inference.py

"""
Example inference script for AIBA BERT-BiLSTM model

This script demonstrates how to use the model for prediction.
"""

import torch
from transformers import AutoTokenizer
from nn_model import load_model_and_tokenizer


def predict(text, model_path="YOUR_USERNAME/aiba-bert-bilstm"):
    """
    Make predictions on input text
    
    Args:
        text: Input text to analyze
        model_path: Hugging Face model repo ID
    
    Returns:
        dict with extracted entities, intent, and language
    """
    # Load model
    model, tokenizer, config = load_model_and_tokenizer(model_path)
    
    # Get label mappings from config
    id2tag = {int(k): v for k, v in config['id2tag'].items()}
    id2intent = {int(k): v for k, v in config['id2intent'].items()}
    id2lang = {int(k): v for k, v in config['id2lang'].items()}
    
    # Tokenize
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        max_length=config['max_length'],
        return_offsets_mapping=True
    )
    
    offset_mapping = inputs.pop('offset_mapping')[0]
    
    # Predict
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Process NER predictions
    ner_predictions = torch.argmax(outputs['ner_logits'], dim=2)[0]
    tokens = tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])
    
    # Extract entities
    entities = []
    current_entity = None
    current_tokens = []
    
    for token, pred_id, offset in zip(tokens, ner_predictions, offset_mapping):
        if token in ['[CLS]', '[SEP]', '[PAD]']:
            continue
        
        tag = id2tag[pred_id.item()]
        
        if tag.startswith('B-'):
            if current_entity:
                entity_text = text[current_tokens[0][0]:current_tokens[-1][1]]
                entities.append({'type': current_entity, 'value': entity_text})
            current_entity = tag[2:]
            current_tokens = [offset.tolist()]
        elif tag.startswith('I-') and current_entity:
            current_tokens.append(offset.tolist())
        else:
            if current_entity:
                entity_text = text[current_tokens[0][0]:current_tokens[-1][1]]
                entities.append({'type': current_entity, 'value': entity_text})
            current_entity = None
            current_tokens = []
    
    if current_entity:
        entity_text = text[current_tokens[0][0]:current_tokens[-1][1]]
        entities.append({'type': current_entity, 'value': entity_text})
    
    # Process intent prediction
    intent_logits = outputs['intent_logits'][0]
    intent_probs = torch.nn.functional.softmax(intent_logits, dim=0)
    intent_id = torch.argmax(intent_probs).item()
    intent = id2intent[intent_id]
    intent_confidence = intent_probs[intent_id].item()
    
    # Process language prediction
    lang_logits = outputs['lang_logits'][0]
    lang_probs = torch.nn.functional.softmax(lang_logits, dim=0)
    lang_id = torch.argmax(lang_probs).item()
    language = id2lang[lang_id]
    lang_confidence = lang_probs[lang_id].item()
    
    return {
        'text': text,
        'entities': entities,
        'intent': intent,
        'intent_confidence': intent_confidence,
        'language': language,
        'language_confidence': lang_confidence
    }


# Example usage
if __name__ == "__main__":
    # Test examples
    examples = [
        "Qabul qiluvchi Omad Biznes MCHJ STIR 123456789 summa 500000 UZS",
        "Получатель ООО Прогресс ИНН 987654321 сумма 1000000 руб",
        "Transfer 5000 USD to Starlight Ventures LLC TIN 555666777"
    ]
    
    for text in examples:
        print(f"\nInput: {text}")
        result = predict(text)
        print(f"Intent: {result['intent']} ({result['intent_confidence']:.2%})")
        print(f"Language: {result['language']} ({result['language_confidence']:.2%})")
        print("Entities:")
        for entity in result['entities']:
            print(f"  - {entity['type']}: {entity['value']}")