^j *dZddlZddlZddlmZmZmZddlZddlmZerddl m Z m Z ddl m Z mZddlmZddlmZdd lmZmZmZmZmZmZmZmZdd lmZdd lmZdd l m!Z!dd l"m#Z#ddl$m%Z%ddl&m'Z'ddl(m)Z)m*Z*e'Z+dejXde-dejXfdZ.dej^de-dej^fdZ0dejbddfdZ2GddejbZ3d%dZ4d%dZ5de fddd ed!d"ede3fd#Z6e fddd d!d"ede7ee#ffd$Z8y)&z$LW-DETR model and criterion classes.N) TYPE_CHECKINGCallableOptional)nn) ModelConfig TrainConfig)MODEL_DEFAULTS ModelDefaults) BuilderArgs)build_backbone) SetCriterion dice_loss dice_loss_jitposition_supervised_losssigmoid_ce_losssigmoid_ce_loss_jitsigmoid_focal_losssigmoid_varifocal_loss)SegmentationHead) build_matcher)MLP) PostProcess)build_transformer) get_logger) NestedTensornested_tensor_from_tensor_listlinear num_classesreturnc.|jjd}ttj||z }|jj j |dd|}|j,|jj j |d|nd}tj|j||du}tj5|jj||'|j|jj|ddd|jj|j_|j1|j%|jj|j_|S#1swYmxYw)uReturn a new :class:`~torch.nn.Linear` resized to *num_classes* outputs. Tiles the existing weight rows when *num_classes* is larger than the current output size, or truncates them when smaller. The returned module has ``out_features == num_classes`` so that ``nn.Linear`` metadata stays consistent with the actual weight shape — a requirement for correct ONNX export and ``torch.jit.trace`` serialisation. Args: linear: Source linear layer whose weights are used as the starting point. num_classes: Target number of output features. Returns: A new :class:`~torch.nn.Linear` with ``in_features`` unchanged and ``out_features == num_classes``. rN)bias)weightshapeintmathceildetachrepeatr"rLinear in_featurestorchno_gradcopy_ requires_grad)rrbase num_repeats new_weightnew_bias new_linears _/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/rfdetr/models/lwdetr.py_resize_linearr68s@ ==  q !Ddii d 234K%%'..{A>| LJIOI`v{{!!#**;7 EfjH6--{QUAUVJ , +  JOO$? OO ! !( +,'-mm&A&AJ# {{:??#>(. (A(A % ,,s AF  F parameternum_rowsc|jd}||k(r|S|dk(r1|jj|g|jdd}nYtt j ||z gdg|j dz z}|jj|d|}tj|j|jS)zQReturn a parameter with the first dimension resized by tiling or truncating rows.rr!N)r/) r$r( new_zerosr%r&r'dimr)r Parametercloner/)r7r8 current_rowsnew_datarepeatss r5_resize_parameter_rowsrAVs??1%Lxq##%//0PIOOABY[,9##%,,g6yA <<( 8O8O PPmodulec|t|dd}|t|dk(ry|d}t|tjr|j dkryt j5|jddj|j|jddjdddy#1swYyxYw)zFReset keypoint precision-Cholesky output rows to unit Gaussian values.layersNr) getattrlen isinstancerr* out_featuresr,r-r#zero_r")rCrE final_layers r5$_reset_keypoint_gaussian_output_rowsrPds VXt ,F ~V)*K k299 -1I1IQ1N *1Q%%'    '   Qq ! ' ' )***s "AB22B;c eZdZdZ d$deedzdeffd ZdeddfdZedeede jfd Z edeede jfd Z deefd Z ed eee jfdeedzfd ZdeedzddfdZd%dZdZde jdedede jfdZde jde jfdZd&defdZdZe j2j4 d&de jde jde jdzde jdzfdZdeej<fdZd e d!eddfd"Z!d#Z"xZ#S)'LWDETRz@This is the Group DETR v3 module that performs object detection.Nnum_keypoints_per_class grouppose_keypoint_dim_downscalec  t|||_||_|j}t j |||_t||dd|_ ||_ d}t j||z||_ t j||z||_ t jj|jj j"d||_||_||_| |_|j*s&|j|jj,_ nd|jj,_ | |_| |_| xsg|_| |_|j0r@t7|j2|kDr(t9dt7|j2d|d|dd |_|j0rt||j4z||j4zd d|_t jj|j<j>d j j"dt jj|j<j>d j@j"dnd|_|jCd |jE|j2d }tGjHd|z |z  }tKjL||z|jj@_t jj|jj>d j j"dt jj|jj>d j@j"d||_'|jNr t jPtS|Dcgc]!}tUjV|j#c}|j_,t jPtS|Dcgc]!}tUjV|j#c}|j_-|j0rd|j<Xt jPtS|Dcgc]!