }

        # Special case: k=1
        if k == 1:
            if n > d + 1:
                return {
                    'success': False,
                    'error': f'For k=1 (equidistant points), maximum n is {d+1} in {d}D.'
                }
            X = self._generate_regular_simplex(n, d)
        else:
            X = self._generate_k_distance_set(n, d, k, seed)

        # Compute pairwise distances
        distances = self._compute_pairwise_distances(X)
        unique_distances = np.unique(np.round(distances[distances > 0], decimals=6))

        return {
            'success': True,
            'points': X.tolist(),
            'n': n,
            'd': d,
            'k': k,
            'actual_k': len(unique_distances),
            'unique_distances': unique_distances.tolist(),
            'distances_min': float(np.min(distances[distances > 0])) if n > 1 else 0,
            'distances_mean': float(np.mean(distances[distances > 0])) if n > 1 else 0,
            'distances_max': float(np.max(distances)),
        }

    def _generate_regular_simplex(self, n, d):
        """Generate regular n-simplex with equal pairwise distances"""
        if n == 1:
            return np.zeros((1, d))

        if n == 2:
            X = np.zeros((2, d))
            X[0, 0] = -0.5
            X[1, 0] = 0.5
            return X

        vertices = np.eye(n)
        vertices = vertices - np.mean(vertices, axis=0)
        vertices = vertices / np.sqrt(2)

        if d >= n - 1:
            X = vertices[:, :min(d, n)]
            if d > n:
                X = np.pad(X, ((0, 0), (0, d - n)), 'constant')
        else:
            X = vertices[:, :d]

        return X

    def _generate_k_distance_set(self, n, d, k, seed):
        """Generate points aiming for k distinct pairwise distances"""
        np.random.seed(seed)

        if n <= 0 or d <= 0:
            return np.zeros((0, max(d, 0)))

        if n == 1:
            return np.zeros((1, d))

        # Exact k=2 constructions
        if k == 2:
            if d >= 2 and n == 5:
                return self._regular_ngon(n=5, d=d)
            if n <= 2 * d:
                return self._cross_polytope(n=n, d=d)
            return self._optimize_k_distance_set(n=n, d=d, k=k, seed=seed)

        # Exact k>=3 constructions
        if k >= 3 and d >= k and k <= 12 and n <= (2 ** k):
            return self._k_cube_k_distance_set(n=n, d=d, k=k)

        if k == 3 and d >= 2 and n in (6, 7):
            return self._regular_ngon(n=n, d=d)

        return self._optimize_k_distance_set(n=n, d=d, k=k, seed=seed)

    def _regular_ngon(self, n, d):
        """Regular n-gon in 2D"""
        X = np.zeros((n, d))
        if d < 2:
            return X
        angles = np.linspace(0, 2 * np.pi, n + 1)[:-1]
        X[:, 0] = np.cos(angles)
        X[:, 1] = np.sin(angles)
        return X

    def _cross_polytope(self, n, d):
        """Cross polytope vertices"""
        X = np.zeros((n, d))
        if n == 1:
            return X
        point_idx = 0
        for i in range(d):
            if point_idx >= n:
                break
            X[point_idx, i] = 1.0
            point_idx += 1
            if point_idx >= n:
                break
            X[point_idx, i] = -1.0
            point_idx += 1
        return X

    def _k_cube_k_distance_set(self, n, d, k):
        """k-dimensional hypercube vertices"""
        vertices = []
        seen = set()

        origin = tuple([0] * k)
        vertices.append(origin)
        seen.add(origin)

        for weight in range(1, k + 1):
            if len(vertices) >= n:
                break
            v = tuple([1] * weight + [0] * (k - weight))
            if v not in seen:
                vertices.append(v)
                seen.add(v)

        for mask in range(1, 2 ** k):
            if len(vertices) >= n:
                break
            v = tuple((mask >> bit) & 1 for bit in range(k))
            if v in seen:
                continue
            vertices.append(v)
            seen.add(v)

