ADD: Filter Duplicate images out

classifier-training/manual-classifier.py

@@ -1,13 +1,15 @@
-# quick_labeler.py
 import random
-import shutil, os
+import shutil
 from pathlib import Path
 from PIL import Image
 import matplotlib.pyplot as plt
+import torch
+import numpy as np
+from sklearn.metrics.pairwise import cosine_similarity
+import open_clip
 
 SOURCE = Path("alle_meine_fotos/")
-images = list(SOURCE.glob("**/*.jpg")) + list(SOURCE.glob("**/*.png"))
-
+THRESHOLD = 0.95
 
 DIRS = [
     "dataset/train/wallpaper", "dataset/train/no_wallpaper",
@@ -16,7 +18,50 @@ DIRS = [
 for d in DIRS:
     Path(d).mkdir(parents=True, exist_ok=True)
 
-for img_path in images:
+# --- Embeddings berechnen ---
+print("Lade CLIP Modell...")
+clip_model, _, preprocess = open_clip.create_model_and_transforms("ViT-B-32", pretrained="openai")
+clip_model.eval()
+
+def get_embedding(img_path):
+    try:
+        img = preprocess(Image.open(img_path).convert("RGB")).unsqueeze(0)
+        with torch.no_grad():
+            return clip_model.encode_image(img).numpy().flatten()
+    except Exception as e:
+        print(f"  Fehler bei {img_path.name}: {e}")
+        return None
+
+all_images = list(SOURCE.glob("**/*.jpg")) + list(SOURCE.glob("**/*.png"))
+
+print(f"{len(all_images)} Bilder gefunden, berechne Embeddings...")
+embeddings, valid_images = [], []
+for i, p in enumerate(all_images):
+    print(f"  {i+1}/{len(all_images)}: {p.name}", end="\r")
+    emb = get_embedding(p)
+    if emb is not None:
+        embeddings.append(emb)
+        valid_images.append(p)
+
+# --- Duplikate rausfiltern ---
+print(f"\nFiltere Duplikate (Threshold={THRESHOLD})...")
+sim_matrix = cosine_similarity(embeddings)
+to_skip = set()
+for i in range(len(valid_images)):
+    if i in to_skip:
+        continue
+    for j in range(i + 1, len(valid_images)):
+        if sim_matrix[i][j] > THRESHOLD:
+            to_skip.add(j)
+
+unique_images = [p for i, p in enumerate(valid_images) if i not in to_skip]
+print(f"{len(valid_images) - len(unique_images)} Duplikate entfernt, {len(unique_images)} verbleiben.\n")
+
+# Modell kann jetzt freigegeben werden
+del clip_model, embeddings, sim_matrix
+
+# --- Labelschleife ---
+for img_path in unique_images:
     img = Image.open(img_path)
     plt.imshow(img)
     plt.title(img_path.name)
@@ -31,4 +76,4 @@ for img_path in images:
     elif label in ("y", "n"):
         folder = "wallpaper" if label == "y" else "no_wallpaper"
         split  = "train" if random.random() < 0.8 else "val"
-        shutil.copy(img_path, f"dataset/{split}/{folder}/{img_path.name}")
+        shutil.copy(img_path, f"dataset/{split}/{folder}/{img_path.name}")