FPhG6dZddlZddlZddlmZddlmZmZddlm Z m Z m Z ddl m Z mZddlmZddlmZmZmZdd lmZmZd ZGd d ej2ZGd deZGddeZGddej2ZGddej2Zy)zModel head modules.N) constant_xavier_uniform_) TORCH_1_10 dist2bbox make_anchors)DFLProto)Conv)MLPDeformableTransformerDecoder!DeformableTransformerDecoderLayer)bias_init_with_prob linear_init_)DetectSegmentPoseClassify RTDETRDecoderceZdZdZdZdZdZejdZ ejdZ dfd Z dZ dZ xZS) rz(YOLOv8 Detect head for detection models.FNrczt|_t|_d_|j dzz_tjj_ td|ddzj dzft|dtjdctjfd|D_tjfd|D_j dkDrt!j _ytj"_y) zUInitializes the YOLOv8 detection layer with specified number of classes and channels.rdc 3K|]S}tjt|dtdtjdjzdUyw)rrN)nn Sequentialr Conv2dreg_max).0xc2selfs fC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\ultralytics/nn/modules/head.py z"Detect.__init__..%sL!lhjcdBMM$q"a.$r2q/299RT\\IY[\;] ^hjsAAc 3K|]P}tjt|dtdtjjdRywrrN)rrr rnc)r!r"c3r$s r%r&z"Detect.__init__..'K wtvoptAr1~tBAPRPYPYZ\^b^e^eghPi!jtvAArN)super__init__r)lennlr notorchzerosstridemaxminr ModuleListcv2cv3r Identitydfl)r$r)chr#r* __class__s` @@r%r.zDetect.__init__s b' t||a''kk$''* b"Q%1*dllQ&678#beSRUEV:WB==!lhj!ll== wtv ww(, q(83t||$bkkmc |dj}t|jD]I}tj|j ||||j |||fd||<K|jr|S|js|j|k7r2dt||jdD\|_ |_ ||_tj|Dcgc]"}|j|d|jd$c}d}|jr?|j dvr1|ddd|j"d zf}|dd|j"d zdf}n.|j%|j"d z|j&fd\}}t)|j+||jj-dd d |jz}|jrs|j d vre|d|jdz} |d |jdz} tj.| | | | g|j0j3dd d} || z}tj||j5fd} |jr| S| |fScc}w)zJConcatenates and returns predicted bounding boxes and class probabilities.rrc3@K|]}|jddyw)rrN) transpose)r!r"s r%r&z!Detect.forward..2s)gEf!++a*;Efs?) saved_modelpbtfliteedgetputfjsNrT)xywhdim)rGrHr)device)shaperanger0r2catr8r9trainingdynamicrr4anchorsstridesviewr1exportformatr splitr)rr; unsqueezetensorrLreshapesigmoid) r$r"rMixix_catboxclsdboximg_himg_wimg_sizeys r%forwardzDetect.forward*s4! twwA99kdhhqk!A$/!QqT1BCQGAaD ==H \\TZZ50)g\RSUYU`U`beEf)g &DL$,DJ AFAb27758TWWb9AFJ ;;4;;*\\,DLL1,,,-C4<BHC# (>(>q(ARSTW[WcWcc ;;4;;*??!Ht{{1~-E!Ht{{1~-E||UE5%$@U]]^_abdefH H D IItS[[]+Q /KKq+aV+%Gs'JcF|}t|j|j|jD]q\}}}d|djj ddt jd|jz d|z dzz |djj d|jsy)zBInitialize Detect() biases, WARNING: requires stride availability.?rCNirD) zipr8r9r4biasdatamathlogr))r$mabss r% bias_initzDetect.bias_initIs~ 155!%%2GAq!!$AbEJJOOA %)XXa!$$h#'a.G%HAbEJJOOEQTT "3r>)P)__name__ __module__ __qualname____doc__rQrUrMr2emptyrRrSr.rfrs __classcell__r=s@r%rrsD2G F Eekk!nGekk!nG L,>Ir>rc*eZdZdZdfd ZdZxZS)rz,YOLOv8 Segment head for segmentation models.cBt|||_|_t |djj_t j_t|ddzjtjfd|D_ y)ziInitialize the YOLO model attributes such as the number of masks, prototypes, and the convolution layers.rrc 3K|]P}tjt|dtdtjjdRywr()rrr rnmr!r"c4r$s r%r&z#Segment.