a yfh@s*dZddlZddlZddlZddlmZddlmZmZddl m Z m Z m Z m Z ddlmZmZmZmZddlmZmZddlmZmZmZdd lmZmZd ZGd d d ejZGd ddeZ GdddeZ!GdddeZ"GdddejZ#GdddeZ$GdddejZ%GdddeZ&dS)zModel head modules.N) constant_xavier_uniform_) TORCH_1_10 dist2bbox dist2rbox make_anchors)DFLBNContrastiveHeadContrastiveHeadProto)ConvDWConv)MLPDeformableTransformerDecoder!DeformableTransformerDecoderLayer)bias_init_with_prob linear_init)DetectSegmentPoseClassifyOBB RTDETRDecoder v10DetectcseZdZdZdZdZdZdZdZe dZ e dZ dfdd Z d d Zd d ZddZddZddZede jeedddZZS)rz(YOLOv8 Detect head for detection models.F,NrPcst|_t|_d_|jd_tj_ t d|ddjdft |dt jdt fdd|D_t fdd|D_jdkrtjnt _jrtj_tj_d S) zUInitializes the YOLOv8 detection layer with specified number of classes and channels.rdc 3s<|]4}tt|dtdtdjdVqdS)rrN)nn Sequentialr Conv2dreg_max.0x)c2selfrW/var/www/html/django/DPS/env/lib/python3.9/site-packages/ultralytics/nn/modules/head.py )sz"Detect.__init__..c 3sX|]P}ttt||dt|dttdtdtjdVqdSr!rN)r"r#rr r$ncr&c3r*rr+r,,s rN)super__init__r.lennlr%notorchZzerosstridemaxminr" ModuleListcv2cv3r ZIdentitydflend2endcopydeepcopy one2one_cv2 one2one_cv3r*r.ch __class__)r)r0r*r+r2 s"  4  zDetect.__init__cCsv|jr||St|jD]4}t|j||||j|||fd||<q|jrZ|S| |}|j rn|S||fS)JConcatenates and returns predicted bounding boxes and class probabilities.r) r>forward_end2endranger4r6catr;r<training _inferenceexport)r*r(iyrrr+forward:s 2 zDetect.forwardcsdd|DfddtjD}tjD]4}tj|||j|||fd||<q2jrx||dS|}| dddj j }j r|S|||dfS)a Performs forward pass of the v10Detect module. Args: x (tensor): Input tensor. Returns: (dict, tensor): If not in training mode, returns a dictionary containing the outputs of both one2many and one2one detections. If in training mode, returns a dictionary containing the outputs of one2many and one2one detections separately. cSsg|] }|qSr)detachr'xirrr+ Qz*Detect.forward_end2end..cs8g|]0}tj||j||fdqS)r)r6rJrArBr'rNr*Zx_detachrr+rTRsr)Zone2manyone2oner) rIr4r6rJr;r<rKrL postprocesspermutemax_detr.rM)r*r(rXrNrOrrWr+rHFs  2  zDetect.forward_end2endc s|djtfdd|Dd}js6jkr\ddt|jdD\___jrj dvr|d d d j d f}|d d j d d f}n| j d j fd \}}jrLj d vrLd}d }tj ||||g|jdd d d }jjd|}||jd|d d d df} n |jdj} t| |fd S)z]Decode predicted bounding boxes and class probabilities based on multiple-level feature maps.rcs g|]}|djdqS)r)viewr5rRr*shaperr+rTbrUz%Detect._inference..rYcss|]}|ddVqdSrrNZ transposer&rrr+r,drUz$Detect._inference..?ZtfjstfliteZ saved_modeledgetpuZpbNrrrerfr!device)r`r6rJdynamicrr7anchorsstridesrMformatr%splitr.tensorrireshape