        Xk = np.array(vertices[:n], dtype=float)
        X = np.zeros((n, d), dtype=float)
        X[:, :k] = Xk
        X = X - X.mean(axis=0, keepdims=True)
        return X

    def _optimize_k_distance_set(self, n, d, k, seed, n_iter=2000, lr=0.02):
        """Heuristic optimization for k distances"""
        rng = np.random.default_rng(seed)
        X = rng.standard_normal((n, d)) * 0.1

        if n < 2:
            return X

        D0 = self._compute_pairwise_distances(X)
        upper = D0[np.triu_indices(n, k=1)]
        if upper.size == 0:
            return X

        r_min = float(np.percentile(upper, 10))
        r_max = float(np.percentile(upper, 90))
        if r_max <= 1e-8:
            r_max = 1.0
        radii = np.linspace(max(r_min, 1e-3), max(r_max, 1e-3), k)

        use_minibatch = n > 150
        batch_size = min(5000, (n * (n - 1)) // 2) if use_minibatch else 0
        ema = 0.15

        for _ in range(n_iter):
            if use_minibatch:
                ii = rng.integers(0, n, size=batch_size)
                jj = rng.integers(0, n, size=batch_size)
                mask = ii != jj
                if not np.any(mask):
                    continue
                ii = ii[mask]
                jj = jj[mask]

                diff = X[ii] - X[jj]
                dist = np.sqrt(np.sum(diff * diff, axis=1))
                dist_safe = np.maximum(dist, 1e-12)

                assign = np.argmin(np.abs(dist[:, np.newaxis] - radii[np.newaxis, :]), axis=1)
                target = radii[assign]

                for m in range(k):
                    m_mask = assign == m
                    if np.any(m_mask):
                        radii[m] = (1 - ema) * radii[m] + ema * float(np.mean(dist[m_mask]))

                err = dist_safe - target
                coef = (2.0 * err / dist_safe)[:, np.newaxis]
                grad_pairs = coef * diff

                grad = np.zeros_like(X)
                np.add.at(grad, ii, grad_pairs)
                np.add.at(grad, jj, -grad_pairs)
            else:
                D = self._compute_pairwise_distances(X)
                iu, ju = np.triu_indices(n, k=1)
                dist = D[iu, ju]
                dist_safe = np.maximum(dist, 1e-12)

                assign = np.argmin(np.abs(dist[:, np.newaxis] - radii[np.newaxis, :]), axis=1)
                target = radii[assign]

                for m in range(k):
                    m_mask = assign == m
                    if np.any(m_mask):
                        radii[m] = float(np.mean(dist[m_mask]))

                err = dist_safe - target
                coef = (2.0 * err / dist_safe)[:, np.newaxis]
                diff = X[iu] - X[ju]
                grad_pairs = coef * diff

                grad = np.zeros_like(X)
                np.add.at(grad, iu, grad_pairs)
                np.add.at(grad, ju, -grad_pairs)

            grad += 1e-3 * X
            X = X - lr * grad
            X = X - X.mean(axis=0, keepdims=True)

        return X

    def _compute_pairwise_distances(self, X):
        """Compute pairwise Euclidean distances"""
        n = X.shape[0]
        distances = np.zeros((n, n))
        for i in range(n):
            for j in range(i+1, n):
                dist = np.linalg.norm(X[i] - X[j])
                distances[i, j] = dist
                distances[j, i] = dist
        return distances

    # ==================== MNIST Dataset ====================

    def load_mnist(self, max_samples=1000, subset='train'):
        """Load MNIST dataset"""
        try:
            if fetch_openml is None:
                return {'success': False, 'error': 'scikit-learn not available'}

            mnist = fetch_openml('mnist_784', version=1, as_frame=False, parser='auto')

            all_images = np.array(mnist.data, dtype=np.float32)

            if isinstance(mnist.target[0], str):
                all_labels = np.array([int(label) for label in mnist.target], dtype=np.int32)
            else:
                all_labels = np.array(mnist.target, dtype=np.int32)

            if subset == 'train':
                images_flat = all_images[:60000]
                labels = all_labels[:60000]
            else:
                images_flat = all_images[60000:]
                labels = all_labels[60000:]

            if max_samples > 0 and max_samples < len(images_flat):
                images_flat = images_flat[:max_samples]
                labels = labels[:max_samples]

