api/predict.py

import torch
import numpy as np
import re
from transformers import (
    DistilBertTokenizer, DistilBertForSequenceClassification,
    RobertaTokenizer, RobertaForSequenceClassification,
    BertTokenizer, BertForSequenceClassification
)
from huggingface_hub import hf_hub_download
import gc

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

models = []
tokenizers = []
maxLengths = []
modelWeights = [0.333, 0.333, 0.333]
optimalThreshold = 0.40
uncertaintyThreshold = 0.67

CLASS_NAMES = ['genuine', 'fake']

models_loaded = False

def validateText(text):
    if not isinstance(text, str):
        return False
    text = text.strip()
    return len(text) > 0 and len(text.split()) > 0

def cleanReview(text):
    if not text or not isinstance(text, str):
        return ""
    text = re.sub(r'http\S+|www\.\S+', '', text)
    text = re.sub(r'<[^>]+>', '', text)
    text = re.sub(r'([!?.])\1+', r'\1', text)
    text = ' '.join(text.split())
    return text.strip()

def loadResources():
    global models, tokenizers, maxLengths, models_loaded
    
    if models_loaded:
        return
    
    print("loading ensemble models...", flush=True)
    
    modelConfigs = [
        {
            'filename': 'ensemble_model_1.pth',
            'type': 'distilbert',
            'name': 'distilbert-base-uncased',
            'maxLen': 128
        },
        {
            'filename': 'ensemble_model_2.pth',
            'type': 'roberta',
            'name': 'roberta-base',
            'maxLen': 192
        },
        {
            'filename': 'ensemble_model_3.pth',
            'type': 'bert',
            'name': 'bert-base-uncased',
            'maxLen': 256
        }
    ]
    
    for i, config in enumerate(modelConfigs, 1):
        try:
            print(f"loading model {i}: {config['type']}", flush=True)
            
            modelPath = hf_hub_download(
                repo_id="codingcoolfun9ed/sentinelcheck-models",
                filename=config['filename']
            )
            
            if config['type'] == 'distilbert':
                tokenizer = DistilBertTokenizer.from_pretrained(config['name'])
                model = DistilBertForSequenceClassification.from_pretrained(
                    config['name'], 
                    num_labels=2
                )
            elif config['type'] == 'roberta':
                tokenizer = RobertaTokenizer.from_pretrained(config['name'])
                model = RobertaForSequenceClassification.from_pretrained(
                    config['name'], 
                    num_labels=2
                )
            elif config['type'] == 'bert':
                tokenizer = BertTokenizer.from_pretrained(config['name'])
                model = BertForSequenceClassification.from_pretrained(
                    config['name'], 
                    num_labels=2
                )
            else:
                raise ValueError(f"unknown model type: {config['type']}")
            
            checkpoint = torch.load(modelPath, map_location=device, weights_only=False)
            
            if 'state_dict' not in checkpoint:
                raise ValueError(f"model {i} missing state_dict")
            
            model.load_state_dict(checkpoint['state_dict'], strict=False)
            model = model.to(device)
            model.eval()
            
            for param in model.parameters():
                param.requires_grad = False
            
            models.append(model)
            tokenizers.append(tokenizer)
            maxLengths.append(config['maxLen'])
            
            del checkpoint
            gc.collect()
            
            print(f"model {i} loaded successfully", flush=True)
            
        except Exception as e:
            print(f"error loading model {i}: {str(e)}", flush=True)
            raise
    
    models_loaded = True
    print("all ensemble models loaded", flush=True)

def ensemblePredict(text):
    if not models_loaded:
        loadResources()
    
    if not isinstance(text, str):
        text = str(text)
    
    text = cleanReview(text)
    
    if not validateText(text):
        return {
            'fakeProb': 0.5,
            'genuineProb': 0.5,
            'isFake': None,
            'agreement': 0.0,
            'error': 'invalid_text'
        }
    
    weightedProbs = torch.zeros(1, 2).to(device)
    allPreds = []
    
