| import os |
| import glob |
| import numpy as np |
| import pandas as pd |
| import matplotlib.pyplot as plt |
| import seaborn as sns |
| from PIL import Image |
| from umap import UMAP |
|
|
| sns.set_theme(style="whitegrid") |
|
|
| def load_ood_renders(method_dir): |
| ood_dir = os.path.join(method_dir, "renders_ood") |
| if not os.path.exists(ood_dir): |
| return None |
| images = sorted(glob.glob(os.path.join(ood_dir, "*.png"))) |
| if not images: |
| return None |
| |
| feature_vector = [] |
| for img_path in images: |
| img = Image.open(img_path).convert("RGB").resize((64, 64)) |
| feature_vector.append(np.array(img).flatten()) |
| |
| return np.concatenate(feature_vector) |
|
|
| def main(): |
| outputs_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "outputs")) |
| plots_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "plots")) |
| os.makedirs(plots_dir, exist_ok=True) |
| |
| features = [] |
| labels = [] |
| |
| for folder in os.listdir(outputs_dir): |
| method_dir = os.path.join(outputs_dir, folder) |
| if not os.path.isdir(method_dir) or folder == "pathology_archive": |
| continue |
| |
| vec = load_ood_renders(method_dir) |
| if vec is not None: |
| features.append(vec) |
| labels.append(folder.split('_')[0]) |
| |
| if not features: |
| return |
| |
| X = np.stack(features) |
| reducer = UMAP(n_components=2, random_state=42) |
| embedding = reducer.fit_transform(X) |
| |
| df = pd.DataFrame({ |
| 'UMAP_1': embedding[:, 0], |
| 'UMAP_2': embedding[:, 1], |
| 'Method': labels |
| }) |
| |
| plt.figure(figsize=(10, 8)) |
| sns.scatterplot(data=df, x='UMAP_1', y='UMAP_2', hue='Method', s=200, palette='tab10', edgecolor='k') |
| plt.title("Functional Space Convergence (OOD Renders)") |
| plt.tight_layout() |
| plt.savefig(os.path.join(plots_dir, "fig3_functional_manifold.png")) |
| plt.close() |
|
|
| if __name__ == "__main__": |
| main() |
|
|
