Dimensionality Reduction Explorer

import os
import json
import numpy as np
import pandas as pd
import plotly.express as px
import ipywidgets as widgets
from IPython.display import display
import IPython.display as ipd
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
import warnings
warnings.filterwarnings('ignore')

try:
    import umap
    HAS_UMAP = True
except ImportError:
    HAS_UMAP = False
    print("⚠️ umap non installé — pip install umap-learn")

outputs_folder = "../outputs"
data = []
for sound_folder in os.listdir(outputs_folder):
    json_path = os.path.join(outputs_folder, sound_folder, "results.json")
    if os.path.exists(json_path):
        with open(json_path, "r") as f:
            result = json.load(f)
            result["group"] = sound_folder[0]
            data.append(result)

df = pd.DataFrame(data)
audio_dir = os.path.abspath("../data/sounds")
df["audio_path"] = df["file"].apply(lambda x: os.path.join(audio_dir, x))
print(f"✓ {len(df)} sons chargés")

X = np.array(df["mfcc_mean"].tolist())
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X)

pca = PCA(n_components=2)
pca_result = pca.fit_transform(X_scaled)
df["PC1"] = pca_result[:, 0]
df["PC2"] = pca_result[:, 1]
print(f"PCA — variance expliquée : {pca.explained_variance_ratio_.round(3)}")

if HAS_UMAP:
    reducer = umap.UMAP(n_neighbors=10, min_dist=0.3, random_state=42)
    umap_result = reducer.fit_transform(X_scaled)
    df["UMAP1"] = umap_result[:, 0]
    df["UMAP2"] = umap_result[:, 1]
    print("UMAP — calculé")

group_selector = widgets.SelectMultiple(
    options=df["group"].unique(),
    value=tuple(df["group"].unique()),
    description="Group"
)
projection_options = ["PCA", "UMAP"] if HAS_UMAP else ["PCA"]
projection_selector = widgets.ToggleButtons(
    options=projection_options,
    description="Projection"
)

def update_plot(groups, projection):
    filtered_df = df[df["group"].isin(groups)].copy()
    if filtered_df.empty:
        print("Aucune donnée")
        return
    x, y = ("PC1", "PC2") if projection == "PCA" else ("UMAP1", "UMAP2")
    fig = px.scatter(
        filtered_df, x=x, y=y, color="group",
        hover_name="file",
        title=f"{projection} of Crowd Sounds"
    )
    fig.show()

plot_out = widgets.interactive_output(
    update_plot,
    {"groups": group_selector, "projection": projection_selector}
)
display(group_selector, projection_selector, plot_out)