Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,51 +1,39 @@
|
|
| 1 |
import os
|
| 2 |
-
|
| 3 |
-
# ⚠️ إجبار TensorFlow على عدم حجز الـ GPU بالكامل أثناء الـ Startup لمنع تصادم ZeroGPU
|
| 4 |
-
os.environ["TF_FORCE_GPU_ALLOW_GROWTH"] = "true"
|
| 5 |
-
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
|
| 6 |
-
|
| 7 |
import cv2
|
| 8 |
import numpy as np
|
| 9 |
-
import tensorflow as tf
|
| 10 |
import gradio as gr
|
| 11 |
-
import spaces
|
| 12 |
-
from tensorflow.keras import datasets, layers, models
|
| 13 |
|
| 14 |
# ============================================
|
| 15 |
-
# 1.
|
| 16 |
# ============================================
|
| 17 |
model_path = 'mnist_cnn_model.keras'
|
| 18 |
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
optimizer='adam',
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
print("✅ تم حفظ النموذج.")
|
| 45 |
-
|
| 46 |
-
# تحميل بيانات الاختبار للأمثلة
|
| 47 |
-
(_, _), (test_images, test_labels) = datasets.mnist.load_data()
|
| 48 |
-
test_images = test_images.reshape((10000, 28, 28, 1)).astype('float32') / 255.0
|
| 49 |
|
| 50 |
# ============================================
|
| 51 |
# 2. دالة معالجة الصورة
|
|
@@ -55,9 +43,7 @@ def process_image_for_mnist(image_input):
|
|
| 55 |
return None, None
|
| 56 |
|
| 57 |
if isinstance(image_input, dict):
|
| 58 |
-
image = image_input.get('composite', None)
|
| 59 |
-
if image is None:
|
| 60 |
-
image = image_input.get('background', None)
|
| 61 |
if image is None and len(image_input) > 0:
|
| 62 |
image = list(image_input.values())[0]
|
| 63 |
else:
|
|
@@ -85,11 +71,15 @@ def process_image_for_mnist(image_input):
|
|
| 85 |
|
| 86 |
return normalized, reshaped
|
| 87 |
|
|
|
|
|
|
|
|
|
|
| 88 |
# ============================================
|
| 89 |
-
# 3. دالة التوقع الم
|
| 90 |
# ============================================
|
| 91 |
@spaces.GPU
|
| 92 |
def predict_sketch(image):
|
|
|
|
| 93 |
try:
|
| 94 |
if image is None:
|
| 95 |
return {}, None, "⚠️ يرجى الرسم في المربع أولاً!"
|
|
@@ -99,8 +89,23 @@ def predict_sketch(image):
|
|
| 99 |
if normalized is None:
|
| 100 |
return {}, None, "⚠️ تعذر معالجة الصورة المدخلة."
|
| 101 |
|
| 102 |
-
#
|
| 103 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
predicted_class = int(np.argmax(preds))
|
| 105 |
confidence = float(np.max(preds))
|
| 106 |
|
|
@@ -112,7 +117,7 @@ def predict_sketch(image):
|
|
| 112 |
top3_idx = np.argsort(preds)[-3:][::-1]
|
| 113 |
top3_str = "\n".join([f" {i+1}. الرقم {idx}: {preds[idx]:.2%}" for i, idx in enumerate(top3_idx)])
|
| 114 |
|
| 115 |
-
debug_info = f"""📊 نتائج التوقع (ZeroGPU):
|
| 116 |
- الرقم المتوقع: {predicted_class}
|
| 117 |
- نسبة الثقة: {confidence:.2%}
|
| 118 |
|
|
@@ -125,17 +130,7 @@ def predict_sketch(image):
|
|
| 125 |
return {}, None, f"❌ خطأ أثناء التوقّع: {str(e)}"
