Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -20,6 +20,9 @@ from transformers import BertTokenizer, BertModel
|
|
| 20 |
import warnings
|
| 21 |
warnings.filterwarnings('ignore')
|
| 22 |
|
|
|
|
|
|
|
|
|
|
| 23 |
try:
|
| 24 |
import pkg_resources
|
| 25 |
print("✅ pkg_resources уже установлен")
|
|
@@ -30,15 +33,6 @@ except ImportError:
|
|
| 30 |
import pkg_resources
|
| 31 |
print("✅ pkg_resources установлен")
|
| 32 |
|
| 33 |
-
# Сохраняем ссылку на класс в глобальной области видимости
|
| 34 |
-
_global_ontology_class = None
|
| 35 |
-
|
| 36 |
-
def _register_ontology_class(cls):
|
| 37 |
-
"""Регистрирует класс OntologyEmotionModel для pickle"""
|
| 38 |
-
import __main__
|
| 39 |
-
__main__.OntologyEmotionModel = cls
|
| 40 |
-
print(f"✅ Класс {cls.__name__} зарегистрирован в __main__")
|
| 41 |
-
|
| 42 |
# Определяем устройство
|
| 43 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
| 44 |
print(f"Используется устройство: {device}")
|
|
@@ -105,7 +99,6 @@ class EmotionBERT(nn.Module):
|
|
| 105 |
def __init__(self, bert_model_name, num_classes, dropout=0.3):
|
| 106 |
super().__init__()
|
| 107 |
self.bert = BertModel.from_pretrained(bert_model_name)
|
| 108 |
-
# Замораживаем все слои кроме последних
|
| 109 |
for p in list(self.bert.parameters())[:-50]:
|
| 110 |
p.requires_grad = False
|
| 111 |
hidden = self.bert.config.hidden_size
|
|
@@ -240,13 +233,11 @@ class OntologyEmotionModel:
|
|
| 240 |
adj = rule_analysis['adjustments']
|
| 241 |
rules = rule_analysis['rules_applied']
|
| 242 |
|
| 243 |
-
# Базовая корректировка
|
| 244 |
conf_mult = 1.0 + adj['arousal'] * 0.2
|
| 245 |
conf_mult = np.clip(conf_mult, 0.5, 1.5)
|
| 246 |
new_confidence = original_confidence * conf_mult
|
| 247 |
new_emotion = original_emotion
|
| 248 |
|
| 249 |
-
# Специальные правила
|
| 250 |
for rule in rules:
|
| 251 |
if rule.startswith("отрицания:"):
|
| 252 |
new_confidence *= 0.8
|
|
@@ -260,7 +251,6 @@ class OntologyEmotionModel:
|
|
| 260 |
def get_ontology_analysis(self, text: str, model_prediction: Dict) -> Dict:
|
| 261 |
rule_analysis = self.apply_linguistic_rules(text)
|
| 262 |
adjusted = self.adjust_prediction_with_rules(model_prediction, rule_analysis)
|
| 263 |
-
|
| 264 |
return {
|
| 265 |
'rule_analysis': rule_analysis,
|
| 266 |
'adjusted_prediction': adjusted
|
|
@@ -321,7 +311,6 @@ class CascadeEmotionClassifier:
|
|
| 321 |
'probabilities': lstm_probs[0].cpu().numpy().tolist()
|
| 322 |
}
|
| 323 |
|
| 324 |
-
# Применяем онтологию к LSTM предсказанию
|
| 325 |
