00_load_and_infer.py

from urllib.request import urlopen

import timm
import torch
from PIL import Image

from imagenet_classes import IMAGENET2012_CLASSES

model_name = "eva02_large_patch14_448.mim_m38m_ft_in22k_in1k"
model = timm.create_model(model_name, pretrained=True).eval()

data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)


img = Image.open(
    urlopen(
        "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png"
    )
)

with torch.inference_mode():
    output = model(transforms(img).unsqueeze(0))

top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)
im_classes = list(IMAGENET2012_CLASSES.values())
class_names = [im_classes[i] for i in top5_class_indices[0]]

print("Top 5 predictions:")
for name, prob in zip(class_names, top5_probabilities[0]):
    print(f"  {name}: {prob:.2f}%")