            X = images_flat / 255.0

            return {
                'success': True,
                'X': X,
                'labels': labels,
                'count': len(images_flat),
                'shape': X.shape,
                'message': f'Loaded {len(images_flat)} MNIST {subset} samples'
            }

        except Exception as e:
            return {'success': False, 'error': str(e)}

    # ==================== t-SNE Implementation ====================

    def run_tsne(self, X, perplexity=30, learning_rate=200, n_iter=1000,

                 early_exaggeration=12, momentum=0.8, seed=42, progress_callback=None):
        """Run t-SNE with transparent internals"""
        try:
            n, d = X.shape

            if n > 1000:
                return {'success': False, 'error': f'Dataset too large ({n} points). Please use n <= 1000.'}

            # Initialize Y
            np.random.seed(seed)
            Y = np.random.randn(n, 2) * 0.0001

            # Compute P
            if progress_callback:
                progress_callback(0, 'Computing P matrix...')
            P = self._compute_P(X, perplexity)

            # Optimize
            if progress_callback:
                progress_callback(0, 'Starting t-SNE optimization...')
            Y, Q, C_history = self._optimize_tsne(
                P, Y, learning_rate, n_iter, early_exaggeration, momentum, progress_callback
            )

            return {
                'success': True,
                'Y': Y,
                'P': P,
                'Q': Q,
                'C_history': C_history,
                'n': n
            }

        except Exception as e:
            return {'success': False, 'error': str(e)}

    def _compute_P(self, X, perplexity):
        """Compute pairwise affinities P_ij"""
        n = X.shape[0]

        sum_X = np.sum(X**2, axis=1)
        D = sum_X[:, np.newaxis] + sum_X[np.newaxis, :] - 2 * X @ X.T
        D = np.maximum(D, 0)

        P = np.zeros((n, n))
        target_entropy = np.log2(perplexity)

        for i in range(n):
            beta_min = -np.inf
            beta_max = np.inf
            beta = 1.0

            for _ in range(50):
                Di = D[i].copy()
                Di[i] = 0

                P_i = np.exp(-Di * beta)
                P_i[i] = 0
                sum_P_i = np.sum(P_i)

                if sum_P_i == 0:
                    P_i = np.ones(n) / n
                    sum_P_i = 1.0

                P_i = P_i / sum_P_i

                P_i_nonzero = P_i[P_i > 1e-12]
                H = -np.sum(P_i_nonzero * np.log2(P_i_nonzero))

                H_diff = H - target_entropy
                if np.abs(H_diff) < 1e-5:
                    break

                if H_diff > 0:
                    beta_min = beta
                    if beta_max == np.inf:
                        beta = beta * 2
                    else:
                        beta = (beta + beta_max) / 2
                else:
                    beta_max = beta
                    if beta_min == -np.inf:
                        beta = beta / 2
                    else:
                        beta = (beta + beta_min) / 2

            P[i] = P_i

        P = (P + P.T) / (2 * n)
        P = np.maximum(P, 1e-12)

        return P

    def _optimize_tsne(self, P, Y, learning_rate, n_iter, early_exaggeration, momentum, progress_callback=None):
        """Optimize t-SNE using gradient descent"""
        n = Y.shape[0]
        Y_velocity = np.zeros_like(Y)
        C_history = []

        P_exag = P * early_exaggeration

        for iteration in range(n_iter):
            P_current = P_exag if iteration < 250 else P

            sum_Y = np.sum(Y**2, axis=1)
            D_low = sum_Y[:, np.newaxis] + sum_Y[np.newaxis, :] - 2 * Y @ Y.T
            D_low = np.maximum(D_low, 0)

            Q = (1 + D_low) ** (-1)
            np.fill_diagonal(Q, 0)
            sum_Q = np.sum(Q)
            if sum_Q < 1e-12:
                sum_Q = 1e-12
            Q = Q / sum_Q
            Q = np.maximum(Q, 1e-12)

            C = np.sum(P_current * np.log((P_current + 1e-12) / (Q + 1e-12)))
            C_history.append(float(C))

            PQ_diff = P_current - Q
            repulsion = (1 + D_low) ** (-1)
            attraction_repulsion = (PQ_diff * repulsion)[:, :, np.newaxis]
            Y_diff = Y[:, np.newaxis, :] - Y[np.newaxis, :, :]
            gradient = 4 * (attraction_repulsion * Y_diff).sum(axis=1)

            Y_velocity = momentum * Y_velocity - learning_rate * gradient
            Y = Y + Y_velocity
            Y = Y - Y.mean(axis=0)

            if progress_callback and iteration % 10 == 0:
                progress_callback(iteration / n_iter, f'Iteration {iteration}/{n_iter}, Cost: {C:.4f}')

