^jUddlZddlZddlmZddlmZmZmZmZm Z m Z m Z m Z m Z ddlZddlmZmZmZmZmZddlmZe dZe ed<eezZe ed<Gd d eZd efd ZeZeed <GddeZGddeZ Gdde Z!Gdde!Z"Gdde!Z#Gdde!Z$Gdde!Z%Gdde Z&Gdde!Z'Gd d!e!Z(Gd"d#e!Z)Gd$d%e!Z*Gd&d'e!Z+Gd(d)e!Z,Gd*d+e!Z-Gd,d-e!Z.Gd.d/eZ/Gd0d1e/Z0Gd2d3e/Z1y)4N)Path) AnyClassVarDictListLiteralMappingOptional TypeAliasUnion) BaseModel ConfigDictFieldfield_validatormodel_validator)PydanticUndefined)dinov2_windowed_smalldinov2_windowed_basedinov2_registers_windowed_small EncoderName PathLikeStrceZdZdZy)#PretrainWeightsCompatibilityWarninguWarning emitted when ``ModelConfig`` overrides are likely to prevent the variant's published pretrained weights from loading into the model — leaving large portions of the model randomly initialized and typically producing much lower accuracy.N)__name__ __module__ __qualname____doc__X/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/config.pyrrsrrreturncxttdd}t|dd}| |d}| t |Sytj jrytjjjryy#t$r|}||jsd}Y}wxYw#t $rYywxYw) u Detect the best available device **without** initialising the CUDA runtime. ``torch.cuda.is_available()`` creates a CUDA driver context that makes ``_is_in_bad_fork()`` return ``True`` in child processes. This breaks fork-based DDP strategies (e.g. ``ddp_notebook``) in notebook environments. We defer to :func:`torch.accelerator.current_accelerator` (PyTorch ≥ 2.4) when available — it queries the driver through NVML without creating a primary context. On older builds we fall back to ``torch.cuda.is_available()``. ``check_available=True`` is required: without it ``current_accelerator()`` only reports the *compile-time* accelerator, so the default CUDA wheel on a machine without an NVIDIA driver yields ``"cuda"`` and every model build crashes with "Found no NVIDIA driver". The runtime check is NVML-backed and still avoids creating a CUDA context. Builds whose ``current_accelerator`` predates the ``check_available`` kwarg get the same runtime verification via ``torch.accelerator.is_available``. acceleratorNcurrent_acceleratorT)check_availablecpucudamps) getattrtorch TypeError is_availablestr RuntimeErrorr'backendsr()r#r$accels r _detect_devicer1s%5K!+/DdK&  !+DA  5z! zz  ~~&&(  !+-$[-E-E-G E !  s. B B-$B*'B-)B**B-- B98B9DEVICEceZdZUdZeddZeeed<ede de d e fd Z d e d e d d ffd Z xZS) BaseConfigzBase configuration class that validates input parameters against the defined model schema. If any unknown fields are provided, a ValueError is raised listing the unknown and available parameters. forbidTextravalidate_assignment model_configbeforemodevaluesr!cnt|ts|S|jjddk7r|St |j j }t |}||z }|rMdjdt|D}djt|}td|d|d|S)Nr7r5z, c3(K|] }d|d yw)'Nr).0params r z/BaseConfig.catch_typo_kwargs..Ws+]UawaL+]szUnknown parameter(s): z. Available parameter(s): .) isinstancer r9getset model_fieldskeysjoinsorted ValueError)clsr=allowed_paramsprovided_paramsunknown_paramsunknown_params_listallowed_params_lists r catch_typo_kwargszBaseConfig.catch_typo_kwargsLs&'*M      (H 4MS--2245f+(>9 "&))+]f^F\+]"] "&))F>,B"C ()<(==WXkWllmn  rnamevalueNc|jds|t|jvrt|||yt d|d)N_zUnknown attribute: 'z'.) startswithtyperHsuper __setattr__rL)selfrTrU __class__s r r[zBaseConfig.