Update app.py
Browse files
app.py
CHANGED
|
@@ -77,8 +77,10 @@ def call_peace_judge(audio_top, vlm_desc):
|
|
| 77 |
# ==========================================
|
| 78 |
@spaces.GPU(duration=120)
|
| 79 |
def analyze_cat_v12_final(video_path):
|
| 80 |
-
if not video_path:
|
| 81 |
-
|
|
|
|
|
|
|
| 82 |
|
| 83 |
tmp_audio = f"temp_{os.getpid()}.wav"
|
| 84 |
start_total = time.time()
|
|
@@ -91,7 +93,8 @@ def analyze_cat_v12_final(video_path):
|
|
| 91 |
if clip.audio:
|
| 92 |
clip.audio.write_audiofile(tmp_audio, fps=16000, logger=None)
|
| 93 |
w, _ = librosa.load(tmp_audio, sr=16000, duration=5.0)
|
| 94 |
-
if len(w) < 48000:
|
|
|
|
| 95 |
mel = librosa.feature.melspectrogram(y=w, sr=16000, n_mels=192)
|
| 96 |
mel_db = (librosa.power_to_db(mel, ref=np.max) + 40) / 40
|
| 97 |
img = cv2.resize((np.vstack([mel_db, np.zeros((10, mel_db.shape[1]))]) * 255).astype(np.uint8), (224, 224))
|
|
@@ -105,7 +108,7 @@ def analyze_cat_v12_final(video_path):
|
|
| 105 |
clip.close()
|
| 106 |
t_audio = time.time() - t_0
|
| 107 |
|
| 108 |
-
# --- B. VISION (
|
| 109 |
t_1 = time.time()
|
| 110 |
|
| 111 |
vlm_prompt = (
|
|
@@ -114,14 +117,11 @@ def analyze_cat_v12_final(video_path):
|
|
| 114 |
"Based on this, what is the cat's mood?"
|
| 115 |
)
|
| 116 |
|
| 117 |
-
# On définit le message
|
| 118 |
messages = [{"role": "user", "content": [{"type": "video", "path": video_path}, {"type": "text", "text": vlm_prompt}]}]
|
| 119 |
|
| 120 |
-
# LA SOLUTION : On utilise le processeur pour transformer les messages en tenseurs directement.
|
| 121 |
-
# Cela injecte automatiquement le bon nombre de jetons vidéo (93) dans le prompt.
|
| 122 |
vlm_inputs = vlm_proc.apply_chat_template(
|
| 123 |
messages,
|
| 124 |
-
add_generation_prompt=True,
|
| 125 |
tokenize=True,
|
| 126 |
return_dict=True,
|
| 127 |
return_tensors="pt"
|
|
@@ -132,25 +132,13 @@ def analyze_cat_v12_final(video_path):
|
|
| 132 |
|
| 133 |
vlm_res = vlm_proc.batch_decode(vlm_out, skip_special_tokens=True)[0]
|
| 134 |
|
| 135 |
-
#
|
| 136 |
if "assistant" in vlm_res.lower():
|
| 137 |
vlm_clean = vlm_res.split("assistant")[-1].strip()
|
| 138 |
else:
|
| 139 |
-
vlm_clean = vlm_res.strip()
|
| 140 |
|
| 141 |
t_vlm = time.time() - t_1
|
| 142 |
-
|
| 143 |
-
# On appelle le processeur directement avec le texte brut
|
| 144 |
-
vlm_inputs = vlm_proc(text=vlm_prompt, videos=video_path, return_tensors="pt").to(DEVICE)
|
| 145 |
-
|
| 146 |
-
with torch.no_grad():
|
| 147 |
-
vlm_out = vlm_model.generate(**vlm_inputs, max_new_tokens=100, do_sample=False)
|
| 148 |
-
|
| 149 |
-
vlm_res = vlm_proc.batch_decode(vlm_out, skip_special_tokens=True)[0]
|
| 150 |
-
|
| 151 |
-
# On nettoie juste le prompt de l'affichage
|
| 152 |
-
vlm_clean = vlm_res.replace(vlm_prompt, "").strip()
|
| 153 |
-
t_vlm = time.time() - t_1
|
| 154 |
|
| 155 |
# --- C. JUGE ---
|
| 156 |
t_2 = time.time()
|
|
@@ -161,7 +149,12 @@ def analyze_cat_v12_final(video_path):
|
|
| 161 |
|
| 162 |
# --- D. VISUELS ---
|
| 163 |
top5 = np.argsort(audio_probs)[-5:][::-1]
|
| 164 |
-
fig = px.bar(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
# --- E. RAPPORT ---
|
| 167 |
t_total = time.time() - start_total
|
|
@@ -171,11 +164,13 @@ def analyze_cat_v12_final(video_path):
|
|
| 171 |
