models.py

# models.py
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
from PIL import Image
from torchvision import transforms
from torchvision.models import efficientnet_b0

# -------------------
# Текстовые модели
# -------------------
class TextModels:
    def __init__(self):
        # модель 1
        self.tokenizer1 = AutoTokenizer.from_pretrained("VSAsteroid/ai-text-detector-hc3")
        self.model1 = AutoModelForSequenceClassification.from_pretrained("VSAsteroid/ai-text-detector-hc3")
        self.model1.eval()

        # модель 2 (для ensemble)
        self.tokenizer2 = AutoTokenizer.from_pretrained("silentone0725/text-detector-model-v2")
        self.model2 = AutoModelForSequenceClassification.from_pretrained("silentone0725/text-detector-model-v2")
        self.model2.eval()

    def predict(self, text):
        # модель 1
        inputs1 = self.tokenizer1(text, return_tensors="pt", truncation=True, max_length=512)
        with torch.no_grad():
            outputs1 = self.model1(**inputs1)
            prob1 = torch.softmax(outputs1.logits, dim=1)[0][1].item()  # AI probability

        # модель 2
        inputs2 = self.tokenizer2(text, return_tensors="pt", truncation=True, max_length=512)
        with torch.no_grad():
            outputs2 = self.model2(**inputs2)
            prob2 = torch.softmax(outputs2.logits, dim=1)[0][1].item()  # AI probability

        # ensemble
        return (prob1 + prob2) / 2


# -------------------
# Изображения
# -------------------
class ImageModel:
    def __init__(self):
        self.model = efficientnet_b0(pretrained=True)
        self.model.eval()
        # трансформация для EfficientNet
        self.transform = transforms.Compose([
            transforms.Resize((224,224)),
            transforms.ToTensor(),
            transforms.Normalize([0.485,0.456,0.406],[0.229,0.224,0.225])
        ])

    def predict(self, image: Image.Image):
        img = self.transform(image).unsqueeze(0)  # batch 1
        with torch.no_grad():
            outputs = self.model(img)
            # для MVP используем sigmoid на один нейрон
            prob = torch.sigmoid(outputs[:,0])[0].item()
        return prob