opendatalab/MinerU · critical · FileNotFoundError

{} is not existed.

Error message

{} is not existed.

What it means

Raised by BaseOCRV20.read_pytorch_weights when the OCR weights file path does not exist on disk, before any safetensors/torch load is attempted. The message ('{} is not existed.') is inherited from PaddleOCR-style code.

Source

Thrown at mineru/model/utils/pytorchocr/base_ocr_v20.py:85

        except TypeError:
            return torch.load(weights_path, map_location="cpu")

    @staticmethod
    def _normalize_ppocrv6_state_dict(weights, weights_path):
        """归一化 HF OCR safetensors 的外层 `model.` 前缀。"""
        if not BaseOCRV20._is_safetensors_path(weights_path):
            return weights
        if not any(key.startswith("model.") for key in weights.keys()):
            return weights
        return {
            key.removeprefix("model."): value
            for key, value in weights.items()
        }

    def read_pytorch_weights(self, weights_path):
        """读取 PyTorch OCR 权重，并兼容 PP-OCRv6 safetensors。"""
        if not os.path.exists(weights_path):
            raise FileNotFoundError('{} is not existed.'.format(weights_path))
        weights = self._load_weight_file(weights_path)
        return self._normalize_ppocrv6_state_dict(weights, weights_path)

    def get_out_channels(self, weights):
        """从权重结构推断识别输出通道数。"""
        if "head.head.weight" in weights:
            # PP-OCRv6 safetensors 的识别分类层固定命名为 head.head。
            return weights["head.head.weight"].shape[0]
        if list(weights.keys())[-1].endswith('.weight') and len(list(weights.values())[-1].shape) == 2:
            out_channels = list(weights.values())[-1].numpy().shape[1]
        else:
            out_channels = list(weights.values())[-1].numpy().shape[0]
        return out_channels

    def load_state_dict(self, weights):
        self.net.load_state_dict(weights)
        # print('weights is loaded.')

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Verify the path exists: ls -l <weights_path>; fix typos or absolute-vs-relative mistakes.
2. Run mineru's model download command (e.g. `mineru-models-download` or pipeline auto-download) so OCR weights are fetched to the models root.
3. If offline, download weights on a connected machine and copy them to the exact expected path.
4. Check the models_path configuration/env used to construct weights_path.

Example fix

# before
ocr = TextDetector(model_path="/models/ocr_det_v6.safetensors")  # file absent

# after
import os
assert os.path.exists(model_path), f"missing OCR weights: {model_path}"
ocr = TextDetector(model_path=model_path)

import os
if not os.path.isfile(weights_path):
    raise FileNotFoundError(f"OCR weights missing: {weights_path}; run model download first")

try:
    ocr_model = BaseOCRV20(...)
except FileNotFoundError as e:
    logger.error("weights not found: %s", e)
    raise SystemExit("Run mineru model download before inference") from e

When it happens

Trigger: Initializing the OCR model with a weights_path pointing to a missing .pth/.safetensors file; path built by joining a model root with a wrong filename; models directory never downloaded.

Common situations: Model auto-download skipped or failed (offline environment, blocked download endpoint); manually specified model dir with typo; running in a container where the models volume is not mounted at the expected path.

Related errors

• {} is not found.
• {} is not existed.
• architecture {file_name} is not in arch_config.yaml
• Language {lang} not supported. Allowed values: {allowed_valu
• Missing middle json file: {middle_json_path}

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/2c2c6cc988812379. Report an issue: GitHub.