📊 AUDIO : {audio_ctx}
|
| 172 |
⏱️ TEMPS : Audio {t_audio:.2f}s | Vision {t_vlm:.2f}s | Total {t_total:.2f}s"""
|
| 173 |
|
| 174 |
-
if os.path.exists(tmp_audio):
|
|
|
|
| 175 |
return report, fig
|
| 176 |
|
| 177 |
except Exception as e:
|
| 178 |
-
if os.path.exists(tmp_audio):
|
|
|
|
| 179 |
return f"❌ Erreur : {str(e)}", None
|
| 180 |
|
| 181 |
# --- Interface Gradio ---
|
|
|
|
| 77 |
# ==========================================
|
| 78 |
@spaces.GPU(duration=120)
|
| 79 |
def analyze_cat_v12_final(video_path):
|
| 80 |
+
if not video_path:
|
| 81 |
+
return "❌ Aucune vidéo.", None
|
| 82 |
+
if torch.cuda.is_available():
|
| 83 |
+
torch.cuda.empty_cache()
|
| 84 |
|
| 85 |
tmp_audio = f"temp_{os.getpid()}.wav"
|
| 86 |
start_total = time.time()
|
|
|
|
| 93 |
if clip.audio:
|
| 94 |
clip.audio.write_audiofile(tmp_audio, fps=16000, logger=None)
|
| 95 |
w, _ = librosa.load(tmp_audio, sr=16000, duration=5.0)
|
| 96 |
+
if len(w) < 48000:
|
| 97 |
+
w = np.pad(w, (0, 48000 - len(w)))
|
| 98 |
mel = librosa.feature.melspectrogram(y=w, sr=16000, n_mels=192)
|
| 99 |
mel_db = (librosa.power_to_db(mel, ref=np.max) + 40) / 40
|
| 100 |
img = cv2.resize((np.vstack([mel_db, np.zeros((10, mel_db.shape[1]))]) * 255).astype(np.uint8), (224, 224))
|
|
|
|
| 108 |
clip.close()
|
| 109 |
t_audio = time.time() - t_0
|
| 110 |
|
| 111 |
+
# --- B. VISION (CORRIGÉ : une seule fois, via apply_chat_template) ---
|
| 112 |
t_1 = time.time()
|
| 113 |
|
| 114 |
vlm_prompt = (
|
|
|
|
| 117 |
"Based on this, what is the cat's mood?"
|
| 118 |
)
|
| 119 |
|
|
|
|
| 120 |
messages = [{"role": "user", "content": [{"type": "video", "path": video_path}, {"type": "text", "text": vlm_prompt}]}]
|
| 121 |
|
|
|
|
|
|
|
| 122 |
vlm_inputs = vlm_proc.apply_chat_template(
|
| 123 |
messages,
|
| 124 |
+
add_generation_prompt=True,
|
| 125 |
tokenize=True,
|
| 126 |
return_dict=True,
|
| 127 |
return_tensors="pt"
|
|
|
|
| 132 |
|
| 133 |
vlm_res = vlm_proc.batch_decode(vlm_out, skip_special_tokens=True)[0]
|
| 134 |
|
| 135 |
+
# Nettoyage robuste de la réponse
|
| 136 |
if "assistant" in vlm_res.lower():
|
| 137 |
vlm_clean = vlm_res.split("assistant")[-1].strip()
|
| 138 |
else:
|
| 139 |
+
vlm_clean = vlm_res.replace(vlm_prompt, "").strip()
|
| 140 |
|
| 141 |
t_vlm = time.time() - t_1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 142 |
|
| 143 |
# --- C. JUGE ---
|
| 144 |
t_2 = time.time()
|
|
|
|
| 149 |
|
| 150 |
# --- D. VISUELS ---
|
| 151 |
top5 = np.argsort(audio_probs)[-5:][::-1]
|
| 152 |
+
fig = px.bar(
|
| 153 |
+
x=[audio_probs[i]*100 for i in top5],
|
| 154 |
+
y=[CATEGORIES[i].upper() for i in top5],
|
| 155 |
+
orientation='h',
|
| 156 |
+
title='Scores Audio'
|
| 157 |
+
)
|
| 158 |
|
| 159 |
# --- E. RAPPORT ---
|
| 160 |
t_total = time.time() - start_total
|
|
|
|
| 164 |
📊 AUDIO : {audio_ctx}
|
| 165 |
⏱️ TEMPS : Audio {t_audio:.2f}s | Vision {t_vlm:.2f}s | Total {t_total:.2f}s"""
|
| 166 |
|
| 167 |
+
if os.path.exists(tmp_audio):
|
| 168 |
+
os.remove(tmp_audio)
|
| 169 |
return report, fig
|
| 170 |
|
| 171 |
except Exception as e:
|
| 172 |
+
if os.path.exists(tmp_audio):
|
| 173 |
+
os.remove(tmp_audio)
|
| 174 |
return f"❌ Erreur : {str(e)}", None
|
| 175 |
|
| 176 |
# --- Interface Gradio ---
|