Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -3,73 +3,69 @@ from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
|
| 3 |
import torch
|
| 4 |
import torch.nn.functional as F
|
| 5 |
import re
|
| 6 |
-
import
|
| 7 |
|
|
|
|
|
|
|
|
|
|
| 8 |
|
| 9 |
-
|
|
|
|
|
|
|
| 10 |
|
| 11 |
-
#
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
except Exception as e:
|
| 17 |
-
print(f"Error loading local model: {e}")
|
| 18 |
-
print("Fallback to base model (Not recommended for final output)...")
|
| 19 |
-
# كود احتياطي لو الملفات مش موجودة
|
| 20 |
-
tokenizer = AutoTokenizer.from_pretrained("UBC-NLP/MARBERTv2")
|
| 21 |
-
model = AutoModelForSequenceClassification.from_pretrained("UBC-NLP/MARBERTv2", num_labels=5)
|
| 22 |
|
| 23 |
-
#
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
|
|
|
| 31 |
|
| 32 |
-
# ---
|
| 33 |
def clean_text(text):
|
| 34 |
if not text: return ""
|
| 35 |
-
text = re.sub(r'[\u064B-\u0652]', '', text)
|
| 36 |
-
text = re.sub(r'[أإآ]', 'ا', text)
|
| 37 |
-
text = re.sub(r'ى', 'ي', text)
|
| 38 |
-
text = re.sub(r'ة', 'ه', text)
|
| 39 |
-
text = re.sub(r'(.)\1+', r'\1', text)
|
|
|
|
| 40 |
return text
|
| 41 |
|
| 42 |
-
# ---
|
| 43 |
def classify_text(text):
|
| 44 |
if not text: return {}
|
| 45 |
-
|
| 46 |
-
#
|
| 47 |
cleaned = clean_text(text)
|
| 48 |
inputs = tokenizer(cleaned, return_tensors="pt", padding=True, truncation=True, max_length=128)
|
| 49 |
|
| 50 |
-
#
|
| 51 |
with torch.no_grad():
|
| 52 |
logits = model(**inputs).logits
|
| 53 |
|
| 54 |
-
#
|
| 55 |
probs = F.softmax(logits, dim=-1)[0].numpy()
|
| 56 |
|
| 57 |
-
#
|
| 58 |
results = {}
|
| 59 |
for i, score in enumerate(probs):
|
| 60 |
-
|
|
|
|
| 61 |
results[label] = float(score)
|
| 62 |
|
| 63 |
return results
|
| 64 |
|
| 65 |
-
#
|
| 66 |
iface = gr.Interface(
|
| 67 |
fn=classify_text,
|
| 68 |
-
inputs=gr.Textbox(label="اكتب النص
|
| 69 |
outputs=gr.Label(label="النتيجة"),
|
| 70 |
-
title="Arabic Toxicity Detection
|
| 71 |
-
description="
|
| 72 |
-
examples=[["انت راجل محترم"], ["يا ابن الكلب"], ["دي حاجة تقرف"]]
|
| 73 |
)
|
| 74 |
|
| 75 |
iface.launch()
|
|
|
|
| 3 |
import torch
|
| 4 |
import torch.nn.functional as F
|
| 5 |
import re
|
| 6 |
+
import json
|
| 7 |
|
| 8 |
+
# --- 1. تحميل الموديل (إجباري من نفس المكان) ---
|
| 9 |
+
# مش هنحط try/except عشان نتأكد انه بيقرأ ملفاتك انت
|
| 10 |
+
model_path = "."
|
| 11 |
|
| 12 |
+
print("Loading model from current directory...")
|
| 13 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
| 14 |
+
model = AutoModelForSequenceClassification.from_pretrained(model_path)
|
| 15 |
|
| 16 |
+
# --- 2. قراءة ترتيب الكلاسات من ملف config.json ---
|
| 17 |
+
# دي أضمن طريقة عشان الترتيب يطلع زي ما اتدرب بالظبط
|
| 18 |
+
with open('config.json', 'r') as f:
|
| 19 |
+
config = json.load(f)
|
| 20 |
+
id2label = config.get('id2label')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
+
# لو الترتيب مش موجود في الملف، هنستخدم الترتيب الافتراضي (تأكد انه مناسب ليك)
|
| 23 |
+
if not id2label:
|
| 24 |
+
id2label = {
|
| 25 |
+
"0": "مسيء / كراهية (Hate)",
|
| 26 |
+
"1": "هجومي (Offensive)",
|
| 27 |
+
"2": "عادي / محايد (Neutral)",
|
| 28 |
+
"3": "إهانة (Insult)",
|
| 29 |
+
"4": "تهديد (Threat)"
|
| 30 |
+
}
|
| 31 |
|
| 32 |
+
# --- 3. دالة التنضيف ---
|
| 33 |
def clean_text(text):
|
| 34 |
if not text: return ""
|
| 35 |
+
text = re.sub(r'[\u064B-\u0652]', '', text)
|
| 36 |
+
text = re.sub(r'[أإآ]', 'ا', text)
|
| 37 |
+
text = re.sub(r'ى', 'ي', text)
|
| 38 |
+
text = re.sub(r'ة', 'ه', text)
|
| 39 |
+
text = re.sub(r'(.)\1+', r'\1', text)
|
| 40 |
+
text = re.sub(r'[^\u0621-\u064A\u0660-\u0669\s]', '', text)
|
| 41 |
return text
|
| 42 |
|
| 43 |
+
# --- 4. التنبؤ ---
|
| 44 |
def classify_text(text):
|
| 45 |
if not text: return {}
|
|
|
|
|
|
|
| 46 |
cleaned = clean_text(text)
|
| 47 |
inputs = tokenizer(cleaned, return_tensors="pt", padding=True, truncation=True, max_length=128)
|
| 48 |
|
|
|
|
| 49 |
with torch.no_grad():
|
| 50 |
logits = model(**inputs).logits
|
| 51 |
|
|
|
|
| 52 |
probs = F.softmax(logits, dim=-1)[0].numpy()
|
| 53 |
|
|
|
|
| 54 |
results = {}
|
| 55 |
for i, score in enumerate(probs):
|
| 56 |
+
# بنجيب الاسم الصح بناء على رقم الكلاس
|
| 57 |
+
label = id2label.get(str(i), f"Class {i}")
|
| 58 |
results[label] = float(score)
|
| 59 |
|
| 60 |
return results
|
| 61 |
|
| 62 |
+
# --- 5. الواجهة ---
|
| 63 |
iface = gr.Interface(
|
| 64 |
fn=classify_text,
|
| 65 |
+
inputs=gr.Textbox(label="اكتب النص"),
|
| 66 |
outputs=gr.Label(label="النتيجة"),
|
| 67 |
+
title="Arabic Toxicity Detection",
|
| 68 |
+
description="تجربة النظام (يجب أن تكون الملفات pytorch_model.bin و config.json موجودة)."
|
|
|
|
| 69 |
)
|
| 70 |
|
| 71 |
iface.launch()
|