}tUjV|j<#c}|j_.d |_/ycc}wcc}wcc}w)aInitializes the model. Parameters: backbone: torch module of the backbone to be used. See backbone.py transformer: torch module of the transformer architecture. See transformer.py num_classes: number of object classes num_queries: number of object queries, ie detection slot. This is the maximal number of objects Conditional DETR can detect in a single image. For COCO, we recommend 100 queries. aux_loss: True if auxiliary decoding losses (loss at each decoder layer) are to be used. group_detr: Number of groups to speed detr training. Default is 1. lite_refpoint_refine: TODO rHrNznum_keypoints_per_class has z) entries but the detection head only has z= classes. Class-logit boosts for keypoint classes with id >= zV would be silently truncated. Increase num_classes or shorten num_keypoints_per_class.FrF_kp_active_maskg{Gz?r!)0super__init__ num_queries transformerd_modelrr* class_embedr bbox_embedsegmentation_head Embeddingrefpoint_embed query_featinit constant_r#databackboneaux_loss group_detrlite_refpoint_refinedecoder bbox_reparamuse_grouppose_keypointsrSrTrK ValueError_kp_zero_pad_warnedkeypoint_embedrEr"register_buffer_create_kp_active_maskr&logr,ones two_stage ModuleListrangecopydeepcopyenc_out_bbox_embedenc_out_class_embedenc_out_keypoint_embed_export)selfrgr\r`rr[rhrirurjrlrmrSrT hidden_dim query_dim prior_prob bias_value_ __class__s r5rZzLWDETR.__init__ws8 && (( 99Z=j*a;!2  ll;+CYO,,{Z'?L $--44991=    $%9!((26//D   $ $ /26D   $ $ /((?$'>'D"$0P-  ' 'C0L0L,MP[,[.s43O3O/P.QR'=(efqershh  $)  ' '"%dCCCdCCC #D  GG  d1188<CCHH! L GG  d1188<AAFF J"&D  .0K0KDLhLh0ij hhJ*<== %*ZZ %>B# >>24--9>z9JKAt/K3D   /46==:? :KLQt//0L4D   0++0C0C0O:<--AFzARSAT]]4#6#67S;  7 LMTs&S&&S&S$rrc t|j||_|jrQtj|j j Dcgc]}t||c}|j _yycc}w)aResize the detection classification head to *num_classes* outputs. Replaces ``self.class_embed`` (and each ``enc_out_class_embed`` when the model uses two-stage detection) with a new :class:`torch.nn.Linear` whose ``out_features`` equals *num_classes*. When *num_classes* is larger than the current head the existing weights are tiled; when smaller they are truncated. Replacing the module (rather than mutating ``.data``) keeps ``nn.Linear.out_features`` consistent with the actual weight shape, which is required for correct ONNX export. Args: num_classes: Target number of output classes (including background). N)r6r^rurrvr\r{)r~rms r5reinitialize_detection_headz"LWDETR.reinitialize_detection_heads`*$*:*:KH >>35==9=9I9I9]9]^A;/^4D   0 ^sA:c|s&tjddtjSt|}tjt ||tj}t |D]\}}d||d|f<|S)zECreate a compact class-by-keypoint active mask for a keypoint schema.rdtypeTN)r,zerosboolmaxrK enumerate)rSmax_kp kp_active class_idx num_keypointss r5rrzLWDETR._create_kp_active_masksx';;q!5::6 6,-KK$; ?7=D   4$**It<  w 9:28/!(2F!M ,bll;-CDVXg-h*w =>335"78&:: .2.N.Nv.V.Y.YZfZmZm.n+ ` W !