Update app.py
Browse files
app.py
CHANGED
|
@@ -8,7 +8,7 @@ import gradio as gr
|
|
| 8 |
# 1. CONFIGURATION MATÉRIEL
|
| 9 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 10 |
|
| 11 |
-
# 2. DICTIONNAIRE DES MODÈLES
|
| 12 |
MODELS = {
|
| 13 |
"Soloba V3 (CTC)": ("RobotsMali/soloba-ctc-0.6b-v3", "ctc"),
|
| 14 |
"Soloba V1.5 (TDT)": ("RobotsMali/soloba-tdt-0.6b-v1.5", "rnnt"),
|
|
@@ -21,19 +21,18 @@ MODELS = {
|
|
| 21 |
_cache = {}
|
| 22 |
|
| 23 |
def clear_memory():
|
| 24 |
-
"""
|
| 25 |
_cache.clear()
|
| 26 |
gc.collect()
|
| 27 |
if torch.cuda.is_available():
|
| 28 |
torch.cuda.empty_cache()
|
| 29 |
|
| 30 |
-
def
|
| 31 |
-
"""Charge le modèle
|
| 32 |
-
if name in _cache:
|
|
|
|
| 33 |
|
| 34 |
-
yield f"⏳ Chargement du modèle {name}..."
|
| 35 |
clear_memory()
|
| 36 |
-
|
| 37 |
repo, mode = MODELS[name]
|
| 38 |
folder = snapshot_download(repo, local_dir_use_symlinks=False)
|
| 39 |
nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
|
|
@@ -45,7 +44,7 @@ def load_model(name):
|
|
| 45 |
|
| 46 |
model.to(DEVICE).eval()
|
| 47 |
if DEVICE == "cuda":
|
| 48 |
-
model.half()
|
| 49 |
|
| 50 |
_cache[name] = model
|
| 51 |
return model
|
|
@@ -58,37 +57,35 @@ def format_srt_time(sec):
|
|
| 58 |
def pipeline(video_in, model_name):
|
| 59 |
tmp_dir = tempfile.mkdtemp()
|
| 60 |
try:
|
| 61 |
-
if not video_in:
|
|
|
|
| 62 |
|
| 63 |
# A. Extraction Audio
|
| 64 |
-
yield "⏳ Extraction de l'audio...", None
|
| 65 |
full_wav = os.path.join(tmp_dir, "full.wav")
|
| 66 |
subprocess.run(f"ffmpeg -y -i {shlex.quote(video_in)} -vn -ac 1 -ar 16000 {full_wav}", shell=True, check=True)
|
| 67 |
|
| 68 |
-
# B. Segmentation
|
| 69 |
segment_pattern = os.path.join(tmp_dir, "seg_%03d.wav")
|
| 70 |
subprocess.run(f"ffmpeg -i {full_wav} -f segment -segment_time 10 -c copy {segment_pattern}", shell=True, check=True)
|
| 71 |
audio_segments = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
|
| 72 |
|
| 73 |
-
# C. Chargement
|
| 74 |
-
|
| 75 |
-
model =
|
| 76 |
-
for update in model_gen:
|
| 77 |
-
if isinstance(update, str): yield update, None, None
|
| 78 |
-
else: model = update
|
| 79 |
|
| 80 |
-
#
|
| 81 |
stride = 0.02
|
| 82 |
if hasattr(model, 'preprocessor') and hasattr(model.preprocessor, 'featurizer'):
|
| 83 |
hop = model.preprocessor.featurizer.hop_length
|
| 84 |
sr = model.preprocessor.featurizer.sample_rate
|
| 85 |
stride = hop / sr
|
| 86 |
|
| 87 |
-
# D. Transcription
|
| 88 |
all_words_ts = []
|
| 89 |
for idx, seg_path in enumerate(audio_segments):
|
| 90 |
base_time = idx * 10.0
|
| 91 |
-
yield f"⏳ IA : Transcription
|
| 92 |
|
| 93 |
hyp = model.transcribe([seg_path], return_hypotheses=True)[0]
|
| 94 |
offsets = getattr(hyp, 'word_offsets', None)
|
|
@@ -99,12 +96,11 @@ def pipeline(video_in, model_name):
|
|
| 99 |
start_t = base_time + (offsets[i] * stride)
|
| 100 |