decode_bboxesr= unsqueezesigmoid) r*r(x_catboxclsgrid_hgrid_w grid_sizenormdboxrr_r+rL^s"  "4 zDetect._inferencecCs|}t|j|j|jD]L\}}}d|djjdd<td|jd|d|djjd|j<q|j rt|j |j |jD]L\}}}d|djjdd<td|jd|d|djjd|j<q|dS)BInitialize Detect() biases, WARNING: requires stride availability.?r]NirY) zipr;r<r7biasdatamathlogr.r>rArBr*mabsrrr+ bias_initzs0zDetect.bias_initcCst|||j ddS)zDecode bounding boxes.r)Zxywhdim)rr>r*Zbboxesrkrrr+rqszDetect.decode_bboxes)predsr\r.c Cs|j\}}}|jd|gdd\}}|jddt||dd}|jd|dddd}|jd|dd|d}|dt||\}}t |d} t j || ||f|d||d gddS)a4 Post-processes YOLO model predictions. Args: preds (torch.Tensor): Raw predictions with shape (batch_size, num_anchors, 4 + nc) with last dimension format [x, y, w, h, class_probs]. max_det (int): Maximum detections per image. nc (int, optional): Number of classes. Default: 80. Returns: (torch.Tensor): Processed predictions with shape (batch_size, min(max_det, num_anchors), 6) and last dimension format [x, y, w, h, max_class_prob, class_index]. rr]rr)rindex).N) r`rnZamaxtopkr9rrgatherrepeatflattenr6arangerJfloat) rr\r.Z batch_sizerk_ZboxesZscoresrrNrrr+rZs "zDetect.postprocess)rr)r)__name__ __module__ __qualname____doc__rjrMr>r\r`r6emptyrkrlr2rPrHrLrrq staticmethodZTensorintrZ __classcell__rrrEr+rs     rcs*eZdZdZd fdd Zdd ZZS) rz,YOLOv8 Segment head for segmentation models.r rcsdt|||_|_t|djj_t|ddjtfdd|D_ dS)ziInitialize the YOLO model attributes such as the number of masks, prototypes, and the convolution layers.rrc 3s8|]0}tt|dtdtjdVqdSr-)r"r#r r$nmr&Zc4r*rr+r,rUz#Segment.__init__..N) r1r2rnprr protor8r"r:cv4)r*r.rrrDrErr+r2s zSegment.__init__csd}|jdtfddtjDd}tjrX||fSj rrt|gd|fStd|gdd||ffS)zgReturn model outputs and mask coefficients if training, otherwise return outputs and mask coefficients.rcs*g|]"}j||jdqSr])rr^rrVbsr*r(rr+rTrUz#Segment.forward..rYr) rr`r6rJrIr4rrPrKrM)r*r(pZmcrrr+rPs $  zSegment.forward)rrrrrrrrr2rPrrrrEr+rs rcs2eZdZdZd fdd ZddZd d ZZS) rz=YOLOv8 OBB detection head for detection with rotation models.rrrcsHt|||_t|ddjtfdd|D_dS)zCInitialize OBB with number of classes `nc` and layer channels `ch`.rrc 3s8|]0}tt|dtdtjdVqdSr-)r"r#r r$ner&rrr+r,rUzOBB.__init__..N)r1r2rr8r"r:r)r*r.rrDrErr+r2sz OBB.__init__csdjdtfddtjDd}|dtj}jsP|_ t jrj|fSj rt|gdStd|gdd|ffS)rGrcs*g|]"}j||jdqSr)rr^rrVrrr+rTrUzOBB.forward..rYg?r) r`r6rJrIr4rsrpirKanglerrPrM)r*r(rrrr+rPs$ z OBB.forwardcCst||j|ddS)zDecode rotated bounding boxes.rr)rrrrrr+rqszOBB.decode_bboxes)rrr)rrrrr2rPrqrrrrEr+rsrcs2eZdZdZd fdd ZddZd d ZZS) rz&YOLOv8 Pose head for keypoints models.rr!rcsZt|||_|d|d_t|ddjtfdd|D_dS)zIInitialize YOLO network with default parameters and Convolutional Layers.rrrc 3s8|]0}tt|dtdtjdVqdSr-)r"r#r r$nkr&rrr+r,rUz Pose.