| # Layout Analysis |
|
|
| D-FINE Stage-2 B2 layout model with 12 classes. Detects page regions; |
| text recognition and reading order are handled by OXR. |
|
|
| ## Files |
|
|
| - `last.pth`: inference-only model weights (no optimizer or training state). |
| - `metadata.json`: model identity, architecture, preprocessing, class IDs, |
| checkpoint size and SHA-256 (schema version 1). |
|
|
| `model_version` is a public release version (`major.minor.patch`), not a |
| training-run identifier. |
|
|
| ## Usage |
|
|
| Use the LA inference service in the OXR code repository at |
| `services/layout_analyze/`. Mount this directory into the service and set |
| `MODEL_PATH` to the mounted `last.pth` path. |
|
|
| The matching LA service reads `metadata.json` next to the checkpoint at startup |
| and validates its supported architecture, preprocessing, class IDs and weight |
| integrity. Both files must be mounted. `MODEL_METADATA_PATH` can override the |
| metadata path. Unsupported metadata or a mismatched checkpoint fails startup. |
| This is not a Transformers `AutoModel` configuration. Device, precision and |
| CUDA Graph settings are selected by the service, not by the model metadata. |
|
|
| Checkpoint verification: |
|
|
| ```bash |
| shasum -a 256 layout_analysis/last.pth |
| # dff78ccf18439955362ca44a69eb28fb0b40222a7d278860589682de0176dc17 |
| ``` |
|
|