|
| 126 |
|
| 127 |
# ============================================
|
| 128 |
-
# 4.
|
| 129 |
-
# ============================================
|
| 130 |
-
def random_example():
|
| 131 |
-
idx = np.random.randint(0, len(test_images))
|
| 132 |
-
img = test_images[idx].reshape(28, 28)
|
| 133 |
-
img_large = cv2.resize(img, (280, 280), interpolation=cv2.INTER_NEAREST)
|
| 134 |
-
img_rgb = np.stack([img_large] * 3, axis=2)
|
| 135 |
-
return img_rgb, f"🎲 تم اختيار رقم عشوائي (الرقم الحقيقي: {test_labels[idx]})"
|
| 136 |
-
|
| 137 |
-
# ============================================
|
| 138 |
-
# 5. بناء واجهة Gradio
|
| 139 |
# ============================================
|
| 140 |
with gr.Blocks(title="MNIST Recognizer on ZeroGPU") as demo:
|
| 141 |
gr.Markdown("# 🧠 التعرف على الأرقام المكتوبة بخط اليد (ZeroGPU)")
|
|
@@ -143,9 +138,7 @@ with gr.Blocks(title="MNIST Recognizer on ZeroGPU") as demo:
|
|
| 143 |
with gr.Row():
|
| 144 |
with gr.Column(scale=1):
|
| 145 |
sketch = gr.Sketchpad(label="✏️ منطقة الرسم")
|
| 146 |
-
|
| 147 |
-
submit_btn = gr.Button("🔮 توقع", variant="primary")
|
| 148 |
-
random_btn = gr.Button("🎲 مثال عشوائي", variant="secondary")
|
| 149 |
info = gr.Textbox(label="📌 معلومات التصحيح", interactive=False, lines=8)
|
| 150 |
|
| 151 |
with gr.Column(scale=1):
|
|
@@ -157,11 +150,5 @@ with gr.Blocks(title="MNIST Recognizer on ZeroGPU") as demo:
|
|
| 157 |
inputs=sketch,
|
| 158 |
outputs=[output, processed_img, info]
|
| 159 |
)
|
| 160 |
-
|
| 161 |
-
random_btn.click(
|
| 162 |
-
fn=random_example,
|
| 163 |
-
inputs=[],
|
| 164 |
-
outputs=[sketch, info]
|
| 165 |
-
)
|
| 166 |
|
| 167 |
demo.launch()
|
|
|
|
| 1 |
import os
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import cv2
|
| 3 |
import numpy as np
|
|
|
|
| 4 |
import gradio as gr
|
| 5 |
+
import spaces
|
|
|
|
| 6 |
|
| 7 |
# ============================================
|
| 8 |
+
# 1. بناء/تدريب النموذج لحفظه محلياً
|
| 9 |
# ============================================
|
| 10 |
model_path = 'mnist_cnn_model.keras'
|
| 11 |
|
| 12 |
+
def ensure_model_exists():
|
| 13 |
+
if not os.path.exists(model_path):
|
| 14 |
+
print("⚠️ جاري تدريب النموذج وحفظه لأول مرة...")
|
| 15 |
+
import tensorflow as tf
|
| 16 |
+
from tensorflow.keras import datasets, layers, models
|
| 17 |
+
|
| 18 |
+
model = models.Sequential([
|
| 19 |
+
layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)),
|
| 20 |
+
layers.MaxPooling2D((2, 2)),
|
| 21 |
+
layers.Conv2D(64, (3, 3), activation='relu'),
|
| 22 |
+
layers.MaxPooling2D((2, 2)),
|
| 23 |
+
layers.Flatten(),
|
| 24 |
+
layers.Dense(128, activation='relu'),
|
| 25 |
+
layers.Dropout(0.5),
|
| 26 |
+
layers.Dense(10, activation='softmax')
|
| 27 |
+
])
|
| 28 |
+
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
|
| 29 |
+
|
| 30 |
+
(x_train, y_train), _ = datasets.mnist.load_data()
|
| 31 |
+
x_train = x_train.reshape((60000, 28, 28, 1)).astype('float32') / 255.0
|
| 32 |
+
model.fit(x_train, y_train, epochs=3, validation_split=0.1, verbose=1)
|
| 33 |
+
model.save(model_path)
|
| 34 |
+
print("✅ تم حفظ النموذج بنجاح.")