        # Final Q computation
        sum_Y = np.sum(Y**2, axis=1)
        D_low = sum_Y[:, np.newaxis] + sum_Y[np.newaxis, :] - 2 * Y @ Y.T
        D_low = np.maximum(D_low, 0)
        Q = (1 + D_low) ** (-1)
        np.fill_diagonal(Q, 0)
        sum_Q = np.sum(Q)
        if sum_Q < 1e-12:
            sum_Q = 1e-12
        Q = Q / sum_Q
        Q = np.maximum(Q, 1e-12)

        if progress_callback:
            progress_callback(1.0, 'Complete!')

        return Y, Q, C_history

    # ==================== Clustering ====================

    def run_clustering(self, Y, method='kmeans', k=3, eps=0.5, min_samples=5):
        """Run clustering on t-SNE results"""
        try:
            if method == 'kmeans':
                labels = self._kmeans(Y, k)
            elif method == 'dbscan':
                labels = self._dbscan(Y, eps, min_samples)
            else:
                return {'success': False, 'error': 'Unknown clustering method'}

            unique_labels = np.unique(labels)
            summary = []
            for label in unique_labels:
                count = np.sum(labels == label)
                summary.append({
                    'label': int(label),
                    'count': int(count)
                })

            return {
                'success': True,
                'labels': labels.tolist(),
                'summary': summary
            }

        except Exception as e:
            return {'success': False, 'error': str(e)}

    def _kmeans(self, X, k, max_iter=100):
        """K-means clustering"""
        n = X.shape[0]
        indices = np.random.choice(n, k, replace=False)
        centroids = X[indices].copy()
        labels = np.zeros(n, dtype=int)

        for _ in range(max_iter):
            distances = np.zeros((n, k))
            for i in range(k):
                distances[:, i] = np.sum((X - centroids[i])**2, axis=1)

            new_labels = np.argmin(distances, axis=1)

            if np.all(labels == new_labels):
                break

            labels = new_labels

            for i in range(k):
                cluster_points = X[labels == i]
                if len(cluster_points) > 0:
                    centroids[i] = cluster_points.mean(axis=0)

        return labels

    def _dbscan(self, X, eps, min_samples):
        """DBSCAN clustering"""
        n = X.shape[0]
        labels = -np.ones(n, dtype=int)
        cluster_id = 0

        for i in range(n):
            if labels[i] != -1:
                continue

            neighbors = self._find_neighbors(X, i, eps)

            if len(neighbors) < min_samples:
                labels[i] = -1
            else:
                self._expand_cluster(X, labels, i, neighbors, cluster_id, eps, min_samples)
                cluster_id += 1

        return labels

    def _find_neighbors(self, X, point_idx, eps):
        """Find neighbors within eps distance"""
        distances = np.sum((X - X[point_idx])**2, axis=1)
        return np.where(distances <= eps**2)[0]

    def _expand_cluster(self, X, labels, point_idx, neighbors, cluster_id, eps, min_samples):
        """Expand cluster from seed point"""
        labels[point_idx] = cluster_id

        i = 0
        while i < len(neighbors):
            neighbor_idx = neighbors[i]

            if labels[neighbor_idx] == -1:
                labels[neighbor_idx] = cluster_id

            if labels[neighbor_idx] != -1:
                i += 1
                continue

            labels[neighbor_idx] = cluster_id

            new_neighbors = self._find_neighbors(X, neighbor_idx, eps)
            if len(new_neighbors) >= min_samples:
                neighbors = np.concatenate([neighbors, new_neighbors])

            i += 1


# ==================== Streamlit UI ====================

def main():
    # Page config
    st.set_page_config(
        page_title="t-SNE Explorer",
        page_icon="📊",
        layout="wide",
        initial_sidebar_state="expanded"
    )

    # Inject custom CSS
    inject_custom_css()

    # Header with MCP indicator
    st.markdown("""

    <div style="background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);

                padding: 40px; text-align: center; border-radius: 16px; margin-bottom: 20px;">

        <h1 style="color: white; font-size: 3em; margin-bottom: 10px;">t-SNE Explorer</h1>

        <p style="color: white; font-size: 1.2em; opacity: 0.95;">

            Transparent t-SNE with synthetic data generation and file uploads

        </p>

        <p style="color: #ffd700; font-size: 0.9em; margin-top: 10px;">

            🔧 MCP Server Active - Android clients can connect!