__init__.._r+r,N) r-r.rnprr protorrfdetectr5rr7cv4)r$r)rrr<rr=s` @r%r.zSegment.__init__Vsv R 2a5$((DGG4 nn A!TWW %== wtv wwr>c |j|d}|jd}tjt |j Dcgc]5}|j |||j||jd7c}d}|j||}|jr|||fS|jrtj||gd|fStj|d|gd|d||ffScc}w)zgReturn model outputs and mask coefficients if training, otherwise return outputs and mask coefficients.rrCrDr) rrMr2rOrNr0rrTrrrPrU)r$r"pbsr\mcs r%rfzSegment.forwardas JJqt  WWQZ YYtwwXA  AaD)..r477B?XZ[ \ KKa  ==b!8O-1[[ 1b'1%q)guyy!A$PRUV?WZ[\]Z^`bdeYf>gg Ys :C8)rt rurvrwrxryr.rfr{r|s@r%rrSs6 x hr>rc0eZdZdZdfd ZdZdZxZS)rz&YOLOv8 Pose head for keypoints models.ct|||_|d|dz_tj _t|ddzjtjfd|D_ y)zIInitialize YOLO network with default parameters and Convolutional Layers.rrrc 3K|]P}tjt|dtdtjjdRywr()rrr rnkrs r%r&z Pose.__init__..xr+r,N) r-r. kpt_shaperrrfrr5rr7r)r$r)rr<rr=s` @r%r.z Pose.__init__psg R "A,1-nn A!TWW %== wtv wwr>c |djd}tjt|jDcgc]5}|j |||j ||jd7c}d}|j||}|jr||fS|j||}|jrtj||gdStj|d|gd|d|ffScc}w)z?Perform forward pass through YOLO model and return predictions.rrCr) rMr2rOrNr0rrTrrrP kpts_decoderU)r$r"rr\kptpred_kpts r%rfz Pose.forwardzs qTZZ]ii%PTPWPW.Y.Q!QqT*//DGGR@.Y[]^ KKa  ==c6M##B,.2kkuyy!X*l 1Q4QYJZ\]@^abcdaegj`k?ll Zs:C5c|jd}|jr|j|g|jd}|ddddddfdz|jdz z|jz}|dk(r2t j ||ddddddfjfd}|j||jdS|j}|dk(r|dddddfj|dddd|fdz|jddz z|jz|dddd|f<|dddd|fdz|jddz z|jz|dddd|f<|S) zDecodes keypoints.rrCNrD@rBrr) rrUrTrRrSr2rOr[rclonesigmoid_)r$rkptsndimrerps r%rzPose.kpts_decodesa~~a  ;; "2t~~2r2A1a!8s"dllS&89T\\IAqyIIq!Aq!A#I,"6"6"891=66"dggr* * Aqy!QTT' ##%q!'T'z]S0DLLOc4IJdllZAaDjMq!'T'z]S0DLLOc4IJdllZAaDjMHr>)rt)rru)rvrwrxryr.rfrr{r|s@r%rrms0xmr>rc*eZdZdZdfd ZdZxZS)rz:YOLOv8 classification head, i.e. x(b,c1,20,20) to x(b,c2).ct|d}t|||||||_t j d|_t jdd|_t j|||_ y)zInitializes YOLOv8 classification head with specified input and output channels, kernel size, stride, padding, and groups. irT)rinplaceN) r-r.r convrAdaptiveAvgPool2dpoolDropoutdropLinearlinear) r$c1r#krrrgc_r=s r%r.zClassify.__init__sa  RAq!, ((+ JJd3 iiB' r>c &t|trtj|d}|j |j |j |j|jd}|jr|S|jdS)z>Performs a forward pass of the YOLO model on input image data.r) isinstancelistr2rOrrrrflattenrPsoftmax)r$r"s r%rfzClassify.forwardsg a  !QA KK $))DIIaL"9"A"A!"DE FMMq3qyy|3r>)rrNrrr|s@r%rrsD (4r>rceZdZdZdZddddddd d d ej d d dddffd ZddZde jddfdZ dZ ddZ dZxZS)ra Real-Time Deformable Transformer Decoder (RTDETRDecoder) module for object detection. This