lstm_onto = self.ontology_model.get_ontology_analysis(text_clean, lstm_pred_dict)
|
| 326 |
lstm_adjusted = lstm_onto['adjusted_prediction']
|
| 327 |
|
|
@@ -330,7 +319,6 @@ class CascadeEmotionClassifier:
|
|
| 330 |
final = lstm_adjusted
|
| 331 |
used_model = "LSTM с онтологией"
|
| 332 |
else:
|
| 333 |
-
# BERT prediction
|
| 334 |
self.stats['bert'] += 1
|
| 335 |
enc = self.tokenizer(
|
| 336 |
text_clean,
|
|
@@ -356,14 +344,12 @@ class CascadeEmotionClassifier:
|
|
| 356 |
'probabilities': bert_probs[0].cpu().numpy().tolist()
|
| 357 |
}
|
| 358 |
|
| 359 |
-
# Применяем онтологию к BERT предсказанию
|
| 360 |
bert_onto = self.ontology_model.get_ontology_analysis(text_clean, bert_pred_dict)
|
| 361 |
bert_adjusted = bert_onto['adjusted_prediction']
|
| 362 |
final = bert_adjusted
|
| 363 |
used_model = "BERT с онтологией"
|
| 364 |
-
lstm_onto = bert_onto
|
| 365 |
|
| 366 |
-
# Формируем результат
|
| 367 |
result = {
|
| 368 |
'text': text,
|
| 369 |
'text_clean': text_clean,
|
|
@@ -380,47 +366,41 @@ class CascadeEmotionClassifier:
|
|
| 380 |
'lstm_confidence': float(lstm_conf.item()),
|
| 381 |
'was_corrected': len(lstm_onto['rule_analysis']['rules_applied']) > 0
|
| 382 |
}
|
| 383 |
-
|
| 384 |
return result
|
| 385 |
|
| 386 |
# ============================================================
|
| 387 |
-
#
|
| 388 |
-
# ============================================================
|
| 389 |
-
_register_ontology_class(OntologyEmotionModel)
|
| 390 |
-
|
| 391 |
-
# ============================================================
|
| 392 |
-
# ЗАГРУЗКА МОДЕЛИ (исправленная версия)
|
| 393 |
# ============================================================
|
| 394 |
-
|
| 395 |
-
print("
|
| 396 |
-
|
| 397 |
-
# Предварительно инициализируем pymorphy3, чтобы словари загрузились
|
| 398 |
-
test_morph = pymorphy3.MorphAnalyzer()
|
| 399 |
-
test_word = test_morph.parse('тест')[0]
|
| 400 |
-
print("✅ pymorphy3 инициализирован")
|
| 401 |
|
| 402 |
-
#
|
| 403 |
-
|
| 404 |
-
|
| 405 |
|
| 406 |
-
# Загружаем
|
| 407 |
-
with open(f'{model_dir}/
|
| 408 |
-
|
| 409 |
-
print("✅ Сохранённая онтология успешно загружена!")
|
| 410 |
|
| 411 |
-
#
|
| 412 |
-
|
| 413 |
-
|
| 414 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 415 |
|
| 416 |
-
|
| 417 |
-
|
| 418 |
-
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
|
| 422 |
-
|
| 423 |
-
# ============================================================
|
| 424 |
|
| 425 |
# Создаем и загружаем LSTM
|
| 426 |
print("📂 Загрузка LSTM модели...")