    try:
        with torch.no_grad():
            for tokenizer, model, maxLen, weight in zip(tokenizers, models, maxLengths, modelWeights):
                inputs = tokenizer(
                    text,
                    return_tensors='pt',
                    truncation=True,
                    max_length=maxLen,
                    padding='max_length'
                )
                inputIds = inputs['input_ids'].to(device)
                attentionMask = inputs['attention_mask'].to(device)
                
                outputs = model(input_ids=inputIds, attention_mask=attentionMask)
                probs = torch.softmax(outputs.logits, dim=1)
                weightedProbs += probs * weight
                
                _, pred = torch.max(probs, 1)
                allPreds.append(pred.item())
                
                del inputs, inputIds, attentionMask, outputs, probs, pred
        
        probs = weightedProbs[0].cpu().numpy()
        genuineProb = float(probs[0])
        fakeProb = float(probs[1])
        
        isFake = fakeProb > optimalThreshold
        
        finalPred = 1 if isFake else 0
        agreementCount = sum(1 for p in allPreds if p == finalPred)
        agreement = agreementCount / len(allPreds)
        
        del weightedProbs, allPreds
        gc.collect()
        
        return {
            'genuineProb': genuineProb,
            'fakeProb': fakeProb,
            'isFake': isFake,
            'agreement': agreement
        }
        
    except Exception as e:
        print(f"prediction error: {str(e)}", flush=True)
        return {
            'fakeProb': 0.5,
            'genuineProb': 0.5,
            'isFake': None,
            'agreement': 0.0,
            'error': str(e)
        }

def getLengthCategory(text):
    if not text:
        return 'empty'
    words = text.split()
    wordCount = len(words)
    if wordCount <= 20:
        return 'short'
    elif wordCount <= 50:
        return 'short-medium'
    elif wordCount <= 100:
        return 'medium'
    elif wordCount <= 200:
        return 'long'
    else:
        return 'very-long'

def predict_review(text):
    if not text or not isinstance(text, str):
        return {
            "prediction": "error",
            "confidence": 0.0,
            "is_fake": None,
            "model_agreement": 0.0,
            "fake_probability": 0.0,
            "genuine_probability": 0.0,
            "length_category": "empty",
            "token_count": 0,
            "error": "invalid input: text must be non-empty string"
        }
    
    cleaned = cleanReview(text)
    
    if not cleaned or len(cleaned.strip()) == 0:
        return {
            "prediction": "error",
            "confidence": 0.0,
            "is_fake": None,
            "model_agreement": 0.0,
            "fake_probability": 0.0,
            "genuine_probability": 0.0,
            "length_category": "empty",
            "token_count": 0,
            "error": "empty text after preprocessing"
        }
    
    result = ensemblePredict(text)
    
    if 'error' in result:
        return {
            "prediction": "error",
            "confidence": 0.0,
            "is_fake": None,
            "model_agreement": result['agreement'] * 100,
            "fake_probability": result['fakeProb'],
            "genuine_probability": result['genuineProb'],
            "length_category": getLengthCategory(cleaned),
            "token_count": len(cleaned.split()),
            "error": result['error']
        }
    
    fakeProb = result['fakeProb']
    genuineProb = result['genuineProb']
    isFake = result['isFake']
    agreement = result['agreement']
    
    confidence = max(fakeProb, genuineProb)
    
    prediction = "fake" if isFake else "genuine"
    isFakeOutput = isFake
    
    if agreement < uncertaintyThreshold:
        prediction = "uncertain"
        isFakeOutput = None
    
    lengthCat = getLengthCategory(cleaned)
    tokenCount = len(cleaned.split())
    
    return {
        "prediction": prediction,
        "confidence": confidence,
        "is_fake": isFakeOutput,
        "model_agreement": agreement * 100,
        "fake_probability": fakeProb,
        "genuine_probability": genuineProb,
        "length_category": lengthCat,
        "token_count": tokenCount
    }