STomoya/docling-layout-heron-101-onnx

docling-layout-heron-101 (ONNX)

ONNX export of docling-layout-heron-101, an RT-DETRv2 document layout detector from the Docling project. Unofficial export, not maintained by docling-project.

Files

• model.onnx -- fp32 model, postprocessing included
• config.json, preprocessor_config.json -- copied from the base model

Input / output

• Inputs: images (float32, [N,3,H,W], output of the HF AutoImageProcessor at its normal defaults -- resize + rescale + normalize all included), orig_target_sizes (int64, [N,2], [width, height])
• Outputs: labels (int64, [N,300]), boxes (float32, [N,300,4], xyxy pixels), scores (float32, [N,300])
• Not thresholded -- filter by scores yourself.

Usage

import numpy as np, onnxruntime as ort
from PIL import Image
from transformers import AutoImageProcessor

processor = AutoImageProcessor.from_pretrained("<this-repo-id>")
session = ort.InferenceSession("model.onnx")

image = Image.open("page.png").convert("RGB")
inputs = processor(images=[image], return_tensors="np")
orig_sizes = np.array([[image.width, image.height]], dtype=np.int64)

labels, boxes, scores = session.run(
    None, {"images": inputs["pixel_values"], "orig_target_sizes": orig_sizes}
)

License

Apache 2.0, inherited from the base model.