|
|
@@ -474,11 +454,8 @@ except Exception as e:
|
|
| 474 |
# FASTAPI ПРИЛОЖЕНИЕ
|
| 475 |
# ============================================================
|
| 476 |
app = FastAPI(title="Emotion Analysis with BERT and Ontology")
|
| 477 |
-
|
| 478 |
-
# Настраиваем шаблоны
|
| 479 |
templates = Jinja2Templates(directory="templates")
|
| 480 |
|
| 481 |
-
# Глобальная переменная для модели
|
| 482 |
classifier = None
|
| 483 |
model_info = None
|
| 484 |
|
|
@@ -510,16 +487,14 @@ async def predict(text: str = Form(...)):
|
|
| 510 |
try:
|
| 511 |
result = classifier.predict(text)
|
| 512 |
|
| 513 |
-
# Форматируем правила для отображения
|
| 514 |
rules_display = []
|
| 515 |
-
for rule in result['rules_applied'][:10]:
|
| 516 |
if ':' in rule:
|
| 517 |
cat, val = rule.split(':', 1)
|
| 518 |
rules_display.append(f"<span class='rule-tag rule-{cat.strip()}'>{cat}: {val.strip()}</span>")
|
| 519 |
else:
|
| 520 |
rules_display.append(f"<span class='rule-tag'>{rule}</span>")
|
| 521 |
|
| 522 |
-
# Форматируем вероятности
|
| 523 |
probs_display = []
|
| 524 |
for emotion, prob in result['class_probabilities'].items():
|
| 525 |
percentage = prob * 100
|
|
|
|
| 20 |
import warnings
|
| 21 |
warnings.filterwarnings('ignore')
|
| 22 |
|
| 23 |
+
# ============================================================
|
| 24 |
+
# Устанавливаем setuptools для pkg_resources
|
| 25 |
+
# ============================================================
|
| 26 |
try:
|
| 27 |
import pkg_resources
|
| 28 |
print("✅ pkg_resources уже установлен")
|
|
|
|
| 33 |
import pkg_resources
|
| 34 |
print("✅ pkg_resources установлен")
|
| 35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
# Определяем устройство
|
| 37 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
| 38 |
print(f"Используется устройство: {device}")
|
|
|
|
| 99 |
def __init__(self, bert_model_name, num_classes, dropout=0.3):
|
| 100 |
super().__init__()
|
| 101 |
self.bert = BertModel.from_pretrained(bert_model_name)
|
|
|
|
| 102 |
for p in list(self.bert.parameters())[:-50]:
|
| 103 |
p.requires_grad = False
|
| 104 |
hidden = self.bert.config.hidden_size
|
|
|
|
| 233 |
adj = rule_analysis['adjustments']
|
| 234 |
rules = rule_analysis['rules_applied']
|
| 235 |
|
|
|
|
| 236 |
conf_mult = 1.0 + adj['arousal'] * 0.2
|
| 237 |
conf_mult = np.clip(conf_mult, 0.5, 1.5)
|
| 238 |
new_confidence = original_confidence * conf_mult
|
| 239 |
new_emotion = original_emotion
|
| 240 |
|
|
|
|
| 241 |
for rule in rules:
|
| 242 |
if rule.startswith("отрицания:"):
|
| 243 |
new_confidence *= 0.8
|
|
|
|
| 251 |
def get_ontology_analysis(self, text: str, model_prediction: Dict) -> Dict:
|
| 252 |
rule_analysis = self.apply_linguistic_rules(text)
|
| 253 |
adjusted = self.adjust_prediction_with_rules(model_prediction, rule_analysis)
|
|
|
|
| 254 |
return {
|
| 255 |
'rule_analysis': rule_analysis,
|
| 256 |
'adjusted_prediction': adjusted
|
|
|
|
| 311 |
'probabilities': lstm_probs[0].cpu().numpy().tolist()
|
| 312 |
}
|
| 313 |
|
|
|
|
| 314 |
lstm_onto = self.ontology_model.get_ontology_analysis(text_clean, lstm_pred_dict)
|
| 315 |
lstm_adjusted = lstm_onto['adjusted_prediction']
|
| 316 |
|
|
|
|
| 319 |
final = lstm_adjusted
|
| 320 |
used_model = "LSTM с онтологией"
|
| 321 |
else:
|
|
|
|
| 322 |
self.stats['bert'] += 1
|
| 323 |
enc = self.tokenizer(
|
| 324 |
text_clean,
|
|
|
|
| 344 |
'probabilities': bert_probs[0].cpu().numpy().tolist()
|
| 345 |
}
|
| 346 |
|
|
|
|
| 347 |
bert_onto = self.ontology_model.get_ontology_analysis(text_clean, bert_pred_dict)
|
| 348 |
bert_adjusted = bert_onto['adjusted_prediction']
|
| 349 |
final = bert_adjusted
|
| 350 |
used_model = "BERT с онтологией"
|
| 351 |
+
lstm_onto = bert_onto
|
| 352 |
|
|
|
|
| 353 |
result = {
|
| 354 |
'text': text,
|
| 355 |
'text_clean': text_clean,
|
|
|
|
| 366 |
'lstm_confidence': float(lstm_conf.item()),
|
| 367 |
'was_corrected': len(lstm_onto['rule_analysis']['rules_applied']) > 0
|
| 368 |
}
|
|
|
|
| 369 |
return result
|
| 370 |
|
| 371 |
# ============================================================
|
| 372 |
+
# ЗАГРУЗКА МОДЕЛИ
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 373 |
# ============================================================
|
| 374 |
+
def load_model():
|
| 375 |
+
print("Загрузка модели...")
