Update app.py

app.py

@@ -1,3 +1,77 @@
+
+# ------------------------------------------------------------
+# NEW — Dimensional‑reduction / clustering feature
+# ------------------------------------------------------------
+def dim_reduction(method='t-SNE', perplexity=30, n_neighbors=15, min_dist=0.1,
+                  use_cnn=False, cluster_eps=0.5, cluster_min_samples=5):
+    """
+    Compute 2‑D embedding (t‑SNE or UMAP) of per‑storm features.
+    Returns a Plotly Figure and an info string.
+    """
+    global typhoon_max, typhoon_data
+    if typhoon_max is None:
+        return None, "No typhoon data loaded."
+
+    df = typhoon_max.copy()
+
+    # Select features
+    if use_cnn and TF_AVAILABLE:
+        MAX_LEN = 100
+        grouped = typhoon_data.groupby('SID')
+        sequences = []
+        for sid in df['SID']:
+            winds = grouped.get_group(sid)['USA_WIND'].fillna(0).values if sid in grouped.groups else []
+            winds = winds[:MAX_LEN]
+            seq = np.zeros(MAX_LEN, dtype=np.float32)
+            seq[:len(winds)] = winds
+            sequences.append(seq)
+        X = np.stack(sequences)[..., None]
+
+        model = models.Sequential([
+            layers.Conv1D(16, 5, activation='relu', input_shape=(MAX_LEN,1)),
+            layers.MaxPool1D(2),
+            layers.Conv1D(32, 3, activation='relu'),
+            layers.GlobalAveragePooling1D(),
+            layers.Dense(32, activation='relu')
+        ])
+        embeddings = model.predict(X, verbose=0)
+        features_for_dr = embeddings
+        feature_text = "CNN latent vectors"
+    else:
+        features = ['USA_WIND','USA_PRES','LAT','LON']
+        features_for_dr = df[features].fillna(0).values
+        feature_text = ", ".join(features)
+
+    from sklearn.preprocessing import StandardScaler
+    X_std = StandardScaler().fit_transform(features_for_dr)
+
+    if method == 't-SNE':
+        from sklearn.manifold import TSNE
+        reducer = TSNE(n_components=2, init='pca', perplexity=perplexity, learning_rate='auto', random_state=0)
+        coords = reducer.fit_transform(X_std)
+    else:
+        reducer = umap.UMAP(n_neighbors=n_neighbors, min_dist=min_dist, n_components=2, random_state=0)
+        coords = reducer.fit_transform(X_std)
+
+    df['DR_X'] = coords[:,0]
+    df['DR_Y'] = coords[:,1]
+
+    from sklearn.cluster import DBSCAN
+    clusterer = DBSCAN(eps=cluster_eps, min_samples=cluster_min_samples)
+    labels = clusterer.fit_predict(coords)
+    df['Cluster'] = labels.astype(str)
+
+    import plotly.express as px
+    fig = px.scatter(
+        df, x='DR_X', y='DR_Y',
+        color='Category', symbol='Cluster',
+        hover_data={'Name': df['NAME'], 'Year': df['Year'], 'Wind': df['USA_WIND'], 'Pressure': df['USA_PRES']},
+        title=f"{method} embedding (features: {feature_text})"
+    )
+    fig.update_traces(marker=dict(size=6, line=dict(width=0.5,color='black')))
+    info=f"Method: {method} | Points: {len(df)} | Clusters: {len(set(labels))-('-1' in set(labels))}"
+    return fig, info
+
 import os
 import argparse
 import logging
@@ -10,6 +84,13 @@ import csv
 import gradio as gr
 import pandas as pd
 import numpy as np
+import umap   # NEW
+try:
+    import tensorflow as tf
+    from tensorflow.keras import layers, models
+    TF_AVAILABLE = True
+except Exception:
+    TF_AVAILABLE = False
 import matplotlib.pyplot as plt
 import matplotlib.animation as animation
 from matplotlib.backends.backend_agg import FigureCanvasAgg as FigureCanvas
@@ -1162,6 +1243,27 @@ def create_interface():
                     outputs=[regression_plot, slopes_text, lon_regression_results]
                 )
             
+            
+# ---- Dimensional Reduction (t‑SNE / UMAP) ----
+with gr.Tab("Dimensional Reduction (t‑SNE / UMAP)"):
+    dr_method   = gr.Dropdown(['t-SNE','UMAP'], label="Method", value='t-SNE')
+    tsne_perp   = gr.Slider(5,100,step=5,value=30,label="t‑SNE Perplexity")
+    umap_nn     = gr.Slider(5,100,step=1,value=15,label="UMAP n_neighbors")
+    umap_md     = gr.Slider(0.0,1.0,step=0.05,value=0.1,label="UMAP min_dist")
+    use_cnn_box = gr.Checkbox(label="Use CNN latent features", value=False, visible=TF_AVAILABLE)
+    cluster_eps = gr.Slider(0.1,5.0,step=0.1,value=0.5,label="DBSCAN ε")
+    cluster_min = gr.Slider(2,20,step=1,value=5,label="DBSCAN min_samples")
+    dr_btn      = gr.Button("Compute Embedding")
+    dr_plot     = gr.Plot()
+    dr_info     = gr.Textbox(label="Info")
+    def _dr_wrapper(method, perp, nn, md, cnn, eps, mns):
+        return dim_reduction(method, perp, nn, md, cnn, eps, mns)
+    dr_btn.click(
+        fn=_dr_wrapper,
+        inputs=[dr_method, tsne_perp, umap_nn, umap_md, use_cnn_box, cluster_eps, cluster_min],
+        outputs=[dr_plot, dr_info]
+    )
+
             with gr.Tab("Tropical Cyclone Path Animation"):
                 with gr.Row():
                     year_dropdown = gr.Dropdown(label="Year", choices=[str(y) for y in range(1950, 2025)], value="2000")