        </p>

    </div>

    """, unsafe_allow_html=True)

    # Initialize backend
    if 'backend' not in st.session_state:
        st.session_state.backend = TSNEExplorer()

    # Initialize session state
    if 'datasets' not in st.session_state:
        st.session_state.datasets = {}
    if 'current_results' not in st.session_state:
        st.session_state.current_results = None

    # Sidebar navigation
    st.sidebar.title("Navigation")
    tab = st.sidebar.radio("Select Section", ["t-SNE", "Upload", "MCP Info"])

    # MCP Connection Info in sidebar
    st.sidebar.markdown("---")
    st.sidebar.markdown("### 🔧 MCP Server")
    st.sidebar.success("✅ Active")
    st.sidebar.caption("Android devices can connect to this server for remote computations")

    if tab == "t-SNE":
        tsne_tab()
    elif tab == "Upload":
        upload_tab()
    else:
        mcp_info_tab()


def tsne_tab():
    """Main t-SNE tab"""
    st.header("t-SNE Analysis")

    # Section A: Synthetic Data Generator
    with st.expander("A) Synthetic Data Generator", expanded=True):
        st.info("Generate n points in d dimensions with k distinct distance types. "
                "Optimal: k=1 (n≤d+1 simplex), k=2 (n=5 pentagon), k=3 (n=7 heptagon).")

        col1, col2, col3, col4 = st.columns(4)
        with col1:
            n = st.number_input("n (points)", min_value=1, max_value=100, value=6)
        with col2:
            d = st.number_input("d (dimensions)", min_value=1, max_value=100, value=10)
        with col3:
            k = st.number_input("k (distance types)", min_value=1, value=2)
        with col4:
            seed = st.number_input("seed", min_value=0, value=42)

        if st.button("Generate Points", key="gen_points"):
            with st.spinner("Generating synthetic data..."):
                result = st.session_state.backend.generate_simplex_points(n, d, k, seed)

                if result['success']:
                    # Store dataset
                    dataset_id = f"synthetic_{len(st.session_state.datasets)}"
                    st.session_state.datasets[dataset_id] = {
                        'type': 'synthetic',
                        'X': np.array(result['points']),
                        'shape': (result['n'], result['d'])
                    }

                    # Display stats
                    st.success(f"Generated {result['n']} points successfully!")

                    col1, col2, col3, col4 = st.columns(4)
                    col1.metric("Points", result['n'])
                    col2.metric("Dimensions", result['d'])
                    col3.metric("Target k", result['k'])
                    col4.metric("Actual k", result['actual_k'])

                    st.write(f"**Unique Distances:** {', '.join([f'{d:.4f}' for d in result['unique_distances']])}")
                    st.write(f"**Range:** min={result['distances_min']:.4f}, "
                            f"mean={result['distances_mean']:.4f}, max={result['distances_max']:.4f}")

                    # Display points table
                    points_df = pd.DataFrame(
                        result['points'],
                        columns=[f'x{i+1}' for i in range(result['d'])]
                    )
                    st.dataframe(points_df.head(10), use_container_width=True)
                else:
                    st.error(result['error'])

    # Section B: MNIST Loader
    with st.expander("B) Load MNIST Dataset"):
        col1, col2 = st.columns(2)
        with col1:
            subset = st.selectbox("Subset", ["train", "test"])
        with col2:
            max_samples = st.number_input("Samples", min_value=100, max_value=10000, value=1000, step=100)

        if st.button("Load MNIST", key="load_mnist"):
            with st.spinner("Loading MNIST dataset..."):
                progress_bar = st.progress(0)
                progress_bar.progress(0.3)

                result = st.session_state.backend.load_mnist(max_samples, subset)
                progress_bar.progress(1.0)

                if result['success']:
                    dataset_id = f"mnist_{len(st.session_state.datasets)}"
                    st.session_state.datasets[dataset_id] = {
                        'type': 'mnist',
                        'X': result['X'],
                        'labels': result['labels'],
                        'count': result['count']
                    }
                    st.success(result['message'])
                else:
                    st.error(result['error'])