__setattr__^sG ??3 44:+B+B#B G e , /vR899r)rrrrrr9r__annotations__r classmethodrrSr-r[ __classcell__)r]s@r r4r4Dsm *4(X\)]L(:&](#ss$ ::C:D::rr4c|eZdZUeed<eeed<eed<dZeed<ee ded<eed<eed <eed <eed <eed <eed <dZ eed<dZ eed<e ddZ eed<dZeed<dZeed<dZeed<dZeed<e ddZeed<dZeed<dZeeed<eZeed <eed!<d"Zeed#<d$Zeed%<d$Zeed&<dZeed'<eed(<dZeed)<d*Z eed+<d$Z!eed,<d$Z"eed-<dZ#eed.<d$Z$eed/<dZ%eed0<d$Z&eed1<d$Z'eed2<e e(3Z)eeed4<d5Z*eed6<d7Z+eed8<d$Z,eed9<d$Z-eed:<d;Z.eed<<e dd=>Z/eeed?<e0d@AdJdCZ1e0d@AdJdDZ2e0d@AdJdEZ3e4ddFAe5dGedzdBedzfdHZ6e4d dFAe5dGe7dBefdIZ8y)K ModelConfigencoderout_feature_indexes dec_layersT two_stageP3P4P5projector_scale hidden_dim patch_size num_windows sa_nheads ca_nheads dec_n_points, num_queries num_selectg?defaultgepostprocess_trace_alpha bbox_reparamlite_refpoint_refine layer_normamp num_channelsZ num_classesNpretrain_weightsdevice resolution group_detrFgradient_checkpointingcompilefused_optimizerpositional_encoding_size ia_bce_loss? cls_loss_coefsegmentation_headuse_grouppose_keypointskeypoint_cross_attninter_instance_kp_attn grouppose_keypoint_dim_downscaledual_projectordual_projector_kp_onlydefault_factorynum_keypoints_per_classrnum_decoder_registersmask_downsample_ratio backbone_lorafreeze_encoderz Apache-2.0licensezName of the model class stored in training checkpoints (e.g. ``"RFDETRLarge"``). Set automatically by ``RFDETR.train()`` before saving. Used by ``RFDETR.from_checkpoint()`` to resolve the correct subclass directly without inspecting ``pretrain_weights``.rw description model_nameafterr;r!cZd|jvrtjdtd|S)aEmit DeprecationWarning when cls_loss_coef is explicitly set on ModelConfig. ``cls_loss_coef`` ownership is moving to ``TrainConfig`` (Item #3, v1.7). Setting it on ``ModelConfig`` is deprecated. Use ``TrainConfig(cls_loss_coef=...)`` instead. rz}ModelConfig.cls_loss_coef is deprecated since v1.7.0 and will be removed in v1.9.0. Set cls_loss_coef on TrainConfig instead. stacklevelmodel_fields_setwarningswarnDeprecationWarning)r\s r $_warn_deprecated_model_config_fieldsz0ModelConfig._warn_deprecated_model_config_fieldss2 d33 3 MM<"    rcd|jvsd|jvr|St|}|jdj}|jdj}|jdj}t |t r t |t rt |t s|S|||zk(r|j |jz|_|S)a?Auto-update positional_encoding_size when resolution is explicitly provided. When a user provides a custom ``resolution`` at construction time (e.g., ``RFDETRLarge(resolution=640)``), ``positional_encoding_size`` is updated proportionally, provided the class-default PE is formula-derived (``default_pe == default_resolution // patch_size``). Configs with a pretrained-specific PE (e.g., ``RFDETRBaseConfig`` with ``positional_encoding_size=37`` for DINOv2's native 518 px grid, while ``resolution=560``) are left unchanged. rrrm) rrYrHrwrEintrrmr)r\rMdefault_resolution default_pedefault_patch_sizes r _sync_pe_with_resolutionz$ModelConfig._sync_pe_with_resolutions t44 48RVZVkVk8kK4j --l;CC%%&@AII  --l;CC -s3j#.0#6K +/AA A,0OOt,ND ) rct|}|j}d|v}|r]|jQ|jdj}|t ur.