|
| 376 |
+
model_dir = 'model'
|
|
|
|
|
|
|
|
|
|
|
|
|
| 377 |
|
| 378 |
+
# Загружаем информацию о модели
|
| 379 |
+
with open(f'{model_dir}/model_info.json', 'r', encoding='utf-8') as f:
|
| 380 |
+
model_info = json.load(f)
|
| 381 |
|
| 382 |
+
# Загружаем vocab
|
| 383 |
+
with open(f'{model_dir}/vocab.json', 'r', encoding='utf-8') as f:
|
| 384 |
+
vocab = json.load(f)
|
|
|
|
| 385 |
|
| 386 |
+
# Загружаем label encoder
|
| 387 |
+
with open(f'{model_dir}/label_encoder.pkl', 'rb') as f:
|
| 388 |
+
label_encoder = pickle.load(f)
|
| 389 |
+
|
| 390 |
+
# Загружаем онтологию
|
| 391 |
+
print("📂 Загрузка сохранённой онтологии...")
|
| 392 |
+
try:
|
| 393 |
+
# Убеждаемся, что класс доступен
|
| 394 |
+
import __main__
|
| 395 |
+
__main__.OntologyEmotionModel = OntologyEmotionModel
|
| 396 |
|
| 397 |
+
# Загружаем онтологию
|
| 398 |
+
with open(f'{model_dir}/ontology_model.pkl', 'rb') as f:
|
| 399 |
+
ontology_model = pickle.load(f)
|
| 400 |
+
print("✅ Сохранённая онтология успешно загружена!")
|
| 401 |
+
except Exception as e:
|
| 402 |
+
print(f"❌ Ошибка загрузки онтологии: {e}")
|
| 403 |
+
raise RuntimeError("Не удалось загрузить онтологию") from e
|
|
|
|
| 404 |
|
| 405 |
# Создаем и загружаем LSTM
|
| 406 |
print("📂 Загрузка LSTM модели...")
|
|
|
|
| 454 |
# FASTAPI ПРИЛОЖЕНИЕ
|
| 455 |
# ============================================================
|
| 456 |
app = FastAPI(title="Emotion Analysis with BERT and Ontology")
|
|
|
|
|
|
|
| 457 |
templates = Jinja2Templates(directory="templates")
|
| 458 |
|
|
|
|
| 459 |
classifier = None
|
| 460 |
model_info = None
|
| 461 |
|
|
|
|
| 487 |
try:
|
| 488 |
result = classifier.predict(text)
|
| 489 |
|
|
|
|
| 490 |
rules_display = []
|
| 491 |
+
for rule in result['rules_applied'][:10]:
|
| 492 |
if ':' in rule:
|
| 493 |
cat, val = rule.split(':', 1)
|
| 494 |
rules_display.append(f"<span class='rule-tag rule-{cat.strip()}'>{cat}: {val.strip()}</span>")
|
| 495 |
else:
|
| 496 |
rules_display.append(f"<span class='rule-tag'>{rule}</span>")
|
| 497 |
|
|
|
|
| 498 |
probs_display = []
|
| 499 |
for emotion, prob in result['class_probabilities'].items():
|
| 500 |
percentage = prob * 100
|