Update app.py

app.py

@@ -1,11 +1,12 @@
 # -*- coding: utf-8 -*-
-import os, shlex, subprocess, tempfile, traceback, time, glob, gc, shutil
+import os, shlex, subprocess, tempfile, traceback, time, glob, gc, shutil, tarfile
 import torch
+import yaml
 from huggingface_hub import snapshot_download
 from nemo.collections import asr as nemo_asr
 import gradio as gr
 
-# 1. CONFIGURATION ET MODÈLES
+# 1. CONFIGURATION
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 SEGMENT_DURATION = 5.0 
 
@@ -23,7 +24,7 @@ MODELS = {
 
 # --- SECTION EXEMPLE ---
 def find_example_video():
-    paths = ["examples/MARALINKE_FIXED.mp4", "examples/MARALINKE.mp4", "MARALINKE.mp4"]
+    paths = ["examples/MARALINKE.mp4", "MARALINKE.mp4"]
     for p in paths:
         if os.path.exists(p): return p
     return None
@@ -31,109 +32,109 @@ def find_example_video():
 EXAMPLE_PATH = find_example_video()
 _cache = {}
 
-# 2. GESTION MÉMOIRE ET CHARGEMENT (NOUVELLE MÉTHODE ANTI-BUG)
+# 2. CHARGEMENT "HARDCORE" (BYPASS DU BUG INIT)
 def get_model(name):
     if name in _cache: return _cache[name]
     gc.collect()
     if torch.cuda.is_available(): torch.cuda.empty_cache()
     
-    repo, model_type = MODELS[name]
+    repo, _ = MODELS[name]
     folder = snapshot_download(repo, local_dir_use_symlinks=False)
     nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
-    
-    print(f"📦 Restauration forcée du modèle : {name}")
-    
-    # MÉTHODE ALTERNATIVE : On bypass le connecteur automatique
+
+    # MÉTHODE DE SECOURS : Chargement par restauration standard
+    # Si ça échoue encore, on utilise une approche par classe explicite
     try:
-        # Tentative avec la méthode classique mais simplifiée au maximum
-        model = nemo_asr.models.ASRModel.restore_from(nemo_file, map_location=torch.device(DEVICE))
-    except TypeError:
-        # Si l'erreur object.init() survient, on utilise le chargement par classe spécifique
-        print("⚠️ Détection du bug NeMo, basculement sur le chargement spécifique...")
-        if "ctc" in name.lower() or model_type == "ctc":
+        from nemo.core.connectors.save_restore_connector import SaveRestoreConnector
+        model = nemo_asr.models.ASRModel.restore_from(
+            nemo_file, 
+            map_location=torch.device(DEVICE),
+            save_restore_connector=SaveRestoreConnector()
+        )
+    except Exception:
+        # Si le bug persiste, on force le type selon le nom
+        print("🔧 Mode de secours activé pour le chargement...")
+        if "ctc" in name.lower() or "Soloba" in name:
             model = nemo_asr.models.EncDecCTCModel.restore_from(nemo_file, map_location=torch.device(DEVICE))
         else:
             model = nemo_asr.models.EncDecHybridRNNTCTCModel.restore_from(nemo_file, map_location=torch.device(DEVICE))
-    
+
     model.eval()
     if DEVICE == "cuda": model = model.half()
     _cache[name] = model
     return model
 
-# 3. UTILITAIRES
+# 3. UTILS & PIPELINE
 def format_srt_time(sec):
     td = time.gmtime(max(0, sec))
     ms = int((sec - int(sec)) * 1000)
     return f"{time.strftime('%H:%M:%S', td)},{ms:03}"
 
-# 4. PIPELINE
 def pipeline(video_in, model_name):
     tmp_dir = tempfile.mkdtemp()
     try:
-        if not video_in: return "❌ Vidéo manquante", None
+        if not video_in: return "❌ Vidéo absente", None
         
-        # Audio
+        # Audio & Segmentation
+        yield "⏳ Extraction & Segmentation...", None
         full_wav = os.path.join(tmp_dir, "full.wav")
         subprocess.run(f"ffmpeg -y -i {shlex.quote(video_in)} -vn -ac 1 -ar 16000 {full_wav}", shell=True, check=True)
-        
-        # Segmentation
-        yield "⏳ Segmentation...", None
         subprocess.run(f"ffmpeg -i {full_wav} -f segment -segment_time {SEGMENT_DURATION} -c copy {os.path.join(tmp_dir, 'seg_%03d.wav')}", shell=True, check=True)
         
         files = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
-        valid_segments = [f for f in files if os.path.getsize(f) > 1500]
+        valid_segments = [f for f in files if os.path.getsize(f) > 1000]
         