|,t j|jd|dtd|S|rS|jG|jdj}|*|t ur"|j|}|j|k7r|S|S|jdj}||t ur|Sg}d} d} | D]p} | |vr|jj| } | | jr6| j} | t urKt|| }|| k7s]|j| || fr| D]} | |vr|jj| } | | jr6| j} | t ust| t s[t|| }t|t sx|| kDs~|j| || fd|vrr|j"rf|jjd}|I|js9|j}|t ur%t|d}||k7r|jd||f|r`|jdj}d j%d |D}t j|jd |d |d td|S)ufWarn when overrides are likely to prevent published pretrained weights from loading. Three cases: 1. ``pretrain_weights`` was explicitly set to ``None`` and the variant has a non-``None`` default → warn that the model is being initialised from scratch. 2. ``pretrain_weights`` was explicitly set to a non-``None`` custom path → suppress the architecture-override check (we cannot know the architecture stored in a user-supplied checkpoint at config time). The load-time partial-load detector in :func:`rfdetr.models.weights.load_pretrain_weights` covers this case by inspecting the checkpoint contents directly. 3. ``pretrain_weights`` is the variant's published default → check architecture-affecting fields against the variant defaults and emit a single consolidated warning listing every load-breaking override. The warning class is :class:`PretrainWeightsCompatibilityWarning` (a :class:`UserWarning` subclass), silenceable via the standard ``warnings.filterwarnings`` machinery. rz was instantiated with pretrain_weights=None. The model will be initialised from scratch, which typically produces lower accuracy than fine-tuning from the published checkpoint (z).rr)rcrlrernrorprqrdrkrzr{r|rfrmrr)rsrr c3<K|]\}}}d|d|d|dyw)z z: z (variant default: )Nr)rArTcurrentrws r rCz;ModelConfig._warn_pretrain_compatibility..`s3LbDRY[b"TF"WK':7+QGsz~ was instantiated with overrides that differ from the variant defaults in ways that prevent the published pretrained weights (z) from loading correctly: aF Loading the checkpoint with this configuration will leave significant portions of the model randomly initialised, which typically produces lower accuracy. To suppress this warning: revert the override(s), pick a variant whose defaults match, or pass pretrain_weights=None to acknowledge that you intend to train from scratch.)rYrrrHrwrrrrr expand_pathrF is_requiredr)appendrErrrJ)r\rM fields_setpretrain_user_setdefault_pretrain_default_pretrain_expanded_default_class_default_pretrain overridesbreaking_fieldsbreaking_on_increaserT field_inforwr _mdr_info _mdr_default _mdr_currentliness r _warn_pretrain_compatibilityz(ModelConfig._warn_pretrain_compatibilitysE(4j** .*< !6!6!>"//0BCKK '88=M=Y ||n%#$4"6b:8 K !6!6!B!$ 0 01C D L L  ,1BJ[1[$'OO4E$F!((,==K #&"2"23E"F"N"N " *.EIZ.ZK02 , (1  $ ;D:%))--d3J!Z%;%;%= ((G++dD)G'!  $!9: ;) ;D:%))--d3J!Z%;%;%= ((G++:gs3KdD)G'3'Gg,=  $!9: ; #j 0T5K5K((,,-DEI$Y-B-B-D(00 '88#*41H#IL#|3!