bundle.json

{
  "architecture_id": "ocr.ppocr-dbnet",
  "bundle_id": "ocr/detector/ppocr-dbnet",
  "license": "Apache-2.0",
  "parity_source": {
    "files": [
      "det_mobile.onnx",
      "det.json",
      "ocr_model/ocr_det.axmodel"
    ],
    "repo": "AXERA-TECH/PPOCR_v5",
    "revision": "2a3427a63c410ee6982250769126328ad85eb24a"
  },
  "preprocessing": {
    "known_mismatches": [
      "source inference.yml uses DetResizeForTest resize_long=960, while current markers preprocessing direct-stretches to 960x960",
      "source preprocessor_config.json and inference.yml list mean/std in opposite channel orders; impl-candle-ocr-detectors must resolve this against source or legacy parity fixtures before changing constants",
      "source model names the main input pixel_values in Transformers and x in Paddle inference metadata"
    ],
    "markers_legacy_compat": {
      "channel_order": "bgr",
      "input_shape": [
        1,
        3,
        960,
        960
      ],
      "mean_u8": [
        123.675,
        116.28,
        103.53
      ],
      "output": "probability_map",
      "postprocess": "decode_prob_map",
      "resize": "direct-stretch-to-square",
      "std_u8": [
        58.395,
        57.12,
        57.375
      ]
    },
    "source_inference_yml": {
      "dynamic_shapes": [
        [
          1,
          3,
          32,
          32
        ],
        [
          1,
          3,
          736,
          736
        ],
        [
          1,
          3,
          4000,
          4000
        ]
      ],
      "normalize_mean": [
        0.485,
        0.456,
        0.406
      ],
      "normalize_order": "hwc",
      "normalize_scale": "1./255.",
      "normalize_std": [
        0.229,
        0.224,
        0.225
      ],
      "paddle_input_name": "x",
      "resize_long": 960,
      "to_chw": true
    },
    "source_processor": {
      "channel_first": false,
      "do_normalize": true,
      "do_rescale": true,
      "do_resize": true,
      "image_mean": [
        0.406,
        0.456,
        0.485
      ],
      "image_mode": "BGR",
      "image_std": [
        0.225,
        0.224,
        0.229
      ],
      "limit_side_len": 960,
      "max_side_limit": 4000,
      "model_input_names": [
        "pixel_values",
        "original_image_size"
      ],
      "rescale_factor": 0.00392156862745098
    }
  },
  "provenance": {
    "conversion": "native safetensors copied byte-for-byte; metadata generated",
    "runtime_onnx": "false",
    "source_model_class": "PPOCRv5MobileDetForObjectDetection",
    "source_output": "sigmoid probability map, not logits",
    "source_paddle_input_name": "x",
    "source_transformers_input_name": "pixel_values",
    "source_transformers_output_name": "last_hidden_state"
  },
  "required_files": [
    {
      "path": "model.safetensors",
      "role": "weights"
    },
    {
      "path": "config.json",
      "role": "config"
    },
    {
      "path": "preprocessor_config.json",
      "role": "preprocessor"
    },
    {
      "path": "inference.yml",
      "role": "source-inference-config"
    },
    {
      "path": "README.md",
      "role": "source-card"
    },
    {
      "path": "ocr_pipeline.py",
      "role": "source-example"
    },
    {
      "path": "tensor_map.json",
      "role": "tensor-map"
    },
    {
      "path": "validation.json",
      "role": "validation"
    }
  ],
  "schema_version": 1,
  "source": {
    "files": [
      "README.md",
      "config.json",
      "inference.yml",
      "model.safetensors",
      "ocr_pipeline.py",
      "preprocessor_config.json"
    ],
    "repo": "PaddlePaddle/PP-OCRv5_mobile_det_safetensors",
    "revision": "c5041d225cf951ff06900ab81a3c7d543c45e2ad"
  },
  "tensor_contract": {
    "inputs": [
      {
        "aliases": [
          "x"
        ],
        "dtype": "f32",
        "name": "pixel_values",
        "shape": [
          "batch",
          3,
          960,
          960
        ]
      }
    ],
    "outputs": [
      {
        "aliases": [
          "last_hidden_state"
        ],
        "dtype": "f32",
        "name": "probability_map",
        "shape": [
          "batch",
          1,
          960,
          960
        ]
      }
    ]
  },
  "validated_runtime": [
    {
      "device": "cpu",
      "dtype": "f32"
    }
  ]
}