Update app.py

app.py

@@ -1,12 +1,15 @@
 # -*- coding: utf-8 -*-
-import os, shlex, subprocess, tempfile, traceback, time, glob, gc, shutil, tarfile
+import os, shlex, subprocess, tempfile, traceback, time, glob, gc, shutil
 import torch
-import yaml
+import logging
 from huggingface_hub import snapshot_download
 from nemo.collections import asr as nemo_asr
 import gradio as gr
 
-# 1. CONFIGURATION
+# Configuration des logs pour voir ce qui se passe sous le capot
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 SEGMENT_DURATION = 5.0 
 
@@ -22,7 +25,6 @@ MODELS = {
     "Traduction Soloni (ST)":    ("RobotsMali/st-soloni-114m-tdt-ctc", "rnnt"),
 }
 
-# --- SECTION EXEMPLE ---
 def find_example_video():
     paths = ["examples/MARALINKE.mp4", "MARALINKE.mp4"]
     for p in paths:
@@ -32,109 +34,130 @@ def find_example_video():
 EXAMPLE_PATH = find_example_video()
 _cache = {}
 
-# 2. CHARGEMENT "HARDCORE" (BYPASS DU BUG INIT)
+# --- CHARGEMENT AVEC LOGS ET BYPASS ---
 def get_model(name):
     if name in _cache: return _cache[name]
-    gc.collect()
-    if torch.cuda.is_available(): torch.cuda.empty_cache()
     
-    repo, _ = MODELS[name]
+    repo, m_type = MODELS[name]
+    print(f"🔍 LOG: Tentative de chargement du modèle: {name}")
+    print(f"🔍 LOG: Repo HF: {repo} | Device: {DEVICE}")
+    
     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)
+    
+    if not nemo_file:
+        print(f"❌ LOG: Erreur - Fichier .nemo introuvable dans {folder}")
+        raise FileNotFoundError("Fichier .nemo manquant.")
 
-    # MÉTHODE DE SECOURS : Chargement par restauration standard
-    # Si ça échoue encore, on utilise une approche par classe explicite
+    # Tentative 1: Standard avec connecteur explicite
     try:
+        print("🔍 LOG: Essai Méthode 1 (Standard Restore)...")
         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))
+        connector = SaveRestoreConnector()
+        model = nemo_asr.models.ASRModel.restore_from(nemo_file, map_location=torch.device(DEVICE), save_restore_connector=connector)
+        print("✅ LOG: Succès avec Méthode 1")
+    
+    except TypeError as e:
+        print(f"⚠️ LOG: Échec Méthode 1 (Erreur init): {e}")
+        # Tentative 2: Forcer la classe selon le type
+        try:
+            print(f"🔍 LOG: Essai Méthode 2 (Forçage Classe {m_type})...")
+            if "ctc" in name.lower() or m_type == "ctc":
+                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))
+            print("✅ LOG: Succès avec Méthode 2")
+        
+        except Exception as e2:
+            print(f"❌ LOG: Échec critique Méthode 2: {e2}")
+            traceback.print_exc()
+            raise RuntimeError(f"Impossible de charger le modèle après 2 tentatives. Erreur: {e2}")
 
     model.eval()
     if DEVICE == "cuda": model = model.half()
     _cache[name] = model
     return model
 
-# 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}"
-
+# --- PIPELINE ---
 def pipeline(video_in, model_name):
     tmp_dir = tempfile.mkdtemp()
+    log_messages = []
+    
+    def add_log(msg):
+        print(f"📋 PIPELINE: {msg}")
+        log_messages.append(msg)
+        return "\n".join(log_messages)
+
     try:
-        if not video_in: return "❌ Vidéo absente", None
-        
-        # Audio & Segmentation
-        yield "⏳ Extraction & Segmentation...", None
+        if not video_in:
+            yield "❌ Vidéo manquante", None
+            return
+
+        yield add_log("Phase 1: Extraction audio..."), 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)
-        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)
+        res = subprocess.run(f"ffmpeg -y -i {shlex.quote(video_in)} -vn -ac 1 -ar 16000 {full_wav}", shell=True, capture_output=True)
+        if res.returncode != 0: raise RuntimeError(f"FFmpeg Audio Error: {res.stderr.decode()}")
+
+        yield add_log(f"Phase 2: Découpage en blocs de {SEGMENT_DURATION}s..."), 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)
         
         files = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
         valid_segments = [f for f in files if os.path.getsize(f) > 1000]
-        
-        # Transcription
-        yield f"🎙️ Transcription ({model_name})...", None
+        yield add_log(f"Segments valides trouvés: {len(valid_segments)}"), None
+
+        yield add_log(f"Phase 3: Initialisation de {model_name}..."), None
         model = get_model(model_name)
         