$"2"24DdKKk9d;G'2<<0&\} }| |z}t|D]&}|d d d d | d d f| d d d d ||zd d f<| dz } (@| S)zDConvert compact GroupPose keypoints to class-padded keypoint layout.rHz8_format_keypoint_output received tensor with batch_size=z, num_queries=z but expected batch_size=zW. Shape mismatch silently bypassed in earlier versions; raise to surface upstream bugs.zE_format_keypoint_output received tensor with total_compact_keypoints=z) but schema expects either compact total=z or padded total=z for num_keypoints_per_class=.rrrNr!)rmrSr;r$rnrrKrr,rrrrrw)r~rrr batch_sizer[total_compact_keypoints keypoint_dimmax_num_keypointstotal_padded_keypointstotal_actual_keypointspadded compact_idxrkeypoint_count class_offset keypoint_idxs r5_format_keypoint_outputzLWDETR._format_keypoint_outputZs++43O3O$ $  "a '$ $IZI`I`F K!8, , , ?S0SJ:,Vdepdqr++>*?~NbMcdhh    < <=!$T%A%A!BEV!V!$T%A%A!B "&< <$ $ "&< <WXoWpq;;Q:RS##9"::WX\XtXtWuuvx     " #))$++   )243O3O)P ! %I~$'88L %n 5 ! ?PQRTUWbdeQe?fq!\L8!;<q  ! !  rBkeypoint_predictionsc |jsRtjg|jdd|jj |j |jSt|j}t|j}|djg|jdd||}||jj|j z}|jd}|jj }|dz }|jd|kr|jd|krI|js=tj!d|jd||jd|dz d |_tj"||j$g|jdd||jdz gd}|S|jd|kDr |d d|f}|S) zIAggregate keypoint class-logit contributions into detection-class logits.Nr).rIrFrr!zKeypoint class-logit boost has %d classes but detection head has %d foreground classes; zero-padding boost for classes %d..%d. Detection classes with no keypoint schema will receive zero boost. This warning is emitted once per model instance.T.)rSr,rr$r^rMrrrKrviewrXrrrologgerwarningcatr:)r~rnum_keypoint_classesr class_contrib class_boostdetection_num_classesforeground_num_classess r5 _aggregate_keypoint_class_logitsz'LWDETR._aggregate_keypoint_class_logitss++;;Q&,,Sb1Q43C3C3P3PQ*00+22  #4#?#?@ < <=9,V499 ! ' ' , .B DU &(<(<(?(? @S@S(TT #''B'/ $ 0 0 = =!6!:   R #8 8  $'== //NNd$))"-.#))"-.204D,)))K))q;+<+|%d }||jA|z}|j. |d jB||jBj:d d}(|d|dd })|j(d|)d <||d|)d<|jDr%|jG|||j(nd||)d<|jHra|jr |jJnd }*|jM|*d }+g},t9|*D]5}-|j jN|-|+|-}.|,j|.7tj6|,d },d}/|j"rC|A|j=||j:d |j:d }/|,|jA|/z},|j4 |d jB|g|jBj:d ddd }0|(|,|d )d<|j0|)dd <|/|/|)dd<|)S|,|d })|j0|)d <|/|/|)d<)S)a8The forward expects a NestedTensor, which consists of: - samples.tensor: batched images, of shape [batch_size x 3 x H x W] - samples.mask: a binary mask of shape [batch_size x H x W], containing 1 on padded pixels It returns a dict with the following elements: - "pred_logits": the classification logits (including no-object) for all queries. Shape= [batch_size x num_queries x num_classes] - "pred_boxes": The normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These values are normalized in [0, 1], relative to the size of each individual image (disregarding possible padding). See PostProcess for information on how to retrieve the unnormalized bounding box. - "aux_outputs": Optional, only returned when auxiliary losses are activated. It is a list of dictionaries containing the two above keys for each decoder layer. Ncross_attn_srcsrH.rrFrKuse_grouppose_keypoints=True requires keypoint_hs from transformer outputs.rrr!) pred_logits pred_boxes pred_maskspred_keypoints aux_outputsT skip_blocks enc_outputs)(rLrr,rrrg decomposeappendtrainingrbr#rcr[r`sparse_forwardrr\rmrlr_expconcatsigmoidr^rprn unsqueezerrwr$rstackrtensorsrh _set_aux_lossrurichunkr{)1r~rtargetsfeaturesposscross_attn_featuressrcsmasksfeatsrcrrefpoint_embed_weightquery_feat_weight seg_head_fwdrfeature cross_srcrtransformer_outputshs ref_unsigmoidhs_encref_enc