Update fake_audio.py

fake_audio.py

@@ -1,15 +1,14 @@
 # -*- coding: utf-8 -*-
 """
 Cleaned fake vs forged vs real audio module (Gradio UI & Notebook Magics removed)
+Fixed for PyTorch 2.6+ compatibility with LAION-CLAP checkpoints
 """
 
-# Note: In pure Python files, install requirements via terminal, not in the code.
-# pip install torch librosa numpy laion-clap groq openai-whisper transformers
-# apt-get install -y ffmpeg
-
 import os
+import urllib.request
 from huggingface_hub import login
 import torch
+import torch.serialization
 import librosa
 import numpy as np
 import numpy
@@ -21,7 +20,6 @@ from transformers import AutoFeatureExtractor, AutoModelForAudioClassification,
 from groq import Groq
 
 # --- FETCH SECRETS & LOGIN ---
-# Hugging Face Secrets se keys fetch karna
 HF_TOKEN = os.environ.get("HF_TOKEN")
 GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
 
@@ -36,7 +34,7 @@ client = Groq(api_key=GROQ_API_KEY)
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 AST_MODEL_PATH = "aneela-pervez/FAKE-AUDIO"
 
-# --- 🏷️ CATEGORIZED FORENSIC LABELS ---
+# --- CATEGORIZED FORENSIC LABELS ---
 FORENSIC_CATEGORIES = {
     "AUTHENTIC": ["Natural human speech", "Human voice with natural breathing", "Consistent room ambiance"],
     "AI_SYNTHETIC": ["AI voice clone robotic smoothness", "Synthetic neural vocoder", "Deepfake metallic resonance"],
@@ -57,46 +55,60 @@ try:
 except Exception as e:
     print(f"⚠️ AST Model issue (using default logic if fails): {e}")
 
-# Note: laion_clap loads weights during initialization. 
-# PyTorch 2.6+ compatibility fix for CLAP checkpoint loading
-import torch.serialization
-torch.serialization.add_safe_globals([numpy.core.multiarray.scalar])
+# --- CLAP MODEL LOADING (PyTorch 2.6+ Full Fix) ---
+# Step 1: Checkpoint pehle local mein download karein
+CLAP_CKPT_URL = "https://huggingface.co/lukewys/laion_clap/resolve/main/music_audioset_epoch_15_esc_90.14.pt"
+CLAP_CKPT_PATH = "/tmp/clap_music_audioset.pt"
 
-# Monkey-patch torch.load to force weights_only=False for CLAP
-_orig_load = torch.load
-def _safe_load(f, map_location=None, pickle_module=None, weights_only=True, **kwargs):
-    return _orig_load(f, map_location=map_location, weights_only=False, **kwargs)
-torch.load = _safe_load
+if not os.path.exists(CLAP_CKPT_PATH):
+    print("Downloading CLAP checkpoint to local disk...")
+    urllib.request.urlretrieve(CLAP_CKPT_URL, CLAP_CKPT_PATH)
+    print("✅ CLAP checkpoint downloaded!")
+else:
+    print("✅ CLAP checkpoint already cached.")
+
+# Step 2: PyTorch 2.6+ safe globals fix
+torch.serialization.add_safe_globals([numpy.core.multiarray.scalar])
 
+# Step 3: CLAP model initialize karein
 clap_model = laion_clap.CLAP_Module(enable_fusion=False, amodel='HTSAT-tiny').to(DEVICE)
-clap_model.load_ckpt(
-    ckpt="https://huggingface.co/lukewys/laion_clap/resolve/main/music_audioset_epoch_15_esc_90.14.pt"
-)
 
-# Restore original torch.load after CLAP is loaded
-torch.load = _orig_load
+# Step 4: Manually load checkpoint with strict=False
+# strict=False: unexpected/missing keys jaise position_ids ko ignore karta hai
+print("🔄 Loading CLAP checkpoint (strict=False)...")
+ckpt = torch.load(CLAP_CKPT_PATH, map_location=DEVICE, weights_only=False)
+
+# Checkpoint structure check
+if "state_dict" in ckpt:
+    state_dict = ckpt["state_dict"]
+elif "model" in ckpt:
+    state_dict = ckpt["model"]
+else:
+    state_dict = ckpt
+
+clap_model.model.load_state_dict(state_dict, strict=False)
+clap_model.eval()
+print("✅ CLAP model loaded successfully!")
+
 
 def extract_dna_embeddings(y):
-    # Ensure standard numpy array shape before processing
     y = np.ascontiguousarray(y)
     inputs = id_extractor(y, sampling_rate=16000, return_tensors="pt")
-    
-    # 🛡️ EXPLICIT TENSOR CONVERSION FIX:
     inputs = {k: (v.clone().detach() if isinstance(v, torch.Tensor) else torch.tensor(v)).to(DEVICE) for k, v in inputs.items()}
-    
     with torch.no_grad():
         logits = id_model(**inputs).logits
         emb = torch.mean(logits, dim=1).cpu().numpy()
     return normalize(emb)
 
