DocumentFigureClassifier-v2.5

import torch
import torchvision.transforms as transforms

from transformers import EfficientNetForImageClassification
from PIL import Image
import requests


urls = [
    'http://images.cocodataset.org/val2017/000000039769.jpg',
    'http://images.cocodataset.org/test-stuff2017/000000001750.jpg',
    'http://images.cocodataset.org/test-stuff2017/000000000001.jpg'
]

image_processor = transforms.Compose(
    [
        transforms.Resize((224, 224)),
        transforms.ToTensor(),
        transforms.Normalize(
            mean=[0.485, 0.456, 0.406],
            std=[0.47853944, 0.4732864, 0.47434163],
        ),
    ]
)

images = []
for url in urls:
    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
    image = image_processor(image)
    images.append(image)


model_id = 'docling-project/DocumentFigureClassifier-v2.5'

model = EfficientNetForImageClassification.from_pretrained(model_id)

labels = model.config.id2label

device = torch.device("cpu")

torch_images = torch.stack(images).to(device)

with torch.no_grad():
    logits = model(torch_images).logits  # (batch_size, num_classes)
    probs_batch = logits.softmax(dim=1)  # (batch_size, num_classes)
    probs_batch = probs_batch.cpu().numpy().tolist()

for idx, probs_image in enumerate(probs_batch):
    preds = [(labels[i], prob) for i, prob in enumerate(probs_image)]
    preds.sort(key=lambda t: t[1], reverse=True)
    print(f"{idx}: {preds}")