thesanogoeffect/PP-DocLayoutV3-ONNX

PP-DocLayout V3 — ONNX (Dynamic Batch)

ONNX export of PaddlePaddle/PP-DocLayoutV3_safetensors with dynamic batch axis support.

Details

• Architecture: RT-DETR with HGNet-v2 backbone
• Input: pixel_values — [batch, 3, 800, 800] (float32)
• Outputs: logits [batch, 300, 25], pred_boxes [batch, 300, 4]
• Opset: 17
• Export method: torch.onnx.export (TorchScript path) from HuggingFace Transformers

Usage

import onnxruntime as ort
import numpy as np

sess = ort.InferenceSession("pp_doclayout_v3_dynbatch.onnx")
images = np.random.randn(4, 3, 800, 800).astype(np.float32)  # batch of 4
logits, pred_boxes = sess.run(None, {"pixel_values": images})[:2]

Labels

ID	Label	ID	Label
0	abstract	13	header
1	algorithm	14	image
2	aside_text	15	formula
3	chart	16	number
4	content	17	paragraph_title
5	formula	18	reference
6	doc_title	19	reference_content
7	figure_title	20	seal
8	footer	21	table
9	footer	22	text
10	footnote	23	text
11	formula_number	24	vision_footnote
12	header