^j>dZddlmZddlZddlmZddlmZddlm Z ddl m Z m Z ddl mZeZd Zd Zed Gd dZGddeZGddeZddZ d ddZdddZy)aModel weights abstraction and download system. Provides forward-compatible pattern for model weights across rf-detr and rf-detr-plus packages. External packages (like rf-detr-plus) should inherit from ModelWeightsBase to ensure compile-time interface compatibility. Critical Strategic Decisions: 1. **Standalone first**: Check local ModelWeights before lazy-importing external packages 2. **Compile-time safety**: Inheritance-based compatibility via ModelWeightsBase 3. **Clean abstraction**: Enum values ARE ModelWeightAsset dataclass instances 4. **Backward compatible**: Legacy OPEN_SOURCE_MODELS dict maintained 5. **Offline testable**: All I/O operations mockable Download Priority Order: 1. Local ModelWeights.from_filename() - rf-detr's built-in models 2. rfdetr_plus.assets.ModelWeights.from_filename() - lazy import if not found locally 3. PLATFORM_MODELS dict - legacy fallback for backward compatibility ) annotationsN) dataclass)Enum)_IS_RFDETR_PLUS_AVAILABLE)_download_file_validate_file_md5) get_loggerRF_HOMEz~/.roboflow/modelsT)frozenc4eZdZUdZded<ded<dZded<y)ModelWeightAssetaDataclass representing a model asset with download information. This is the standard format for model assets across rf-detr packages. Both rf-detr and rf-detr-plus should use this structure for compatibility. Attributes: filename: The local filename for the model weights url: The download URL md5_hash: The expected MD5 hash for integrity validation (None if not available) Example: >>> asset = ModelWeightAsset( ... filename='rf-detr-base.pth', ... url='https://storage.googleapis.com/rfdetr/rf-detr-base-coco.pth', ... md5_hash='b4d3ce46099eaed50626ede388caf979' ... ) strfilenameurlN str | Nonemd5_hash)__name__ __module__ __qualname____doc____annotations__rf/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/assets/model_weights.pyr r (s$M HHjrr ceZdZdZd dZed dZed dZed dZe ddZ e ddZ e ddZ e dd Z y )ModelWeightsBaseaBase class for model weight registries. This base class ensures compile-time compatibility between rf-detr and rf-detr-plus. Both packages should inherit from this class to ensure they have the same interface. Each enum member's value must be a ModelWeightAsset instance. Example inheritance: >>> from rfdetr.assets import ModelWeightAsset >>> class MyModelWeights(ModelWeightsBase): ... MODEL_NAME = ModelWeightAsset( ... "model.pth", ... "https://example.com/model.pth", ... "abc123" ... ) Example usage: >>> from rfdetr.assets.model_weights import ModelWeights >>> asset = ModelWeights.from_filename("rf-detr-base.pth") >>> asset.filename 'rf-detr-base.pth' >>> asset.url # doctest: +SKIP 'https://storage.googleapis.com/rfdetr/rf-detr-base-coco.pth' c>tj|}||_|S)N)object__new___value_)clsassetobjs rrzModelWeightsBase.__new__[snnS!  rc.|jjS)z6Get the filename from the underlying ModelWeightAsset.valuerselfs rrzModelWeightsBase.filenameazz"""rc.|jjS)z1Get the URL from the underlying ModelWeightAsset.)r&rr's rrzModelWeightsBase.urlfszz~~rc.|jjS)z6Get the MD5 hash from the underlying ModelWeightAsset.)r&rr's rrzModelWeightsBase.md5_hashkr)rc`|D])}|jj|k(s|jcSy)aGet ModelWeightAsset by filename. Args: filename: The model filename (e.g., 'rf-detr-base.pth') Returns: ModelWeightAsset instance if found, None otherwise Example: >>> asset = ModelWeights.from_filename('rf-detr-base.pth') >>> asset.url 'https://storage.googleapis.com/rfdetr/rf-detr-base-coco.pth' Nr%)r!rmembers r from_filenamezModelWeightsBase.from_filenameps4 $F||$$0||# $rcD|j|}|r |jSdS)zGet download URL for a model by filename. Args: filename: The model filename Returns: URL string if found, None otherwise N)r.rr!rr"s rget_urlzModelWeightsBase.get_urls%!!(+!uyy+t+rcD|j|}|r |jSdS)zGet expected MD5 hash for a model by filename. Args: filename: The model filename Returns: MD5 hash string if available, None otherwise N)r.rr0s rget_md5zModelWeightsBase.get_md5s%!!(+!