decoder module utilizes Transformer architecture along with deformable convolutions to predict bounding boxes and class labels for objects in an image. It integrates features from multiple layers and runs through a series of Transformer decoder layers to output the final predictions. Frt)iiri,rrrrCrrBrhc t||_||_t ||_||_||_||_tjfd|D|_ t||| | |j |}t||| |_tj||_| |_| |_||_||_|rtj||_t-ddzd|_tj0tj2tj4|_tj2||_t-dd|_tjt=|Dcgc]}tj2|c}|_tjt=|Dcgc]}t-ddc}|_ |jCycc}wcc}w)a| Initializes the RTDETRDecoder module with the given parameters. Args: nc (int): Number of classes. Default is 80. ch (tuple): Channels in the backbone feature maps. Default is (512, 1024, 2048). hd (int): Dimension of hidden layers. Default is 256. nq (int): Number of query points. Default is 300. ndp (int): Number of decoder points. Default is 4. nh (int): Number of heads in multi-head attention. Default is 8. ndl (int): Number of decoder layers. Default is 6. d_ffn (int): Dimension of the feed-forward networks. Default is 1024. dropout (float): Dropout rate. Default is 0. act (nn.Module): Activation function. Default is nn.ReLU. eval_idx (int): Evaluation index. Default is -1. nd (int): Number of denoising. Default is 100. label_noise_ratio (float): Label noise ratio. Default is 0.5. box_noise_scale (float): Box noise scale. Default is 1.0. learnt_init_query (bool): Whether to learn initial query embeddings. Default is False. c 3K|]D}tjtj|ddtjFyw)rF)rkN)rrr BatchNorm2d)r!r"hds r%r&z)RTDETRDecoder.__init__..s<'wtvop bii2qu6UWYWeWefhWi(jtvsA A rrD) num_layersrN)"r-r. hidden_dimnheadr/r0r) num_queriesnum_decoder_layersrr7 input_projrr decoder Embeddingdenoising_class_embed num_denoisinglabel_noise_ratiobox_noise_scalelearnt_init_query tgt_embedr query_pos_headrr LayerNorm enc_outputenc_score_head enc_bbox_headrNdec_score_head dec_bbox_head_reset_parameters)r$r)r<rnqndpnhndld_ffndropoutacteval_idxndrrr decoder_layer_r=s ` r%r.zRTDETRDecoder.__init__sL  b'"%--'wtv'ww :"b%RUW[W^W^`cd 3B sHU &(\\"b%9"!2."3 \\"b1DN!!QVRA>-- "b(92<<;KL iiB/ Rq9!mmc ,S 1RYYr2-> ,ST]]RWX[R\+]R\QCBa,HR\+]^  -T+]s HH c zddlm}|j|\}}|||j|j|j j |j|j|j|j\}}}} |j||||\} } } } |j| | |||j|j|j|\}}||| | | f}|jr|St!j"|j%d|j%dj'fd}|j(r|S||fS)zdRuns the forward pass of the module, returning bounding box and classification scores for the input.r) get_cdn_group) attn_maskrC)ultralytics.models.utils.opsr_get_encoder_inputr)rrweightrrrrP_get_decoder_inputrrrrr2rOsqueezer[rU)r$r"batchrfeatsshapesdn_embeddn_bboxrdn_metaembed refer_bbox enc_bboxes enc_scores dec_bboxes dec_scoresres r%rfzRTDETRDecoder.forwardsD>//2 v %''**44;;,,00..