models/efficient_net.py

import torch
from torchvision.models import efficientnet_b7
from torchvision.transforms import InterpolationMode
from torchvision import transforms
from PIL import Image
from torch.nn import functional as F

class EfficientNetB7:
    def __init__(self, weights_path: str):
        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        self.transform = transforms.Compose([
            transforms.Resize((600), interpolation=InterpolationMode.BICUBIC),
            transforms.CenterCrop(600),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                std=[0.229, 0.224, 0.225])
        ])

        self.model = efficientnet_b7()
        self.model.classifier[1] = torch.nn.Linear(2560, 7)

        state_dict = torch.load(weights_path, map_location=torch.device(self.device))
        self.model.load_state_dict(state_dict)
        self.model.eval()
    
    def make_prediction(self, image: Image):
        image = self.transform(image).unsqueeze(0)
        with torch.no_grad():
            pred = F.softmax(self.model(image), dim=1)
        return pred