((*A<d?Z3e#ed@<dZ4e#edA<dZ5e edB<dCZ6e edD<dZ7e edE<ed dFGZ8e dHedI<d?Z9e#edJ<dZ:e edK<dLZ;e edM<d?Ze#edQ<d?Z?e#edR<d?Z@e#edS<d?ZAe#edT<dZBeeCedU<dZDeeCedV<dZEee(eCedW<d?ZFe#edX<d?ZGe#edY<dZZHe ed[<dZIe ed\<dZJe#ed]<dZKeeLeCeMfed^<d_ZNe d`eda<d?ZOe#edb<eddcGZPeeMedd<eQdefddhZReSdPdifeTdjeMdgeMfdkZUeSdIdifeTdjeMdgeMfdlZVd ZWeCedm<dnZXe edo<dZYee edp<d?ZZe#edq<d Z[eCedr<dZ\e]e eCfeds<dZ^e edt<d?Z_e#edu<dvZ`e dwedx<d!Zae edy<d?Zbe#edz<d?Zce#ed{<d?Zde#ed|<d?Zee#ed}<dZfe#ed~<dZge#ed<dZhee#ed<dZiee#ed<dZjee ed<eSd defeTde e d zdge e d zfdZkeSd dddefeTde dge fdZleSddefeTde dge fdZmeSd"defeTde dge fdZneSdBd\defeTde dge fdZoeSddefeTdee dgee fdZpeSd:d= 1, or 'auto'.rLrMrs r validate_batch_sizezTrainConfig.validate_batch_size+s( ;H q5BC Crc&|dkr td|S)zHValidate accumulation, target-effective batch, and max targets are >= 1.rzagrad_accum_steps, auto_batch_target_effective, and auto_batch_max_targets_per_image must be >= 1.r}r~s r validate_positive_train_stepsz)TrainConfig.validate_positive_train_steps5s" q5s rcJd|cxkrdkstdtd|S)z.Validate auto_batch_ema_headroom is in (0, 1].rrz*auto_batch_ema_headroom must be in (0, 1].r}r~s r validate_ema_headroomz!TrainConfig.validate_ema_headroomAs1A  IJ JIJ JrcJd|cxkrdkstdtd|S)z'Validate smooth_alpha is in [0.0, 1.0).rurz#smooth_alpha must be in [0.0, 1.0).r}r~s r validate_smooth_alphaz!TrainConfig.validate_smooth_alphaIs1q3BC CBC Crc&|dkr td|S)z"Validate interval fields are >= 1.rzInterval fields must be >= 1.r}r~s r validate_positive_intervalsz'TrainConfig.validate_positive_intervalsQs q5<= =rc*||dkr td|S)z)Validate prefetch_factor is None or >= 1.rz+prefetch_factor must be >= 1 when provided.r}r~s r validate_prefetch_factorz$TrainConfig.validate_prefetch_factorYs  =QUJK Krc||Stjjtjjtj|S)ujExpand and normalize dataset/output directory paths via ``os.fspath`` → ``expanduser`` → ``realpath``.)rrrrrr~s r expand_pathszTrainConfig.expand_pathsas: 9Hww 2 2299Q< @AArc4||Stj|S)aNormalise the resume checkpoint value to ``str`` without resolving it. Unlike ``dataset_dir``/``output_dir``, ``resume`` is forwarded verbatim to PyTorch Lightning's ``trainer.fit(ckpt_path=...)``, which also accepts sentinel values such as ``"last"``. Running ``os.path.realpath`` would rewrite those sentinels into spurious absolute paths, so this validator only coerces the type (``Path`` -> ``str``) and leaves the value untouched. )rrr~s r _coerce_resume_pathzTrainConfig._coerce_resume_pathis 9Hyy|r)r!r)srrrrrr9rr^rrrrrrrrrrr r!r rr"r#r$rr&r'r(r)r*r+r,rrrrrtrr-rr.r/r0r1r2r7r8r;r<r=r>r?r@rArBrFrGrHrIrJrMrNrOrPrQrRr-rSrTrUrrVrWrXrYrrr[r\r]rrarr_rdrgr#rhrirjrkrlr rmrnrqrrrsrtrurvrwrxryrzr{rrrrrrrrrrr rrwsa *4(X\)]L(:&]BJ()Jgfo%)c')),/$c/%(U(FC$(FH[ !(IuGSGS$RA66!!2c2L%M5 ## ##IuJKM5J%*4%@c@#$5$)**())$%E%.2et+2@JL'<=J!%$%+&&&J &K OT ). $.GT  KL%16 Q 2Iw-. !ND #%S%&+e+#(D(6:L(7>23:KE4FDGT!GXc]!C#'+K$s)$+Hd#t#M3M3"&4&+/Jc3h(/:?'"67?$$  0 E8C= '"#2^(34[x0cc1$KM5D(3-GTHc GU38_ IsIt.4L'*+4M5#t# %%#t#"'4'!d!"t"!