|
| 35 |
+
|
| 36 |
+
ensure_model_exists()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
# ============================================
|
| 39 |
# 2. دالة معالجة الصورة
|
|
|
|
| 43 |
return None, None
|
| 44 |
|
| 45 |
if isinstance(image_input, dict):
|
| 46 |
+
image = image_input.get('composite', None) or image_input.get('background', None)
|
|
|
|
|
|
|
| 47 |
if image is None and len(image_input) > 0:
|
| 48 |
image = list(image_input.values())[0]
|
| 49 |
else:
|
|
|
|
| 71 |
|
| 72 |
return normalized, reshaped
|
| 73 |
|
| 74 |
+
# Variable لحفظ النموذج داخل الـ GPU Worker
|
| 75 |
+
loaded_model = None
|
| 76 |
+
|
| 77 |
# ============================================
|
| 78 |
+
# 3. دالة التوقع ببيئة ZeroGPU المضمونة
|
| 79 |
# ============================================
|
| 80 |
@spaces.GPU
|
| 81 |
def predict_sketch(image):
|
| 82 |
+
global loaded_model
|
| 83 |
try:
|
| 84 |
if image is None:
|
| 85 |
return {}, None, "⚠️ يرجى الرسم في المربع أولاً!"
|
|
|
|
| 89 |
if normalized is None:
|
| 90 |
return {}, None, "⚠️ تعذر معالجة الصورة المدخلة."
|
| 91 |
|
| 92 |
+
# تحميل TensorFlow وتكوين الـ GPU ديناميكياً داخل بيئة ZeroGPU فقط
|
| 93 |
+
import tensorflow as tf
|
| 94 |
+
|
| 95 |
+
# تهيئة نمو الذاكرة لتفادي انهيار CUDA memory allocation
|
| 96 |
+
gpus = tf.config.list_physical_devices('GPU')
|
| 97 |
+
if gpus:
|
| 98 |
+
try:
|
| 99 |
+
for gpu in gpus:
|
| 100 |
+
tf.config.experimental.set_memory_growth(gpu, True)
|
| 101 |
+
except RuntimeError:
|
| 102 |
+
pass
|
| 103 |
+
|
| 104 |
+
if loaded_model is None:
|
| 105 |
+
loaded_model = tf.keras.models.load_model(model_path)
|
| 106 |
+
|
| 107 |
+
# التوقع
|
| 108 |
+
preds = loaded_model(reshaped, training=False).numpy()[0]
|
| 109 |
predicted_class = int(np.argmax(preds))
|
| 110 |
confidence = float(np.max(preds))
|
| 111 |
|
|
|
|
| 117 |
top3_idx = np.argsort(preds)[-3:][::-1]
|
| 118 |
top3_str = "\n".join([f" {i+1}. الرقم {idx}: {preds[idx]:.2%}" for i, idx in enumerate(top3_idx)])
|
| 119 |
|
| 120 |
+
debug_info = f"""📊 نتائج التوقع (ZeroGPU Active):
|
| 121 |
- الرقم المتوقع: {predicted_class}
|
| 122 |
- نسبة الثقة: {confidence:.2%}
|
| 123 |
|
|
|
|
| 130 |
return {}, None, f"❌ خطأ أثناء التوقّع: {str(e)}"
|
| 131 |
|
| 132 |
# ============================================
|
| 133 |
+
# 4. بناء واجهة Gradio
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
# ============================================
|
| 135 |
with gr.Blocks(title="MNIST Recognizer on ZeroGPU") as demo:
|
| 136 |
gr.Markdown("# 🧠 التعرف على الأرقام المكتوبة بخط اليد (ZeroGPU)")
|
|
|
|
| 138 |
with gr.Row():
|
| 139 |
with gr.Column(scale=1):
|
| 140 |
sketch = gr.Sketchpad(label="✏️ منطقة الرسم")
|
| 141 |
+
submit_btn = gr.Button("🔮 توقع", variant="primary")
|
|
|
|
|
|
|
| 142 |
info = gr.Textbox(label="📌 معلومات التصحيح", interactive=False, lines=8)
|
| 143 |
|
| 144 |
with gr.Column(scale=1):
|
|
|
|
| 150 |
inputs=sketch,
|
| 151 |
outputs=[output, processed_img, info]
|
| 152 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 153 |
|
| 154 |
demo.launch()
|