aoiandroid/PP-DocLayoutV3-safetensors-CoreML

PP-DocLayoutV3 safetensors Core ML (mlprogram)

Private distribution of the Core ML export for PaddlePaddle/PP-DocLayoutV3_safetensors, built for TranslateBlue on-device layout detection.

Source weights: PaddlePaddle/PP-DocLayoutV3 / PP-DocLayoutV3_safetensors.

Files

File	Description
PP-DocLayoutV3-safetensors.mlpackage/	Core ML mlprogram (iOS 18+, float32)
preprocessor_config.json	HF image processor contract (800×800)
conversion-status.json	Export metadata
parity-report.json	PyTorch vs Core ML tensor diff (smoke image)

Input / output (in-package)

Tensor	Shape	Dtype
pixel_values (in)	[1, 3, 800, 800]	float32
logits (out)	[1, 300, 25]	float32
pred_boxes (out)	[1, 300, 4]	float32 (normalized cxcywh)
order_logits (out)	[1, 300, 300]	float32

Preprocessing

Match PPDocLayoutV3ImageProcessor:

• Resize to 800×800 (bicubic, antialias=False in Transformers export)
• Rescale: pixel / 255
• Normalize: mean [0,0,0], std [1,1,1]

See preprocessor_config.json.

Post-processing (not in Core ML)

Apply the same logic as PPDocLayoutV3ImageProcessor.post_process_object_detection in Python or Swift:

• Sigmoid + top-k on logits
• Reading order from order_logits
• Optional polygon masks require full PyTorch path (masks omitted in this export)

Conversion notes

• Export script: TranslateBlue Scripts/export_pp_doclayout_v3_safetensors_coreml/
• numpy<2.4 required for coremltools convert (coremltools#2633)
• compute_precision=float32 (fp16 MIL cast diverges)
• Fixed 800×800 input (trace); other sizes are unsupported

Parity (reference)

On bus.jpg (TranslateBlue smoke):

Tensor	max abs diff vs export wrapper
logits	7.2e-5
pred_boxes	9.7e-6
order_logits	0.04

License

Apache 2.0 (same family as PaddlePaddle/PP-DocLayoutV3).