bundle.json

{
  "architecture_id": "ocr.ppocr-ctc",
  "bundle_id": "ocr/recognizer/ppocr-ctc",
  "implementation_notes": {
    "current_markers_decode": "argmax, collapse repeats, drop blank index 0",
    "embedded_config": "runtime should load preprocessor_config.json character_list as-is; do not prepend another blank token, append another space token, or depend on AXERA ppocrv5_dict.txt",
    "head_out_channels": 18385,
    "hidden_size": 120,
    "source_dynamic_shapes": [
      [
        1,
        3,
        48,
        160
      ],
      [
        1,
        3,
        48,
        320
      ],
      [
        8,
        3,
        48,
        3200
      ]
    ]
  },
  "license": "Apache-2.0",
  "parity_source": {
    "files": [
      "rec_mobile_sim_static.onnx",
      "rec.json",
      "ppocrv5_dict.txt"
    ],
    "repo": "AXERA-TECH/PPOCR_v5",
    "revision": "2a3427a63c410ee6982250769126328ad85eb24a"
  },
  "preprocessing": {
    "known_mismatches": [
      "source preprocessor_config.json does not explicitly serialize image_mean or image_std; current markers keeps PP-OCR/AXERA (-0.5)/0.5 normalization until source parity fixtures confirm a change",
      "source preprocessor_config.json says do_convert_rgb=true while inference.yml DecodeImage uses BGR; current markers converts to RGB and then packs BGR-NCHW",
      "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,
        48,
        320
      ],
      "mean_u8": [
        127.5,
        127.5,
        127.5
      ],
      "output": "ctc_probabilities",
      "postprocess": "decode_ctc",
      "resize": "aspect-preserving height resize with right padding",
      "std_u8": [
        127.5,
        127.5,
        127.5
      ]
    },
    "source_inference_yml": {
      "decode_image_mode": "BGR",
      "dynamic_shapes": [
        [
          1,
          3,
          48,
          160
        ],
        [
          1,
          3,
          48,
          320
        ],
        [
          8,
          3,
          48,
          3200
        ]
      ],
      "keep_keys": [
        "image",
        "label_ctc",
        "label_gtc",
        "length",
        "valid_ratio"
      ],
      "paddle_input_name": "x",
      "rec_resize_image_shape": [
        3,
        48,
        320
      ]
    },
    "source_processor": {
      "do_convert_rgb": true,
      "do_normalize": true,
      "do_pad": true,
      "do_rescale": true,
      "do_resize": true,
      "max_image_width": 3200,
      "pad_size": {
        "height": 48,
        "width": 320
      },
      "size": {
        "height": 48,
        "width": 320
      }
    }
  },
  "provenance": {
    "conversion": "native safetensors copied byte-for-byte; metadata generated",
    "runtime_onnx": "false",
    "source_model_class": "PPOCRV5MobileRecForTextRecognition",
    "source_output": "softmax CTC probabilities, not raw 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-and-charset"
    },
    {
      "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_rec_safetensors",
    "revision": "485a9a9781535ad477de3b1eea5f96a2e7300d8d"
  },
  "tensor_contract": {
    "inputs": [
      {
        "aliases": [
          "x"
        ],
        "dtype": "f32",
        "name": "pixel_values",
        "shape": [
          "batch",
          3,
          48,
          320
        ]
      }
    ],
    "outputs": [
      {
        "aliases": [
          "last_hidden_state"
        ],
        "dtype": "f32",
        "name": "ctc_probabilities",
        "sequence_length": "input_width / 8 for the fixed 320-wide smoke contract",
        "shape": [
          "batch",
          40,
          18385
        ]
      }
    ]
  },
  "validated_runtime": [
    {
      "device": "cpu",
      "dtype": "f32"
    }
  ],
  "vocab": {
    "blank_index": 0,
    "blank_token": "blank",
    "body_size": 18383,
    "ctc": {
      "blank_index": 0,
      "collapse_repeated_non_blank": true,
      "confidence": "mean selected timestep probability",
      "output_kind": "softmax probabilities"
    },
    "first_tokens": [
      "blank",
      "\u3000",
      "\u4e00",
      "\u4e59",
      "\u4e8c",
      "\u5341",
      "\u4e01",
      "\u5382"
    ],
    "head_out_channels": 18385,
    "inference_body_matches": true,
    "inference_character_dict_size": 18383,
    "last_tokens": [
      "\ud83d\udd61",
      "\ud83d\udd62",
      "\ud83d\udd63",
      "\ud83d\udd64",
      "\ud83d\udd65",
      "\ud83d\udd66",
      "\ud83d\udd67",
      " "
    ],
    "max_token_unicode_scalars": 5,
    "multi_scalar_token_count": 11,
    "runtime_policy": "embedded_config",
    "source": "preprocessor_config.json character_list",
    "space_token_index": 18384,
    "vocab_size": 18385
  }
}