Update app.py
Browse files
app.py
CHANGED
|
@@ -5,15 +5,10 @@ import google.generativeai as genai
|
|
| 5 |
from ultralytics import YOLO
|
| 6 |
import tempfile
|
| 7 |
import torch
|
|
|
|
| 8 |
|
| 9 |
# =============================
|
| 10 |
-
#
|
| 11 |
-
# =============================
|
| 12 |
-
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
| 13 |
-
print(f"🚀 Using device: {DEVICE}")
|
| 14 |
-
|
| 15 |
-
# =============================
|
| 16 |
-
# إعداد مفتاح Gemini (مكتوب صريح في الكود)
|
| 17 |
# =============================
|
| 18 |
|
| 19 |
GEMINI_API_KEY = "AIzaSyAvm28ZnTMaZ1Jtg9sYM-EO4qlAN2W4BIQ"
|
|
@@ -31,7 +26,6 @@ SYSTEM_PROMPT = (
|
|
| 31 |
)
|
| 32 |
|
| 33 |
|
| 34 |
-
|
| 35 |
def fix_with_gemini(raw_text: str) -> str:
|
| 36 |
if not raw_text:
|
| 37 |
return ""
|
|
@@ -45,18 +39,18 @@ def fix_with_gemini(raw_text: str) -> str:
|
|
| 45 |
|
| 46 |
|
| 47 |
# =============================
|
| 48 |
-
# إعدادات YOLO +
|
| 49 |
# =============================
|
| 50 |
|
| 51 |
WEIGHTS_PATH = "best.pt"
|
| 52 |
IMG_SIZE = 320
|
| 53 |
-
CONF_THRESHOLD = 0.25 #
|
| 54 |
|
| 55 |
# إعدادات تجميع الحروف
|
| 56 |
-
MIN_STABLE_FRAMES = 1
|
| 57 |
-
FRAME_SKIP = 1
|
| 58 |
-
MAX_FRAMES = 1000
|
| 59 |
-
WORD_GAP_FRAMES = 10
|
| 60 |
|
| 61 |
arabic_map = {
|
| 62 |
"aleff": "ا",
|
|
@@ -93,21 +87,41 @@ arabic_map = {
|
|
| 93 |
"la": "لا",
|
| 94 |
}
|
| 95 |
|
| 96 |
-
|
| 97 |
-
|
|
|
|
| 98 |
|
| 99 |
-
# ننقل الموديل إلى كرت الشاشة لو موجود
|
| 100 |
-
try:
|
| 101 |
-
model.to(DEVICE)
|
| 102 |
-
print("✅ YOLO model moved to", DEVICE)
|
| 103 |
-
except Exception as e:
|
| 104 |
-
print("⚠️ تعذر نقل الموديل إلى الجهاز:", e)
|
| 105 |
|
| 106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
|
| 108 |
|
| 109 |
# =============================
|
| 110 |
-
# ضغط الفيديو قبل المعالجة
|
| 111 |
# =============================
|
| 112 |
|
| 113 |
def preprocess_video(input_path: str, target_width: int = 640, target_fps: int = 8) -> str:
|
|
@@ -131,7 +145,6 @@ def preprocess_video(input_path: str, target_width: int = 640, target_fps: int =
|
|
| 131 |
frame_step = max(1, int(round(orig_fps / target_fps)))
|
| 132 |
out_fps = orig_fps / frame_step
|
| 133 |
|
| 134 |
-
# ارتفاع النسخة 360p تقريباً حسب نسبة الأبعاد
|
| 135 |
target_height = int(target_width * h / w)
|
| 136 |
|
| 137 |
fd, tmp_path = tempfile.mkstemp(suffix=".mp4")
|
|
@@ -146,7 +159,6 @@ def preprocess_video(input_path: str, target_width: int = 640, target_fps: int =
|
|
| 146 |
if not ret:
|
| 147 |
break
|
| 148 |
|
| 149 |
-
# نأخذ كل frame_step فريم واحد فقط
|
| 150 |
if frame_idx % frame_step == 0:
|
| 151 |
resized = cv2.resize(frame, (target_width, target_height), interpolation=cv2.INTER_AREA)
|
| 152 |
out.write(resized)
|
|
@@ -160,17 +172,19 @@ def preprocess_video(input_path: str, target_width: int = 640, target_fps: int =
|
|
| 160 |
|
| 161 |
|
| 162 |
# =============================
|
| 163 |
