src/inference.py

import torch
import torchvision.transforms as T

from typing import Optional

from src.dataset import generate_image
from src.models import CrossAttentionClassifier, VGGLikeEncode


class CrossAttentionInference:
    def __init__(
        self,
        model_path: str,
        shape_params: Optional[dict] = None,
        device: torch.device = torch.device("cpu"),
    ):
        if not shape_params:
            self.shape_params = {}
        else:
            self.shape_params = shape_params
        self.device = device

        self.encoder = VGGLikeEncode(
            in_channels=1,
            out_channels=128,
            feature_dim=32,
            apply_pooling=False
        )
        self.model = CrossAttentionClassifier(encoder=self.encoder)

        state_dict = torch.load(model_path, map_location=device)
        self.model.load_state_dict(state_dict)

        self.model.eval()
        self.model.to(device)

        self.transform = T.Compose([
            T.ToTensor(),
            T.Normalize(mean=(0.5,), std=(0.5,))
        ])

    def pil_to_tensor(self, img):
        return self.transform(img).unsqueeze(0).to(self.device)


    def predict_random_pair(self):
        img1, _ = generate_image(**self.shape_params)
        img2, _ = generate_image(**self.shape_params)

        img1_tensor = self.pil_to_tensor(img1)
        img2_tensor = self.pil_to_tensor(img2)

        with torch.no_grad():
            logits, _ = self.model(img1_tensor, img2_tensor)

        preds = (torch.sigmoid(logits) > 0.5).float()
        predicted_label = int(preds.item())

        return predicted_label, (img1, img2)