-- ) .'9g  # #E68W E 2z:z"&e.8.3.4.2.@.@.2.A.A.2.A.A8A"."C J  J G C ==H IIz))!,j.@.@.C.K.K.MNPR SKKq+aV+r>g?cpu{Gz?c.g}t|D]\}\}} tj|||} tj| ||} trtj| | dntj| | \} } tj | | gd}tj || g||}|jddz|z }tj||||zd|zz}|jtj||gdjd|| zd tj|d }||kD|d |z kzjdd }tj|d |z z }|j|td }||fS)z\Generates anchor bounding boxes for given shapes with specific grid size and validates them.)enddtyperLij)indexingrCrrLrrBrrrT)keepdiminf) enumerater2arangermeshgridstackrYrX ones_likeappendrOrTallrn masked_fillfloat)r$r grid_sizerrLepsrRr\hwsysxgrid_ygrid_xgrid_xyvalid_WHwh valid_masks r%_generate_anchorszRTDETRDecoder._generate_anchors)su"6*IAv1!5@B!5@BFPU^^BTBV[VdVdegikVlNFFkk66"2B7G||QF%GH((+c1X=GfE QUX\]U]^B NN599gr]B7<cdt|Dcgc]\}}|j||}}}g}g}|D]X}|jdd\}}|j|j dj ddd|j||gZt j|d}||fScc}}w)zfProcesses and returns encoder inputs by getting projection features from input and concatenating them.rDNrr)rrrMrrpermuter2rO)r$r"r\featrrrrs r%rz RTDETRDecoder._get_encoder_input=s6?q\ B\'!T T__Q  %\ BD::ab>DAq LLa00Aq9 : MM1a& !  %#f} CsB,ct|}|j||j|j\}}|j ||z}|j |} t j| jdj|jdjjd} t j|| jjdjd|jjd} || | fj||jd} |dd| fj||jd} |j!| | z}|j#}|t j$||gd}| | | fj||jd}|j&r6|j(j*jdj|ddn| }|j,r,|j/}|j&s|j/}|t j$||gd}||||fS)z`Generates and prepares the input required for the decoder from the provided features and shapes.rrCr)rK)rrNr)r/rrrLrrr2topkr5valuesrindicesrTrrXrepeatrr[rOrrrrPdetach)r$rrrrrrRr featuresenc_outputs_scorestopk_ind batch_indtop_k_features top_k_anchorsrrr embeddingss r%rz RTDETRDecoder._get_decoder_inputOs Z"44V5;;W\WcWc4d??:#56!00:::044R8??AQAQWXYaaffgijLLRx~~>HHLSSTUW[WgWghmmnpq ")X"56;;B@P@PRTU8 ,11"d6F6FK ''7-G '')  GZ#8!>r1aHhv ==#**,J))'..0  Hj#91=J:z:==r>ctddz |jz}t|jj|t|j j djdt|j j djdt|j|jD]a\}}t|j|t|j djdt|j djdct|jdt|jdj|jrt|jjt|j j djt|j j dj|j"D]}t|djy)zjInitializes or resets the parameters of the model's various components with predefined weights and biases.rrtrCrrrN)rr)rrrkrlayersrrjrrrrrrrrr)r$bias_clscls_reg_layers r%rzRTDETRDecoder._reset_parametersvsw't,r1DGG; $%%**H5$$$++B/66;$$$++B/44b9d1143E3EFJD$ dii * dkk"o,,b 1 dkk"o**B / G T__Q'(*112  ! ! DNN11 2++2215<<=++2215<<=__E E!HOO ,%r>)N)NN)rvrwrxryrUrReLUr.rfr2float32rrrrr{r|s@r%rrsyF  !##L!\#,J37emmTY_c#($$>N-r>r)ryrmr2torch.nnr torch.nn.initrrultralytics.utils.talrrrblockr r rr transformerr r rutilsrr__all__Modulerrrrrrur>r%r/s~ 4EE]]4 B<IRYY<I~hfh4&6&R4ryy4,`-BII`-r>