all_words_ts.append({"word": word, "start": start_t, "end": start_t + 0.45})
|
| 101 |
else:
|
| 102 |
-
# Fallback : Répartition linéaire si les offsets manquent
|
| 103 |
gap = 10.0 / max(len(words), 1)
|
| 104 |
for i, w in enumerate(words):
|
| 105 |
all_words_ts.append({"word": w, "start": base_time + (i * gap), "end": base_time + ((i+1) * gap)})
|
| 106 |
|
| 107 |
-
# E. Génération
|
| 108 |
srt_path = os.path.join(tmp_dir, "final.srt")
|
| 109 |
words_per_line = 6
|
| 110 |
with open(srt_path, "w", encoding="utf-8") as f:
|
|
@@ -114,11 +110,10 @@ def pipeline(video_in, model_name):
|
|
| 114 |
f.write(f"{format_srt_time(chunk[0]['start'])} --> {format_srt_time(chunk[-1]['end'])}\n")
|
| 115 |
f.write(" ".join([c['word'] for c in chunk]) + "\n\n")
|
| 116 |
|
| 117 |
-
# F. Encodage Vidéo
|
| 118 |
-
yield "⏳ Rendu vidéo final
|
| 119 |
out_path = os.path.abspath(f"robotsmali_final_{int(time.time())}.mp4")
|
| 120 |
|
| 121 |
-
# Protection des chemins pour FFmpeg (indispensable pour Windows/Linux)
|
| 122 |
safe_srt = srt_path.replace("\\", "/").replace(":", "\\:")
|
| 123 |
cmd_ffmpeg = (
|
| 124 |
f"ffmpeg -y -i {shlex.quote(video_in)} "
|
|
@@ -127,28 +122,27 @@ def pipeline(video_in, model_name):
|
|
| 127 |
)
|
| 128 |
subprocess.run(cmd_ffmpeg, shell=True, check=True)
|
| 129 |
|
| 130 |
-
yield "✅
|
| 131 |
|
| 132 |
except Exception as e:
|
| 133 |
traceback.print_exc()
|
| 134 |
-
yield f"❌ Erreur : {str(e)}", None
|
| 135 |
|
| 136 |
-
# 6. INTERFACE
|
| 137 |
-
with gr.Blocks(theme=gr.themes.Soft()
|
| 138 |
-
gr.HTML("<h1 style='text-align:center; color:#EAB308;'>🤖 ROBOTSMALI TRANSCRIPTION
|
| 139 |
|
| 140 |
with gr.Row():
|
| 141 |
-
with gr.Column(
|
| 142 |
-
v_in = gr.Video(label="
|
| 143 |
-
m_sel = gr.Dropdown(choices=list(MODELS.keys()), value="Soloba V3 (CTC)", label="
|
| 144 |
-
btn_run = gr.Button("🚀 GÉNÉRER
|
| 145 |
|
| 146 |
-
with gr.Column(
|
| 147 |
-
status = gr.Markdown("### État\nPrêt
|
| 148 |
-
v_out = gr.Video(label="Vidéo Finale
|
| 149 |
-
f_srt = gr.File(label="Fichier Sous-titres (.SRT)")
|
| 150 |
|
| 151 |
-
btn_run.click(pipeline, [v_in, m_sel], [status, v_out
|
| 152 |
|
| 153 |
if __name__ == "__main__":
|
| 154 |
-
demo.launch(debug=True
|
|
|
|
| 8 |
# 1. CONFIGURATION MATÉRIEL
|
| 9 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 10 |
|
| 11 |
+
# 2. DICTIONNAIRE DES MODÈLES
|
| 12 |
MODELS = {
|
| 13 |
"Soloba V3 (CTC)": ("RobotsMali/soloba-ctc-0.6b-v3", "ctc"),
|
| 14 |
"Soloba V1.5 (TDT)": ("RobotsMali/soloba-tdt-0.6b-v1.5", "rnnt"),
|
|
|
|
| 21 |
_cache = {}
|
| 22 |
|
| 23 |
def clear_memory():
|
| 24 |
+
"""Libère la VRAM et la RAM."""
|
| 25 |
_cache.clear()
|
| 26 |
gc.collect()
|
| 27 |
if torch.cuda.is_available():
|
| 28 |
torch.cuda.empty_cache()
|
| 29 |
|
| 30 |
+
def get_model(name):
|
| 31 |
+
"""Charge le modèle et le retourne directement."""