    # Section C: t-SNE Runner
    with st.expander("C) t-SNE Runner", expanded=True):
        # Dataset selector
        dataset_options = {f"{k} ({v['type']})": k for k, v in st.session_state.datasets.items()}

        if len(dataset_options) == 0:
            st.warning("No datasets available. Generate synthetic data or load MNIST first.")
            return

        selected_dataset_key = st.selectbox(
            "Select Dataset",
            options=list(dataset_options.keys())
        )
        selected_dataset_id = dataset_options[selected_dataset_key]

        # t-SNE parameters
        col1, col2, col3 = st.columns(3)
        with col1:
            perplexity = st.number_input("Perplexity", min_value=5, max_value=50, value=30)
            learning_rate = st.number_input("Learning Rate", min_value=10, max_value=1000, value=200)
        with col2:
            iterations = st.number_input("Iterations", min_value=100, max_value=5000, value=1000)
            early_exag = st.number_input("Early Exaggeration", min_value=1, max_value=50, value=12)
        with col3:
            momentum = st.number_input("Momentum", min_value=0.0, max_value=1.0, value=0.8, step=0.1)
            tsne_seed = st.number_input("Seed", min_value=0, value=42, key="tsne_seed")

        if st.button("Run t-SNE", key="run_tsne"):
            dataset = st.session_state.datasets[selected_dataset_id]
            X = dataset['X']

            progress_bar = st.progress(0)
            progress_text = st.empty()

            def progress_callback(progress, message):
                progress_bar.progress(progress)
                progress_text.text(message)

            result = st.session_state.backend.run_tsne(
                X, perplexity, learning_rate, iterations,
                early_exag, momentum, tsne_seed, progress_callback
            )

            if result['success']:
                st.session_state.current_results = result
                st.session_state.current_results['dataset_id'] = selected_dataset_id
                st.session_state.current_results['labels'] = dataset.get('labels')
                st.success("t-SNE completed successfully!")
                st.rerun()
            else:
                st.error(result['error'])

    # Section D: Results Display
    if st.session_state.current_results:
        display_results()


def display_results():
    """Display t-SNE results"""
    st.header("Results & Internals")

    results = st.session_state.current_results
    Y = np.array(results['Y'])
    P = np.array(results['P'])
    Q = np.array(results['Q'])
    C_history = results['C_history']
    labels = results.get('labels')

    # 2D Scatter Plot
    st.subheader("2D t-SNE Embedding")

    if labels is not None:
        # Color by labels
        fig = go.Figure()

        unique_labels = np.unique(labels)
        colors = ['#e74c3c', '#3498db', '#2ecc71', '#f39c12', '#9b59b6',
                  '#1abc9c', '#e67e22', '#95a5a6', '#34495e', '#c0392b']

        for label in unique_labels:
            mask = labels == label
            fig.add_trace(go.Scatter(
                x=Y[mask, 0],
                y=Y[mask, 1],
                mode='markers',
                name=f'Digit {label}',
                marker=dict(size=8, color=colors[int(label) % len(colors)],
                           line=dict(color='white', width=1))
            ))
    else:
        # Default plot
        fig = go.Figure(data=go.Scatter(
            x=Y[:, 0],
            y=Y[:, 1],
            mode='markers+text',
            text=[f'y{i+1}' for i in range(len(Y))],
            textposition='top center',
            marker=dict(size=10, color='#667eea', line=dict(color='white', width=1))
        ))

    fig.update_layout(
        title="t-SNE Embedding",
        xaxis_title="Dimension 1",
        yaxis_title="Dimension 2",
        height=500
    )
    st.plotly_chart(fig, use_container_width=True)

    # Cost Plot
    st.subheader("Cost (KL Divergence) Over Iterations")
    fig_cost = go.Figure(data=go.Scatter(
        y=C_history,
        mode='lines',
        line=dict(color='#e74c3c', width=2)
    ))
    fig_cost.update_layout(
        xaxis_title="Iteration",
        yaxis_title="Cost (KL Divergence)",
        height=400
    )
    st.plotly_chart(fig_cost, use_container_width=True)

    # Matrices
    col1, col2 = st.columns(2)

    with col1:
        st.subheader("P Matrix (High-D Affinities)")
        fig_p = go.Figure(data=go.Heatmap(z=P, colorscale='Viridis'))
        fig_p.update_layout(height=400)
        st.plotly_chart(fig_p, use_container_width=True)

    with col2:
        st.subheader("Q Matrix (Low-D Affinities)")
        fig_q = go.Figure(data=go.Heatmap(z=Q, colorscale='Viridis'))
        fig_q.update_layout(height=400)
        st.plotly_chart(fig_q, use_container_width=True)