-
# معالجة فريم واحد
|
| 164 |
# =============================
|
| 165 |
|
| 166 |
def detect_frame(frame_bgr):
|
|
|
|
|
|
|
| 167 |
frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
|
| 168 |
result = model.predict(
|
| 169 |
frame_rgb,
|
| 170 |
conf=CONF_THRESHOLD,
|
| 171 |
imgsz=IMG_SIZE,
|
| 172 |
verbose=False,
|
| 173 |
-
device=DEVICE #
|
| 174 |
)[0]
|
| 175 |
|
| 176 |
boxes = result.boxes
|
|
@@ -222,7 +236,7 @@ def extract_and_render(video_path: str):
|
|
| 222 |
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
| 223 |
|
| 224 |
if fps <= 0:
|
| 225 |
-
fps = 8.0
|
| 226 |
|
| 227 |
out = cv2.VideoWriter(out_path, fourcc, fps, (width, height))
|
| 228 |
|
|
@@ -285,16 +299,16 @@ def extract_and_render(video_path: str):
|
|
| 285 |
|
| 286 |
|
| 287 |
# =============================
|
| 288 |
-
# Gradio
|
| 289 |
# =============================
|
| 290 |
|
|
|
|
| 291 |
def run(file):
|
| 292 |
if file is None:
|
| 293 |
return "لم يتم رفع فيديو", "", None
|
| 294 |
|
| 295 |
video_path = file.name
|
| 296 |
|
| 297 |
-
# خطوة تسريع الفيديو قبل التحليل (360p + ~8fps)
|
| 298 |
light_path = preprocess_video(video_path, target_width=640, target_fps=8)
|
| 299 |
|
| 300 |
raw, processed_path = extract_and_render(light_path)
|
|
@@ -307,7 +321,7 @@ def run(file):
|
|
| 307 |
|
| 308 |
|
| 309 |
with gr.Blocks() as demo:
|
| 310 |
-
gr.Markdown("## 🤟 ASL → Arabic (YOLO + Gemini)
|
| 311 |
|
| 312 |
inp = gr.File(label="ارفع فيديو الإشارة")
|
| 313 |
raw = gr.Textbox(label="النص الخام", lines=3)
|
|
|
|
| 5 |
from ultralytics import YOLO
|
| 6 |
import tempfile
|
| 7 |
import torch
|
| 8 |
+
import spaces # مهم لـ @spaces.GPU
|
| 9 |
|
| 10 |
# =============================
|
| 11 |
+
# إعداد مفتاح Gemini (حطه هنا)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
# =============================
|
| 13 |
|
| 14 |
GEMINI_API_KEY = "AIzaSyAvm28ZnTMaZ1Jtg9sYM-EO4qlAN2W4BIQ"
|
|
|
|
| 26 |
)
|
| 27 |
|
| 28 |
|
|
|
|
| 29 |
def fix_with_gemini(raw_text: str) -> str:
|
| 30 |
if not raw_text:
|
| 31 |
return ""
|
|
|
|
| 39 |
|
| 40 |
|
| 41 |
# =============================
|
| 42 |
+
# إعدادات YOLO + متغيرات عامة
|
| 43 |
# =============================
|
| 44 |
|
| 45 |
WEIGHTS_PATH = "best.pt"
|
| 46 |
IMG_SIZE = 320
|
| 47 |
+
CONF_THRESHOLD = 0.25 # تخفيض للسهولة
|
| 48 |
|
| 49 |
# إعدادات تجميع الحروف
|
| 50 |
+
MIN_STABLE_FRAMES = 1
|
| 51 |
+
FRAME_SKIP = 1
|
| 52 |
+
MAX_FRAMES = 1000
|
| 53 |
+
WORD_GAP_FRAMES = 10
|
| 54 |
|
| 55 |
arabic_map = {
|
| 56 |
"aleff": "ا",
|
|
|
|
| 87 |
"la": "لا",
|
| 88 |
}
|
| 89 |
|
| 90 |
+
# هنستخدم موديل عالمي لكن نحمّله عند أول استخدام فقط
|
| 91 |
+
yolo_model = None
|
| 92 |
+
DEVICE = "cpu"
|
| 93 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
|
| 95 |
+
def get_model():
|
| 96 |
+
"""
|
| 97 |
+
يحمّل YOLO مرة واحدة، ويحاول نقله لـ GPU لو متوفر.
|
| 98 |
+
يُستدعى داخل دالة عليها @spaces.GPU بعد ما الـ GPU يشتغل فعليًا.