|
| 32 |
+
if name in _cache:
|
| 33 |
+
return _cache[name]
|
| 34 |
|
|
|
|
| 35 |
clear_memory()
|
|
|
|
| 36 |
repo, mode = MODELS[name]
|
| 37 |
folder = snapshot_download(repo, local_dir_use_symlinks=False)
|
| 38 |
nemo_file = next((os.path.join(folder, f) for f in os.listdir(folder) if f.endswith(".nemo")), None)
|
|
|
|
| 44 |
|
| 45 |
model.to(DEVICE).eval()
|
| 46 |
if DEVICE == "cuda":
|
| 47 |
+
model.half()
|
| 48 |
|
| 49 |
_cache[name] = model
|
| 50 |
return model
|
|
|
|
| 57 |
def pipeline(video_in, model_name):
|
| 58 |
tmp_dir = tempfile.mkdtemp()
|
| 59 |
try:
|
| 60 |
+
if not video_in:
|
| 61 |
+
return "❌ Source vide", None
|
| 62 |
|
| 63 |
# A. Extraction Audio
|
| 64 |
+
yield "⏳ Extraction de l'audio...", None
|
| 65 |
full_wav = os.path.join(tmp_dir, "full.wav")
|
| 66 |
subprocess.run(f"ffmpeg -y -i {shlex.quote(video_in)} -vn -ac 1 -ar 16000 {full_wav}", shell=True, check=True)
|
| 67 |
|
| 68 |
+
# B. Segmentation (10s)
|
| 69 |
segment_pattern = os.path.join(tmp_dir, "seg_%03d.wav")
|
| 70 |
subprocess.run(f"ffmpeg -i {full_wav} -f segment -segment_time 10 -c copy {segment_pattern}", shell=True, check=True)
|
| 71 |
audio_segments = sorted(glob.glob(os.path.join(tmp_dir, "seg_*.wav")))
|
| 72 |
|
| 73 |
+
# C. Chargement du modèle
|
| 74 |
+
yield f"⏳ Chargement du modèle {model_name}...", None
|
| 75 |
+
model = get_model(model_name)
|
|
|
|
|
|
|
|
|
|
| 76 |
|
| 77 |
+
# Facteur de temps dynamique
|
| 78 |
stride = 0.02
|
| 79 |
if hasattr(model, 'preprocessor') and hasattr(model.preprocessor, 'featurizer'):
|
| 80 |
hop = model.preprocessor.featurizer.hop_length
|
| 81 |
sr = model.preprocessor.featurizer.sample_rate
|
| 82 |
stride = hop / sr
|
| 83 |
|
| 84 |
+
# D. Transcription
|
| 85 |
all_words_ts = []
|
| 86 |
for idx, seg_path in enumerate(audio_segments):
|
| 87 |
base_time = idx * 10.0
|
| 88 |
+
yield f"⏳ IA : Transcription {idx+1}/{len(audio_segments)}...", None
|
| 89 |
|
| 90 |
hyp = model.transcribe([seg_path], return_hypotheses=True)[0]
|
| 91 |
offsets = getattr(hyp, 'word_offsets', None)
|
|
|
|
| 96 |
start_t = base_time + (offsets[i] * stride)
|
| 97 |
all_words_ts.append({"word": word, "start": start_t, "end": start_t + 0.45})
|
| 98 |
else:
|
|
|
|
| 99 |
gap = 10.0 / max(len(words), 1)
|
| 100 |
for i, w in enumerate(words):
|
| 101 |
all_words_ts.append({"word": w, "start": base_time + (i * gap), "end": base_time + ((i+1) * gap)})
|
| 102 |
|
| 103 |
+
# E. Génération SRT (Fichier temporaire interne)
|
| 104 |
srt_path = os.path.join(tmp_dir, "final.srt")
|
| 105 |
words_per_line = 6
|
| 106 |
with open(srt_path, "w", encoding="utf-8") as f:
|
|
|
|
| 110 |
f.write(f"{format_srt_time(chunk[0]['start'])} --> {format_srt_time(chunk[-1]['end'])}\n")
|
| 111 |
f.write(" ".join([c['word'] for c in chunk]) + "\n\n")
|
| 112 |
|
| 113 |
+
# F. Encodage Vidéo
|
| 114 |
+
yield "⏳ Rendu vidéo final...", None
|
| 115 |
out_path = os.path.abspath(f"robotsmali_final_{int(time.time())}.mp4")
|
| 116 |
|
|
|
|
| 117 |
safe_srt = srt_path.replace("\\", "/").replace(":", "\\:")
|
| 118 |
cmd_ffmpeg = (
|
| 119 |
f"ffmpeg -y -i {shlex.quote(video_in)} "
|
|
|
|
| 122 |
)
|
| 123 |
subprocess.run(cmd_ffmpeg, shell=True, check=True)
|
| 124 |
|
| 125 |
+
yield "✅ Terminé !", out_path
|
| 126 |
|
| 127 |
except Exception as e:
|
| 128 |
traceback.print_exc()
|
| 129 |
+
yield f"❌ Erreur : {str(e)}", None
|
| 130 |
|
| 131 |
+
# 6. INTERFACE
|
| 132 |
+
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
| 133 |
+
gr.HTML("<h1 style='text-align:center; color:#EAB308;'>🤖 ROBOTSMALI TRANSCRIPTION</h1>")
|
| 134 |
|
| 135 |
with gr.Row():
|
| 136 |
+
with gr.Column():
|
| 137 |
+
v_in = gr.Video(label="Source")
|
| 138 |
+
m_sel = gr.Dropdown(choices=list(MODELS.keys()), value="Soloba V3 (CTC)", label="Modèle IA")
|
| 139 |
+
btn_run = gr.Button("🚀 GÉNÉRER", variant="primary")
|
| 140 |
|
| 141 |
+
with gr.Column():
|
| 142 |
+
status = gr.Markdown("### État\nPrêt.")
|
| 143 |
+
v_out = gr.Video(label="Vidéo Finale")
|
|
|
|
| 144 |
|
| 145 |
+
btn_run.click(pipeline, [v_in, m_sel], [status, v_out])
|
| 146 |
|
| 147 |
if __name__ == "__main__":
|
| 148 |
+
demo.launch(debug=True)
|