keypoint_hsenc_kp_predictionsoutputs_coord_deltaoutputs_coord_cxcyoutputs_coord_wh outputs_coord outputs_classoutputs_keypointsoutputs_keypoints_deltaref_whref_xy keypoints_xykeypoints_otheroutputs_keypoints_compactlayer_outputs_keypoints layer_idx compact_preds outputs_masksoutri hs_enc_listcls_encg_idx cls_enc_gidx keypoints_enc masks_encs1 r5rzLWDETR.forwardsL gell3 44W=G.2mmG.D+$+ $D(IC KK  LL # ##  $ ==$($7$7$>$> ! $ 6 6 %)$7$7$>$>?QAQAQ$R ! $ 6 67I9I9I J   ! ! -DHMM411@@W[WmWmWuWuL  * O. 2&002 1&&y1 2#..    ! + /   ' 'Uh RB vw =OQR1DRa1H .B vwK!%  >  &*oob&9#%8bqb%AMRUWXWYRYDZ%Z]jknprqrprkr]s%s"#6sABw#?#C#C#E VY[\[]V]H^#^ % .@BR-SY[ \ !%!4}!D M M O  ,,R0M $ ++0C0C0O&$%rss*.*=*=k*J'&sABw/99"=&sBQBw/99"=6sBQBw?&H6Q "9#qr'"B,1II|_6U[],^)*,'!&'@'F'Fq'I!JI$=i$HM+2244))//2)//2%*KK0GQ$O! -0U0UVg0h h %%1 ,Xa[-@-@"gooF[F[\^\_F` a "/"3=QSCTUC%%1$1"$5L! ,(9"(=$%}}%)%7%7!!%)%;%;%GMT% &M" >>,0MMqJ ,,zq,9KGz* -Jt//CCEJ;W\K]^ |, -iiQ/G M++0B0N $ < <&&,,Q/&,,Q/! "D$I$I-$XX%%1(QK''OO))"#. $  ~5  &*oob&9#%8bqb%AMRUWXWYRYDZ%Z]jknprqrprkr]s%s"#6sABw#?#C#C#E VY[\[]V]H^#^ % .@BR-SY[ \ !%!4}!D M M O  ,,R0M++0C0C0O&$%rss*.*=*=k*J'&sABw/99"=&sBQBw/99"=6sBQBw?&H6Q "9#qr'"B$)II|_.MSU$V!$((*a/(9"(=%$($@$@%%++A.%++A.%! !.0U0UVg0h h %%1 $ 6 6GMM"#& !  ! >> Q#Q Q>CD,,@@CFKM#M++0B0N$($@$@&&,,Q/&,,Q/%! !.0U0UVg0h h %%1 $ 6 6GMM"#& $ !7!!   $ -> >  ( -1BB B -/ /rBrrr rc Dddg}|dd|ddg}|%|jd|j|dd|%|jd|j|ddt|D cgc]"}t||D cic]\}} ||  c} }$c} }}Scc} }wcc} }}w)NrrrFrr)rzip) r~rrr rnamesvalues layer_valuesrvalues r5rzLWDETR._set_aux_losss -$mCR&89  $ LL & MM-, -  ( LL) * MM+CR0 1^aci^jkklE<0HIuuIkkIks)B= B BBc|jdj}t|dr |jSt|dr,t|jdr|jjSt|drVt|jdr@t|jjdr |jjj Sy)a_Resolve the list of transformer blocks/layers from backbone[0].encoder. Supports multiple backbone architectures: - encoder.blocks (standard ViT) - encoder.trunk.blocks (aimv2) - encoder.encoder.encoder.layer (HuggingFace DinoV2) Returns: List of transformer layers, or None if not found. rblockstrunkencoderlayerN)rgr3rr1r2r4)r~encs r5_get_backbone_encoder_layersz#LWDETR._get_backbone_encoder_layerssmmA&& 3 !::  3 WSYY%A99## # 3 "ws{{I'F7SVS^S^SfSfhoKp;;&&,, ,rBdrop_path_ratevit_encoder_num_layerscd|j}|yt|t|}tjd||Dcgc]}|j }}t |D]C}t||dst||jds-||||j_ Eycc}w)aUpdate drop_path rates for backbone encoder layers with linear schedule. Applies a linear schedule where the first layer has drop_path_rate=0 and the last layer has drop_path_rate=drop_path_rate. Intermediate layers are interpolated linearly. Args: drop_path_rate: Maximum drop path rate (applied to last layer). vit_encoder_num_layers: Number of encoder layers to update. Nr drop_path drop_prob) r6minrKr,linspaceitemrwrr:r;)r~r7r8rEnxdp_ratesis r5update_drop_pathzLWDETR.update_drop_paths224 >  &F 4&+nnQ&JKAFFHKKq !>56**7H: {""D$.. t##!$"6"67D#D)K  ! !  OO OO!77  $$??..!66&& '.G O '.G L)07Y[\)] E LrBctj|j}t|}|j|jd}|j |d<|j r|j|d<|j|d<t|dd}|r@t|dd|d <t|d d|d <t|d d|d <t|dd|d<|jri}t|jdz D];}|j|jDcic]\}}|d|z|c}}=|jr6|j|jDcic] \}}|dz| c}}|j|gd} |j r| j!d|r| j!dt|dd} |j rlt#|j$dz|||j&| |j(| |j*|j,|j.