|
| 99 |
+
"""
|
| 100 |
+
global yolo_model, DEVICE
|
| 101 |
+
|
| 102 |
+
if yolo_model is None:
|
| 103 |
+
print("🔹 Loading YOLO model...")
|
| 104 |
+
yolo_model = YOLO(WEIGHTS_PATH)
|
| 105 |
+
print("📚 Classes:", yolo_model.names)
|
| 106 |
+
|
| 107 |
+
# نعيد التحقق من CUDA هنا (بعد ما GPU يشتغل في Spaces)
|
| 108 |
+
if torch.cuda.is_available():
|
| 109 |
+
if DEVICE != "cuda":
|
| 110 |
+
DEVICE = "cuda"
|
| 111 |
+
try:
|
| 112 |
+
yolo_model.to(DEVICE)
|
| 113 |
+
print("✅ YOLO model moved to cuda")
|
| 114 |
+
except Exception as e:
|
| 115 |
+
print("⚠️ تعذر نقل الموديل إلى cuda:", e)
|
| 116 |
+
else:
|
| 117 |
+
DEVICE = "cpu"
|
| 118 |
+
print("⚠️ CUDA غير متوفر، سيتم استخدام CPU.")
|
| 119 |
+
|
| 120 |
+
return yolo_model
|
| 121 |
|
| 122 |
|
| 123 |
# =============================
|
| 124 |
+
# ضغط الفيديو قبل المعالجة
|
| 125 |
# =============================
|
| 126 |
|
| 127 |
def preprocess_video(input_path: str, target_width: int = 640, target_fps: int = 8) -> str:
|
|
|
|
| 145 |
frame_step = max(1, int(round(orig_fps / target_fps)))
|
| 146 |
out_fps = orig_fps / frame_step
|
| 147 |
|
|
|
|
| 148 |
target_height = int(target_width * h / w)
|
| 149 |
|
| 150 |
fd, tmp_path = tempfile.mkstemp(suffix=".mp4")
|
|
|
|
| 159 |
if not ret:
|
| 160 |
break
|
| 161 |
|
|
|
|
| 162 |
if frame_idx % frame_step == 0:
|
| 163 |
resized = cv2.resize(frame, (target_width, target_height), interpolation=cv2.INTER_AREA)
|
| 164 |
out.write(resized)
|
|
|
|
| 172 |
|
| 173 |
|
| 174 |
# =============================
|
| 175 |
+
# معالجة فريم واحد
|
| 176 |
# =============================
|
| 177 |
|
| 178 |
def detect_frame(frame_bgr):
|
| 179 |
+
model = get_model() # نتأكد الموديل جاهز وعلى الجهاز الصحيح
|
| 180 |
+
|
| 181 |
frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
|
| 182 |
result = model.predict(
|
| 183 |
frame_rgb,
|
| 184 |
conf=CONF_THRESHOLD,
|
| 185 |
imgsz=IMG_SIZE,
|
| 186 |
verbose=False,
|
| 187 |
+
device=DEVICE, # cuda أو cpu حسب المتاح
|
| 188 |
)[0]
|
| 189 |
|
| 190 |
boxes = result.boxes
|
|
|
|
| 236 |
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
| 237 |
|
| 238 |
if fps <= 0:
|
| 239 |
+
fps = 8.0
|
| 240 |
|
| 241 |
out = cv2.VideoWriter(out_path, fourcc, fps, (width, height))
|
| 242 |
|
|
|
|
| 299 |
|
| 300 |
|
| 301 |
# =============================
|
| 302 |
+
# Gradio + @spaces.GPU
|
| 303 |
# =============================
|
| 304 |
|
| 305 |
+
@spaces.GPU # هذا اللي يرضي Hugging Face ويشغّل GPU on demand
|
| 306 |
def run(file):
|
| 307 |
if file is None:
|
| 308 |
return "لم يتم رفع فيديو", "", None
|
| 309 |
|
| 310 |
video_path = file.name
|
| 311 |
|
|
|
|
| 312 |
light_path = preprocess_video(video_path, target_width=640, target_fps=8)
|
| 313 |
|
| 314 |
raw, processed_path = extract_and_render(light_path)
|
|
|
|
| 321 |
|
| 322 |
|
| 323 |
with gr.Blocks() as demo:
|
| 324 |
+
gr.Markdown("## 🤟 ASL → Arabic (YOLO + Gemini) — نسخة GPU على Hugging Face Spaces")
|
| 325 |
|
| 326 |
inp = gr.File(label="ارفع فيديو الإشارة")
|
| 327 |
raw = gr.Textbox(label="النص الخام", lines=3)
|