         # Transcription
-        yield f"🎙️ Chargement & Transcription ({model_name})...", None
+        yield f"🎙️ Transcription ({model_name})...", None
         model = get_model(model_name)
         
         with torch.inference_mode():
-            batch_hypotheses = model.transcribe(valid_segments, batch_size=16 if DEVICE=="cuda" else 2, return_hypotheses=True)
+            results = model.transcribe(valid_segments, batch_size=16 if DEVICE=="cuda" else 2, return_hypotheses=True)
         
-        # SRT Generation
-        all_words_ts = []
-        for idx, hyp in enumerate(batch_hypotheses):
-            base_time = idx * SEGMENT_DURATION
+        # SRT
+        all_words = []
+        for idx, hyp in enumerate(results):
             text = hyp.text if hasattr(hyp, 'text') else str(hyp)
             words = text.split()
             if not words: continue
-            gap = SEGMENT_DURATION / len(words)
+            step = SEGMENT_DURATION / len(words)
             for i, w in enumerate(words):
-                all_words_ts.append({"word": w, "start": base_time + (i * gap), "end": base_time + ((i+1) * gap)})
+                all_words.append({"w": w, "s": (idx * SEGMENT_DURATION) + (i * step), "e": (idx * SEGMENT_DURATION) + ((i+1) * step)})
 
-        # Encodage
-        srt_path = os.path.join(tmp_dir, "final.srt")
+        srt_path = os.path.join(tmp_dir, "f.srt")
         with open(srt_path, "w", encoding="utf-8") as f:
-            for i in range(0, len(all_words_ts), 6):
-                chunk = all_words_ts[i:i+6]
-                f.write(f"{(i//6)+1}\n{format_srt_time(chunk[0]['start'])} --> {format_srt_time(chunk[-1]['end'])}\n")
-                f.write(" ".join([c['word'] for c in chunk]) + "\n\n")
+            for i in range(0, len(all_words), 6):
+                c = all_words[i:i+6]
+                f.write(f"{(i//6)+1}\n{format_srt_time(c[0]['s'])} --> {format_srt_time(c[-1]['e'])}\n{' '.join([x['w'] for x in c])}\n\n")
 
-        out_path = os.path.abspath(f"resultat_{int(time.time())}.mp4")
+        # Vidéo
+        out_path = os.path.abspath(f"output_{int(time.time())}.mp4")
+        # Formatage du chemin SRT pour ffmpeg (Linux/Windows)
         safe_srt = srt_path.replace("\\", "/").replace(":", "\\:")
-        cmd = f"ffmpeg -y -i {shlex.quote(video_in)} -vf \"subtitles='{safe_srt}':force_style='Alignment=2,FontSize=18'\" -c:v libx264 -preset ultrafast -c:a copy {out_path}"
+        
+        cmd = f"ffmpeg -y -i {shlex.quote(video_in)} -vf \"subtitles='{safe_srt}'\" -c:v libx264 -preset ultrafast -c:a copy {out_path}"
         subprocess.run(cmd, shell=True, check=True)
         
-        yield "✅ Succès !", out_path
+        yield "✅ Terminé !", out_path
 
     except Exception as e:
         yield f"❌ Erreur : {str(e)}", None
     finally:
         if os.path.exists(tmp_dir): shutil.rmtree(tmp_dir)
 
-# 5. INTERFACE
-with gr.Blocks() as demo:
-    gr.HTML("<h1 style='text-align:center;'>RobotsMali Speech Lab</h1>")
+# 4. INTERFACE
+with gr.Blocks(theme=gr.themes.Default()) as demo:
+    gr.Markdown("# 🤖 RobotsMali Speech Lab")
     with gr.Row():
         with gr.Column():
-            v_input = gr.Video(label="Source")
-            m_input = gr.Dropdown(choices=list(MODELS.keys()), value="Soloni V3 (TDT-CTC)", label="Modèle")
-            run_btn = gr.Button("🚀 GÉNÉRER", variant="primary")
-            if EXAMPLE_PATH: gr.Examples([[EXAMPLE_PATH, "Soloni V3 (TDT-CTC)"]], [v_input, m_input])
+            v_in = gr.Video()
+            m_in = gr.Dropdown(choices=list(MODELS.keys()), value="Soloni V3 (TDT-CTC)")
+            btn = gr.Button("🚀 GÉNÉRER", variant="primary")
+            if EXAMPLE_PATH: gr.Examples([[EXAMPLE_PATH, "Soloni V3 (TDT-CTC)"]], [v_in, m_in])
         with gr.Column():
-            status = gr.Markdown("Prêt.")
-            v_output = gr.Video(label="Résultat")
+            stat = gr.Markdown("Prêt.")
+            v_out = gr.Video()
 
-    run_btn.click(pipeline, [v_input, m_input], [status, v_output])
+    btn.click(pipeline, [v_in, m_in], [stat, v_out])
 
 demo.launch()