|j0t|dg } n`t#|j$dz|||j&| |j(| |j*|j,|j.t|dg } | j3|t5|j6t|dgt|dd} | | fScc}}wcc}}w)N)loss_ce loss_bbox loss_giou loss_mask_celoss_mask_dicermFkeypoint_l1_loss_coefgloss_keypoints_l1keypoint_findable_loss_coefloss_keypoints_findablekeypoint_visible_loss_coefloss_keypoints_visiblekeypoint_nll_loss_coefloss_keypoints_nllr!r_enc)labelsboxes cardinalityr keypointssum_group_lossesrS) matcher weight_dict focal_alphalossesriruse_varifocal_lossuse_position_supervised_loss ia_bce_lossmask_point_sample_ratiorS) rrrrrirrrrrSpostprocess_trace_alphag?) num_selectrS trace_alpha)r,rr cls_loss_coefbbox_loss_coefgiou_loss_coefr`mask_ce_loss_coefmask_dice_loss_coefrJrhrwrrupdateitemsrurr rrrirrrrrrr) rtrrr has_keypointsaux_weight_dictrBkvrr criterion postprocesss r5"build_criterion_and_postprocessorsrAs \\$++ &FD!G"00t?R?RSK#22K  &*&<&< N#(,(@(@ $%D";UCM+249PRU+V '(18?\^a1b -.07>Z\_0` ,-,3D:RTW,X () }}t*+ UA  " "{?P?P?R#Stq!A!A3KN#S T U >>  " "k>O>O>Q#RdaAJM#R S?+ /F  g k"t%7?    q #((-#66)-)J)J(($($@$@$+D2KR$P  !   q #((-#66)-)J)J(($+D2KR$P  LL?? '.G LD";SA K k !!c$T#Rs K K% model_configr train_configrdefaultscddlm}|js |jr t d|ddlm}|dd}||||}t|S)uQBuild an LWDETR model directly from a ModelConfig. A config-native alternative to ``build_model(build_namespace(mc, tc))``. Constructs the namespace internally from ``model_config``, an optional ``train_config``, and ``defaults``, then delegates to :func:`build_model`. Note: The internal ``SimpleNamespace`` bridge is transitional — it will be eliminated once all builder functions accept config objects directly. Callers should not rely on the namespace shape or pass it externally. Args: model_config: Architecture configuration. train_config: Training hyperparameter configuration. If ``None``, a minimal dummy ``TrainConfig(dataset_dir=".", output_dir=".")`` is constructed, matching the previous default behavior. defaults: Hardcoded architectural constants. Defaults to ``MODEL_DEFAULTS``. Returns: Fully initialised LWDETR model. Raises: ValueError: If ``defaults`` request ``encoder_only`` or ``backbone_only``, which would make the return type differ from ``LWDETR``. r_namespace_from_configsz`build_model_from_config() requires defaults.encoder_only=False and defaults.backbone_only=False.)rr) dataset_dir output_dir)rfdetr._namespacerrorprn rfdetr.configrrv)rrrrrnss r5build_model_from_configrsW8: 6 6 n  -"ssC |X FB r?rBc8ddlm}||||}t|S)aBuild criterion and postprocessor directly from config objects. A config-native alternative to ``build_criterion_and_postprocessors(build_namespace(mc, tc))``. Args: model_config: Architecture configuration. train_config: Training hyperparameter configuration. defaults: Hardcoded architectural constants. Defaults to ``MODEL_DEFAULTS``. Returns: A 2-tuple of ``(SetCriterion, PostProcess)``. rr)rrr)rrrrrs r5build_criterion_from_configrs ": |X FB -b 11rB)rtr )9rNrxr&typingrrrr,rrrrrfdetr.models._defaultsr r rfdetr.models._typesr rfdetr.models.backboner rfdetr.models.criterionr rrrrrrr rfdetr.models.heads.segmentationrrfdetr.models.matcherrrfdetr.models.mathrrfdetr.models.postprocessrrfdetr.models.transformerrrfdetr.utilities.loggerrrfdetr.utilities.tensorsrrrr*r%r6r<rAModulerPrRrvrrtuplerrmrBr5rsG$+ 44 6A,1   >/"17.Q 2993299< Qbll Qc Qbll Q * *t * | %RYY| %~K\D"R-1,))=))) )^-2222 < $% 2rB