+        yield add_log("Phase 4: Transcription en cours..."), None
         with torch.inference_mode():
-            results = model.transcribe(valid_segments, batch_size=16 if DEVICE=="cuda" else 2, return_hypotheses=True)
+            batch_hyp = model.transcribe(valid_segments, batch_size=8, return_hypotheses=True)
         
-        # SRT
+        # Traitement SRT
         all_words = []
-        for idx, hyp in enumerate(results):
+        for idx, hyp in enumerate(batch_hyp):
             text = hyp.text if hasattr(hyp, 'text') else str(hyp)
             words = text.split()
             if not words: continue
-            step = SEGMENT_DURATION / len(words)
+            gap = SEGMENT_DURATION / len(words)
             for i, w in enumerate(words):
-                all_words.append({"w": w, "s": (idx * SEGMENT_DURATION) + (i * step), "e": (idx * SEGMENT_DURATION) + ((i+1) * step)})
+                all_words.append({"w": w, "s": (idx * SEGMENT_DURATION) + (i * gap), "e": (idx * SEGMENT_DURATION) + ((i+1) * gap)})
 
-        srt_path = os.path.join(tmp_dir, "f.srt")
+        yield add_log("Phase 5: Création de la vidéo finale..."), None
+        srt_path = os.path.join(tmp_dir, "sub.srt")
         with open(srt_path, "w", encoding="utf-8") as f:
             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")
+                chunk = all_words[i:i+6]
+                start_f = time.strftime('%H:%M:%S', time.gmtime(chunk[0]['s'])) + f",{int((chunk[0]['s']%1)*1000):03d}"
+                end_f = time.strftime('%H:%M:%S', time.gmtime(chunk[-1]['e'])) + f",{int((chunk[-1]['e']%1)*1000):03d}"
+                f.write(f"{(i//6)+1}\n{start_f} --> {end_f}\n{' '.join([x['w'] for x in chunk])}\n\n")
 
-        # Vidéo
-        out_path = os.path.abspath(f"output_{int(time.time())}.mp4")
-        # Formatage du chemin SRT pour ffmpeg (Linux/Windows)
+        out_path = os.path.abspath(f"result_{int(time.time())}.mp4")
         safe_srt = srt_path.replace("\\", "/").replace(":", "\\:")
+        subprocess.run(f"ffmpeg -y -i {shlex.quote(video_in)} -vf \"subtitles='{safe_srt}'\" -c:v libx264 -preset superfast -c:a copy {out_path}", shell=True, check=True)
         
-        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 "✅ Terminé !", out_path
+        yield add_log("✅ Terminé avec succès !"), out_path
 
     except Exception as e:
-        yield f"❌ Erreur : {str(e)}", None
+        err_msg = f"❌ ERREUR: {str(e)}\n{traceback.format_exc()}"
+        print(err_msg)
+        yield add_log(err_msg), None
     finally:
         if os.path.exists(tmp_dir): shutil.rmtree(tmp_dir)
 
-# 4. INTERFACE
-with gr.Blocks(theme=gr.themes.Default()) as demo:
-    gr.Markdown("# 🤖 RobotsMali Speech Lab")
+# --- INTERFACE ---
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# 🚀 RobotsMali Speech Lab (Debug Mode)")
     with gr.Row():
         with gr.Column():
-            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])
+            v_input = gr.Video(label="Vidéo")
+            m_input = gr.Dropdown(choices=list(MODELS.keys()), value="Soloni V3 (TDT-CTC)", label="Modèle")
+            run_btn = gr.Button("DÉMARRER LA TRANSCRIPTION", variant="primary")
+            if EXAMPLE_PATH:
+                gr.Examples([[EXAMPLE_PATH, "Soloni V3 (TDT-CTC)"]], [v_input, m_input])
         with gr.Column():
-            stat = gr.Markdown("Prêt.")
-            v_out = gr.Video()
+            status_box = gr.Textbox(label="Logs d'exécution", lines=10, interactive=False)
+            v_output = gr.Video(label="Résultat")
 
-    btn.click(pipeline, [v_in, m_in], [stat, v_out])
+    run_btn.click(pipeline, [v_input, m_input], [status_box, v_output])
 
-demo.launch()
+demo.launch(debug=True)