__init__..N)r1r2 kpt_shaperr8r"r:r)r*r.rrDrErr+r2s z Pose.__init__csdjdtfddtjDd}tjrL|fS|}j rnt|gdStd|gdd|ffS)z?Perform forward pass through YOLO model and return predictions.rcs*g|]"}j||jdqSr)rr^rrVrrr+rTrUz Pose.forward..r]r) r`r6rJrIr4rrPrK kpts_decoderM)r*r(ZkptZpred_kptrrr+rPs$  z Pose.forwardcCs\|jd}|jr|j|g|jdR}|ddddddfd|jd|j}|dkrt||ddddddffd}|||jdS| }|dkr|dddddf|dddddf<|dddd|fd|jdd|j|dddd|f<|dddd|fd|jdd|j|dddd|f<|SdS) zDecodes keypoints.rr]NrY@rcr!r) rrMr^rkrlr6rJrsrclone)r*rZkptsndimrOrrrr+rs .*,@@zPose.kpts_decode)rrr)rrrrr2rPrrrrrEr+rs  rcs*eZdZdZdfdd ZddZZS) rz:YOLOv8 classification head, i.e. x(b,c1,20,20) to x(b,c2).rNcsPtd}t|||||||_td|_tjddd|_t |||_ dS)zcInitializes YOLOv8 classification head to transform input tensor from (b,c1,20,20) to (b,c2) shape.irT)rZinplaceN) r1r2r convr"ZAdaptiveAvgPool2dpoolZDropoutdropLinearlinear)r*c1r)krrgZc_rErr+r2s   zClassify.__init__c CsLt|trt|d}|||||d}|j rB|S| dS)z>Performs a forward pass of the YOLO model on input image data.r) isinstancelistr6rJrrrrrrKZsoftmax)r*r(rrr+rPs  "zClassify.forward)rrNrrrrrEr+rs rcs2eZdZdZd fdd Zdd Zd d ZZS) WorldDetectz^Head for integrating YOLOv8 detection models with semantic understanding from text embeddings.rFrcs`t||t|dt|jdtfdd|D|_tfdd|D|_dS)zHInitialize YOLOv8 detection layer with nc classes and layer channels ch.rr c 3s6|].}tt|dtdtdVqdSr-)r"r#r r$r&)r0embedrr+r,rUz'WorldDetect.__init__..c3s |]}rtntVqdS)N)r r r'r)rwith_bnrr+r,rUN) r1r2r8r9r.r"r:r<r)r*r.rrrDrE)r0rrr+r2szWorldDetect.__init__c stjD]@}tj|||j|j||||fd||<q jrV|S|djtfdd|Dd}j sjkrddt |j dD\_ _ _jrjd vr|d d d jd f}|d d jd d f}n|jd jfd\}}jrjd vrd}d }tj||||g|jddd d} j j d| } || j d| d d d df} n |j dj } t| |fd} jr| S| |fS)rGrrcs*g|]"}|djjddqS)rrr])r^r.r%rRr_rr+rT*rUz'WorldDetect.forward..rYcss|]}|ddVqdSrarbr&rrr+r,,rUz&WorldDetect.forward..rcrdNrrgr!rh)rIr4r6rJr;rr<rKr`rjrr7rkrlrMrmr%rnr.rorirprqr=rrrs) r*r(textrNrtrurvrwrxryrzr{rOrr_r+rP!s,>  "4 zWorldDetect.forwardcCs:|}t|j|j|jD]\}}}d|djjdd<qdS)r|r}r]N)rr;r<r7rrrrrr+rCszWorldDetect.bias_init)rrFr)rrrrr2rPrrrrrEr+rs"rcseZdZdZdZddddddd d d ed d dddffdd Zd ddZde j ddfddZ ddZ d!ddZ ddZZS)"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. Fr)rirrrrrr]r rcr}csFt|_||_t||_|_||_||_t fdd|D|_ t ||| | |j|}t ||| |_t |_| |_| |_||_||_|rt ||_tdddd|_t t t |_t |_tddd|_t fddt|D|_t fd dt|D|_ |!d S) 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 3s.