%J%)--%)OXc])\0C'&/$9cGFO>S19;]dkcc.W=e> ^'2e3 *O'JCCK &W5#8C=6 ]Lx@B[4/BC$JBAB XH- K$$6 3: . rrcleZdZUdZdZeeed<dZeed<dZ e ed<dZ e ed<d Z e ed <d Z eed <y) SegmentationTrainConfigaTraining configuration for instance segmentation models. Extends :class:`TrainConfig` with segmentation-specific loss coefficients. Attributes: num_select: Maximum number of predictions to keep per image. ``None`` uses the model default. mask_point_sample_ratio: Number of points sampled per mask for point-based mask loss computation. mask_ce_loss_coef: Cross-entropy loss weight for mask prediction. mask_dice_loss_coef: Dice loss weight for mask prediction. cls_loss_coef: Classification loss weight. Defaults to ``1.0`` to match the effective pre-v1.7 value (the v1.7 TrainConfig ownership migration silently activated a dormant ``5.0``; this field restores the correct weight). To reproduce pre-fix segmentation behaviour pass ``cls_loss_coef=5.0`` explicitly. segmentation_head: Whether to attach the segmentation head. Nrtrmask_point_sample_ratiog@mask_ce_loss_coefmask_dice_loss_coefrrTr)rrrrrtr rr^rrrrrrrrrr rrxsM&!%J $#%S%"u"!$$M5"t"rrceZdZUdZdZeed<dZeed<dZeed<dZ eed<dZ eed <d Z eed <e d d Z eed<y)KeypointTrainConfigaETraining configuration for keypoint detection models. Extends :class:`TrainConfig` with keypoint-specific loss coefficients and metric-smoothing defaults tuned for the NLL-Cholesky keypoint head, which produces noisy per-epoch OKS metrics during early fine-tuning. Attributes: cls_loss_coef: Classification loss weight. keypoint_l1_loss_coef: L1 regression loss weight for keypoint coordinates. keypoint_findable_loss_coef: Loss weight for the keypoint visibility head. keypoint_visible_loss_coef: Loss weight for the keypoint visibility score. keypoint_nll_loss_coef: NLL-Cholesky loss weight. Restored to ``1.0`` to align with the other keypoint loss terms (``keypoint_l1_loss_coef``, ``keypoint_findable_loss_coef``, ``keypoint_visible_loss_coef``). Previously set to ``0.5`` to dampen OKS@75 oscillation; reverted as the under-weighting was not beneficial in practice. smooth_alpha: EMA smoothing factor for :class:`BestModelCallback` metric comparison. Overrides the :class:`TrainConfig` default of ``0.0`` (disabled) to ``0.5``, which balances responsiveness and noise suppression for noisy keypoint mAP curves. skip_best_epochs: Number of epochs to skip before checkpoint selection begins. Overrides the :class:`TrainConfig` default of ``0`` to ``10`` because ``val/keypoint_map_50_95`` under the NLL-Cholesky loss is noisy in early fine-tuning and can lock checkpoint selection to a transient peak. g@rrr.r/r0rr1g?r(r%rrvr'N)rrrrrrr^r.r/r0r1r(rr'rrrr rrs[4M5#$5$)**())$'E'L%!"3c3rr)2rrpathlibrtypingrrrrrr r r r r*pydanticr rrrr pydantic_corerrr^r-rrfrr1r2r4rbrrrrrrrrrrr r rrrrrrrr rsl ZZZ SS+ !st Ytt Y#+ ##L::B{E*{E| '{'( B"2 B A' A B( B C) C . -  *  +  ,  +  ,  - "2,~*~B#k#8!4+!4r