Base model: torchvision.models.swin_b (ImageNet pretrained)
Optimizer: AdamW (lr=9.5e-4, weight_decay=7.2e-3, backbone lr multiplier=0.15)
Scheduler: Cosine annealing (eta_min=5.3e-7)
Augmentation: RandAugment + MixUp (α=0.2) + CutMix (α=1.0)
Regularization: EMA (decay=0.9998), label smoothing=0.1, dropout=0.183
Epochs: 35 (early stopping, patience=10)
Hardware: Single CUDA GPU

Hyperparameters were found via Optuna (TPE sampler, 75-epoch budget per trial).

Usage

python

import torch
import torch.nn as nn
import torchvision.models as tv_models
from torchvision import transforms
from huggingface_hub import hf_hub_download
from PIL import Image

# Build model
backbone = tv_models.swin_b(weights=None)
backbone.head = nn.Sequential(
    nn.Flatten(1),
    nn.Dropout(p=0.183),
    nn.Linear(1024, 350),
    nn.ReLU(inplace=True),
    nn.Linear(350, 101),
)
backbone.load_state_dict(
    torch.load(hf_hub_download("karlghosn/swin-b-food101", "best_model.pt"), map_location="cpu")
)
backbone.eval()

# Preprocess
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

image = Image.open("your_food_image.jpg").convert("RGB")
with torch.inference_mode():
    logits = backbone(transform(image).unsqueeze(0))
    probs = torch.softmax(logits, dim=1)[0]