    # Coordinates Table
    st.subheader("2D Coordinates")
    coords_df = pd.DataFrame(Y, columns=['Dim 1', 'Dim 2'])
    coords_df.index = [f'y{i+1}' for i in range(len(Y))]
    st.dataframe(coords_df.head(20), use_container_width=True)

    # Export
    if st.button("Export Results (CSV)"):
        csv = coords_df.to_csv()
        st.download_button(
            label="Download CSV",
            data=csv,
            file_name="tsne_results.csv",
            mime="text/csv"
        )
        st.success("Results exported!")

    # Clustering section
    if labels is not None:
        st.subheader("Clustering")
        cluster_method = st.selectbox("Method", ["kmeans", "dbscan"])

        if cluster_method == "kmeans":
            k = st.number_input("k (clusters)", min_value=2, max_value=10, value=3)
            if st.button("Run K-Means"):
                result = st.session_state.backend.run_clustering(Y, 'kmeans', k=k)
                if result['success']:
                    st.success("Clustering complete!")
                    st.write("**Cluster Summary:**")
                    st.json(result['summary'])
        else:
            col1, col2 = st.columns(2)
            with col1:
                eps = st.number_input("eps", min_value=0.1, value=0.5, step=0.1)
            with col2:
                min_samples = st.number_input("min_samples", min_value=1, value=5)
            if st.button("Run DBSCAN"):
                result = st.session_state.backend.run_clustering(Y, 'dbscan', eps=eps, min_samples=min_samples)
                if result['success']:
                    st.success("Clustering complete!")
                    st.write("**Cluster Summary:**")
                    st.json(result['summary'])


def mcp_info_tab():
    """MCP Server Information tab"""
    st.header("🔧 MCP Server Information")

    st.success("✅ MCP Server is running on this Hugging Face Space!")

    st.markdown("""

    This deployment provides both:

    1. **Web UI** - Access via browser (what you're using now)

    2. **MCP API Server** - For android.py to connect remotely

    """)

    st.subheader("📱 Connect from android.py")

    st.markdown("""

    Your Android device can use this Hugging Face Space as the computation server!



    **Setup Instructions:**



    1. On your Android device, edit `android.py`

    2. Set the MCP server URL:

    """)

    base_url = "https://euler314-t-sne.hf.space"
    st.code(f"""

# In android.py, modify the MCP client connection:



MCP_SERVER_URL = "{base_url}"

MCP_PORT = 8501



# Or set environment variable:

export MCP_SERVER_URL={base_url}

export MCP_PORT=8501

python android.py

    """, language="python")

    st.subheader("🌐 Available MCP Endpoints")

    endpoints = [
        {
            "endpoint": "/mcp/health",
            "method": "GET",
            "description": "Check MCP server status"
        },
        {
            "endpoint": "/mcp/generate_simplex_points",
            "method": "POST",
            "description": "Generate synthetic data",
            "params": "n, d, k, seed"
        },
        {
            "endpoint": "/mcp/load_mnist",
            "method": "POST",
            "description": "Load MNIST dataset",
            "params": "max_samples, subset"
        },
        {
            "endpoint": "/mcp/run_tsne",
            "method": "POST",
            "description": "Run t-SNE computation",
            "params": "X, perplexity, learning_rate, n_iter, etc."
        },
        {
            "endpoint": "/mcp/run_clustering",
            "method": "POST",
            "description": "Run clustering",
            "params": "Y, method, k, eps, min_samples"
        }
    ]

    for ep in endpoints:
        with st.expander(f"**{ep['method']}** `{ep['endpoint']}`"):
            st.write(f"**Description:** {ep['description']}")
            if 'params' in ep:
                st.write(f"**Parameters:** {ep['params']}")
            st.code(f"""

# Example usage:

import requests



url = "{base_url}{ep['endpoint']}"

response = requests.{ep['method'].lower()}(url, json={{"param": "value"}})

result = response.json()