+
 def analyze_voice_forensics(audio_path):
     try:
         y, sr = librosa.load(audio_path, sr=16000)
         y_48, _ = librosa.load(audio_path, sr=48000)
         y = y / (np.max(np.abs(y)) + 1e-9)
-        y = np.ascontiguousarray(y) # Data alignment safety
+        y = np.ascontiguousarray(y)
 
-        # 1. 🛡️ SEGMENT-WISE CROSS VALIDATION
+        # 1. SEGMENT-WISE CROSS VALIDATION
         segments = np.array_split(y, 3)
         seg_embs = [extract_dna_embeddings(s) for s in segments]
         cos_sim = cosine_similarity(seg_embs[0], seg_embs[-1])[0][0]
@@ -105,23 +117,18 @@ def analyze_voice_forensics(audio_path):
         # 2. FEATURE EXTRACTION
         transcription = whisper_model.transcribe(audio_path)["text"]
         inputs = ast_extractor(y, sampling_rate=16000, return_tensors="pt")
-        
-        # 🛡️ EXPLICIT TENSOR CONVERSION FIX:
         inputs = {k: (v.clone().detach() if isinstance(v, torch.Tensor) else torch.tensor(v)).to(DEVICE) for k, v in inputs.items()}
-        
+
         with torch.no_grad():
             ast_prob = torch.nn.functional.softmax(ast_model(**inputs).logits, dim=-1).cpu().numpy()[0][1]
             ast_score = ast_prob * 100
 
-        # 🛡️ CLAP TENSOR FIX: Reverted to your strict `torch.from_numpy` logic to prevent errors
+        # CLAP EMBEDDINGS
         audio_emb_np = clap_model.get_audio_embedding_from_data(x=[y_48])
         text_emb_np = clap_model.get_text_embedding(ALL_LABELS)
-        
-        # Explicitly converting NumPy arrays to PyTorch Tensors before calculating similarity
         audio_emb = torch.from_numpy(audio_emb_np).to(DEVICE)
         text_emb = torch.from_numpy(text_emb_np).to(DEVICE)
         c_sims = torch.nn.functional.cosine_similarity(audio_emb, text_emb)
-        
         top_idx = torch.argmax(c_sims).item()
         clap_label = ALL_LABELS[top_idx]
 
@@ -133,14 +140,14 @@ def analyze_voice_forensics(audio_path):
 
         num_edits = len(np.where(np.diff(librosa.onset.onset_strength(y=y, sr=16000)) > 6.5)[0])
 
-        # --- 🧠 STEP 3: FUZZY LOGIC MEMBERSHIP SCORING ---
+        # FUZZY LOGIC MEMBERSHIP SCORING
         mu_ai = min(1.0, max(0.0, (ast_score - 40) / 45))
         mu_spliced = min(1.0, max(0.0, (num_edits - 8) / 10))
         dna_conf = (cos_sim - 0.85) / 0.10
         cat_conf = 1.0 if category == "AUTHENTIC" else 0.0
         mu_auth = min(1.0, max(0.0, (dna_conf + cat_conf) / 2))
 
-        # --- 🛡️ STEP 4: BALANCED FUZZY DECISION ENGINE (V4.0) ---
+        # FUZZY DECISION ENGINE (V4.0)
         if mu_auth > 0.80 and mu_spliced < 0.75:
             verdict = "✅ AUTHENTIC HUMAN VOICE"
             category_override = "AUTHENTIC"
@@ -151,7 +158,7 @@ def analyze_voice_forensics(audio_path):
             verdict = "⚠️ MANIPULATED REAL VOICE (Spliced)"
             category_override = "MANIPULATED"
 
-        # --- 🔥 ALIGNED PROFESSIONAL REASONING ---
+        # LLM REASONING
         audit_context = (f"Verdict: {verdict}. Category: {category_override}. Metrics: ID {cos_sim:.2f}, "
                          f"Texture {ast_score:.1f}%, Cuts {num_edits}. "
                          f"Fuzzy Scores: AI={mu_ai:.2f}, Spliced={mu_spliced:.2f}, Auth={mu_auth:.2f}")
@@ -161,7 +168,10 @@ def analyze_voice_forensics(audio_path):
                   "emphasize that DNA integrity confirms human origin despite high spectral texture. "
                   "Give a 2-line direct forensic conclusion.")
 
-        res = client.chat.completions.create(model="llama-3.3-70b-versatile", messages=[{"role": "user", "content": prompt}])
+        res = client.chat.completions.create(
+            model="llama-3.3-70b-versatile",
+            messages=[{"role": "user", "content": prompt}]
+        )
         reasoning = res.choices[0].message.content
 
         evidence = f"Category: {category} | ID: {cos_sim:.2f} | Texture: {ast_score:.1f}% | Cuts: {num_edits}"