&u~~0D0rcT|Dcgc]}|jjc}Scc}w)zbList all available model filenames. Returns: List of model filenames r%)r!r-s r list_modelszModelWeightsBase.list_modelss"588& %%888s%N)r"r returnz'ModelWeightsBase'r6r)r6r)rrr6zModelWeightAsset | None)rrr6r)r6z list[str])rrrrrpropertyrrr classmethodr.r1r3r5rrrrrAs2 ####& , , 1 199rrcPeZdZdZedddZedddZedd d Zed d d ZedddZ edddZ edddZ edddZ edddZ edddZed d!d"Zed#d$d%Zed&d'd(Zed)d*d+Zed,d-d.Zed/d0d1Zy2)3 ModelWeightsa-Enumeration of available RF-DETR model assets. Inherits from ModelWeightsBase to ensure compatibility with rf-detr-plus. Each enum member's value is a ModelWeightAsset instance containing: - filename: The local filename for the model weights - url: The download URL - md5_hash: The expected MD5 hash for integrity validation Example: >>> asset = ModelWeights.RF_DETR_BASE >>> asset.filename 'rf-detr-base.pth' >>> asset.url 'https://storage.googleapis.com/rfdetr/rf-detr-base-coco.pth' zrf-detr-base.pthz;https://storage.googleapis.com/rfdetr/rf-detr-base-coco.pth b4d3ce46099eaed50626ede388caf979zrf-detr-base-o365.pthz]https://storage.googleapis.com/rfdetr/top-secret-1234/lwdetr_dinov2_small_o365_checkpoint.pth d93f4921ccbb0f0a2e4364bed290892bzrf-detr-base-2.pthz8https://storage.googleapis.com/rfdetr/rf-detr-base-2.pth 462f4d9df407ddc1812f42614040e913zrf-detr-large.pthz7https://storage.googleapis.com/rfdetr/rf-detr-large.pth 992c8e862aa733a7bb2777e45d49f1a0zrf-detr-large-2026.pthz>> import os >>> _ = os.environ.pop("RF_HOME", None) # ensure default >>> expected = os.path.normpath(os.path.expanduser("~/.roboflow/models")) >>> get_model_cache_dir() == expected True >>> os.environ["RF_HOME"] = "~/rfdetr_cache" >>> expected = os.path.normpath(os.path.expanduser("~/rfdetr_cache")) >>> get_model_cache_dir() == expected True >>> del os.environ["RF_HOME"] )osenvironget_RF_HOME_ENV_VAR_DEFAULT_CACHE_DIRpathabspath expanduser) cache_dirs rget_model_cache_dirrfs<: /1CDI 77??277--i8 99rcd}tjj|}tj |}|t r ddlm}|j |}||j}|r |jnd}n ddl m } || vry| |}d}tjj|rF|sD|rA|r?t!||st"j%d|dyt"j'd|d yt"j'd |t)||| y#t$r}|jdvrYd}~d}~wwxYw#ttf$rYywxYw) aDownload pretrained weights with optional MD5 validation. Download Priority Order: The function searches for models in the following order, stopping at the first match: 1. **Local ModelWeights** (primary source): - Checks rf-detr's built-in ModelWeights enum - Ensures rf-detr works completely standalone - No unnecessary imports or performance overhead 2. **External packages** (lazy import): - Only attempts import if model not found locally - Tries rf-detr-plus.assets.ModelWeights if installed - Gracefully handles missing packages (ImportError/AttributeError) 3. **Legacy platform models** (backward compatibility): - Falls back to PLATFORM_MODELS dict for older models - Maintains compatibility with existing deployments - No MD5 validation for legacy models Args: pretrain_weights: Name of the pretrained weights file (e.g., 'rf-detr-base.pth') redownload: Force re-download even if file exists validate_md5: Whether to validate MD5 hash of downloaded file Example: >>> download_pretrain_weights('rf-detr-base.pth') # doctest: +SKIP Downloading pretrained weights for rf-detr-base.pth Nr)r;>rfdetr_plus.assets rfdetr_plus)PLATFORM_MODELSzExisting file u has incorrect MD5 hash. It may be a user-provided checkpoint or a corrupted/tampered file — skipping re-download to avoid overwriting it. To force a fresh download of the original weights, pass redownload=True.zFile z& already exists with correct MD5 hash.z#Downloading pretrained weights for )rr expected_md5)r]rbbasenamer;r.rrfdetr_plus.assetsModuleNotFoundErrornamerrrfdetr.platform.downloadsrj ImportErrorKeyErrorexistsrloggerwarninginfor) pretrain_weights redownload validate_md5r" model_namePlusModelWeightsexrrkrjs rdownload_pretrain_weightsr}1seD&*E!!"23J  & &z 2E }2  K$22:>E  ii)5u~~4   A0!*-CL  ww~~&' L%&6 E$%5$67__  e$4#55[\] KK56F5GHI !!I# wwCCD $X&   s0D7 D:D: D7D22D7:E  E ctjj|s|rtd|ytjj |}t j |}| |jtjd|dyt||js,d|d|d}|r t|tj|ytjd |y) aValidate MD5 hash of pretrained weights file. Args: pretrain_weights: Path to the pretrained weights file strict: If True, raise error on validation failure. If False, just warn. Returns: True if validation passes or no hash is available, False otherwise Raises: ValueError: If strict=True and validation fails FileNotFoundError: If strict=True and file doesn't exist z#Pretrained weights file not found: FzNo MD5 hash available for z, skipping validationTzMD5 hash validation failed for zf. The file may be corrupted or tampered with. Consider re-downloading with download_pretrain_weights('z', redownload=True)zMD5 validation passed for ) r]rbrsFileNotFoundErrorrlr;r.rrtdebugr ValueErrorru)rwstrictrzr" error_msgs rvalidate_pretrain_weightsrs 77>>* + #&IJZI[$\] ]!!"23J  & &z 2E }. 1*=RST . ?-.>-?@GGQlRe g  Y' ' NN9 % LL-.>-?@A rr7)FT)rwrrxboolryrr6None)F)rwrrrr6r)r __future__rr] dataclassesrenumrrfdetr.platformrrfdetr.utilities.filesrrrfdetr.utilities.loggerr rtr`rar rr;rfr}rrrrrs $# !5E. ) $   0d9td9Nd#dP:FYYYY Yx)r