            """, language="python")

    st.subheader("🧪 Test MCP Connection")

    if st.button("Test MCP Health Endpoint"):
        try:
            import requests
            response = requests.get(f"{base_url}/mcp/health", timeout=5)
            if response.status_code == 200:
                st.success("✅ MCP Server is responding!")
                st.json(response.json())
            else:
                st.error(f"❌ Server returned status {response.status_code}")
        except Exception as e:
            st.warning(f"⚠️ Could not connect: {str(e)}")
            st.info("This is normal if the MCP port (8501) is not exposed. Check Hugging Face Spaces settings.")

    st.subheader("📖 Architecture")

    st.markdown("""

    ```

    ┌─────────────────────┐

    │  Android Device     │

    │  (android.py)       │ ← Your mobile device

    └──────────┬──────────┘

               │ HTTP/REST

               ▼

    ┌─────────────────────────────────┐

    │  Hugging Face Space             │

    │  (This deployment)              │

    │                                 │

    │  🌐 Streamlit UI (Port 7860)   │ ← Web browser

    │  🔧 MCP Server (Port 8501)     │ ← Android connection

    └─────────────────────────────────┘

    ```

    """)

    st.subheader("⚡ Benefits")

    col1, col2 = st.columns(2)
    with col1:
        st.markdown("""

        **For Android:**

        - Offload heavy computations

        - No local sklearn needed

        - Faster t-SNE execution

        - Save battery & memory

        """)
    with col2:
        st.markdown("""

        **For This Server:**

        - Free Hugging Face compute

        - Always available (24/7)

        - Shared by all clients

        - Auto-scaling

        """)

    st.info("💡 **Tip:** You can deploy your own MCP server on any cloud platform if you need more resources!")


def upload_tab():
    """Upload tab for CSV and images"""
    st.header("Upload Data")

    st.subheader("CSV Files")
    uploaded_csv = st.file_uploader("Upload CSV", type=['csv'], accept_multiple_files=False)

    if uploaded_csv:
        try:
            df = pd.read_csv(uploaded_csv)

            dataset_id = f"csv_{len(st.session_state.datasets)}"
            numeric_columns = df.select_dtypes(include=[np.number]).columns.tolist()

            st.success(f"Uploaded {uploaded_csv.name}")
            st.write(f"Shape: {df.shape}")
            st.write(f"Numeric columns: {', '.join(numeric_columns)}")

            st.dataframe(df.head())

            # Column selector
            selected_cols = st.multiselect("Select numeric columns", numeric_columns, default=numeric_columns)
            handle_missing = st.selectbox("Handle missing values", ["drop", "mean", "zero"])

            if st.button("Prepare Dataset"):
                if selected_cols:
                    df_subset = df[selected_cols]

                    if handle_missing == 'drop':
                        df_subset = df_subset.dropna()
                    elif handle_missing == 'mean':
                        df_subset = df_subset.fillna(df_subset.mean())
                    elif handle_missing == 'zero':
                        df_subset = df_subset.fillna(0)

                    X = df_subset.values

                    st.session_state.datasets[dataset_id] = {
                        'type': 'csv',
                        'X': X,
                        'shape': X.shape,
                        'name': uploaded_csv.name
                    }

                    st.success(f"Dataset prepared: {X.shape[0]} rows × {X.shape[1]} columns")
                else:
                    st.warning("Please select at least one column")

        except Exception as e:
            st.error(f"Error uploading CSV: {str(e)}")

    # Dataset list
    st.subheader("Uploaded Datasets")
    if len(st.session_state.datasets) == 0:
        st.info("No datasets uploaded yet")
    else:
        for dataset_id, dataset in st.session_state.datasets.items():
            if dataset['type'] == 'synthetic':
                st.write(f"🔢 Synthetic: {dataset['shape'][0]}×{dataset['shape'][1]}")
            elif dataset['type'] == 'csv':
                st.write(f"📊 CSV: {dataset.get('name', 'Unknown')} ({dataset['shape'][0]}×{dataset['shape'][1]})")
            elif dataset['type'] == 'mnist':
                st.write(f"✏️ MNIST: {dataset['count']} samples")


# ==================== Note: MCP Server ====================
# The MCP server is now standalone (mcp_flask_server.py) when deployed to HF Spaces
# It runs separately via supervisor and is routed through nginx
# This keeps the Streamlit app clean and allows proper port routing


if __name__ == '__main__':
    main()