|]&}ttj|dddtVqdS)rF)rN)r"r#r$Z BatchNorm2dr&hdrr+r,rUz)RTDETRDecoder.__init__..rrYZ num_layersr!csg|]}tqSr)r"rrrr.rr+rTrUz*RTDETRDecoder.__init__..csg|]}tdddqS)rr!r)rrrrr+rTrUN)"r1r2Z hidden_dimZnheadr3r4r. num_queriesZnum_decoder_layersr"r: input_projrrdecoderZ Embeddingdenoising_class_embed num_denoisinglabel_noise_ratiobox_noise_scalelearnt_init_query tgt_embedrquery_pos_headr#rZ LayerNorm enc_outputenc_score_head enc_bbox_headrIdec_score_head dec_bbox_head_reset_parameters)r*r.rDrZnqZndpZnhZndlZd_ffnZdropoutZactZeval_idxndrrrZ decoder_layerrErr+r2Xs0'   zRTDETRDecoder.__init__Nc Csddlm}||\}}|||j|j|jj|j|j|j |j \}}}} | ||||\} } } } |j | | |||j |j|j|d\}}||| | | f}|j r|St|d|dfd}|jr|S||fS)zdRuns the forward pass of the module, returning bounding box and classification scores for the input.r) get_cdn_group) attn_maskr])Zultralytics.models.utils.opsr_get_encoder_inputr.rrweightrrrrK_get_decoder_inputrrrrr6rJZsqueezersrM)r*r(batchrfeatsshapesdn_embeddn_bboxrZdn_metar refer_bbox enc_bboxes enc_scoresZ dec_bboxesZ dec_scoresrOrrr+rPs8    zRTDETRDecoder.forwardg?cpu{Gz?cCs*g}t|D]\}\}} tj|||d} tj| ||d} trLtj| | ddn t| | \} } t| | gd}tj| |g||d}|dd|}tj|||d|d|}| t ||gd d|| d q t |d }||k|d |k@j dd d }t |d |}||td }||fS)z\Generates anchor bounding boxes for given shapes with specific grid size and validates them.)enddtyperiZij)Zindexingr]rrirrcrrrT)Zkeepdiminf) enumerater6rrZmeshgridstackrorrZ ones_likeappendrJr^allrZ masked_fillr)r*rryrriZepsrkrNhwZsysxZgrid_yZgrid_xZgrid_xyZvalid_WHZwh valid_maskrrr+_generate_anchorss$& zRTDETRDecoder._generate_anchorscsvfddt|D}g}g}|D]>}|jdd\}}||dddd|||gq"t|d}||fS)zfProcesses and returns encoder inputs by getting projection features from input and concatenating them.csg|]\}}j||qSr)r)r'rNfeatr*rr+rTrUz4RTDETRDecoder._get_encoder_input..rYNrr)rr`rrr[r6rJ)r*r(rrrrrrrr+rs z RTDETRDecoder._get_encoder_inputcCsf|jd}|j||j|jd\}}|||}||} tj| dj |j ddj d} tj || jddd|j  d} || | f ||j d} |dd| f ||j d} || | }|}|durt||gd}| | | f ||j d}|jr|jjd|ddn| }|jr@|}|js@|}|durZt||gd}||||fS)z`Generates and prepares the input required for the decoder from the provided features and shapes.rrr]rr)rrN)r`rrrirrr6rr8valuesrindicesr^rrrrrrsrJrrrrKrQ)r*rrrrrrkrfeaturesZenc_outputs_scoresZtopk_indZ batch_indZtop_k_featuresZ top_k_anchorsrrrZ embeddingsrrr+rs*  "&$ z RTDETRDecoder._get_decoder_inputcCstdd|j}t|jj|t|jjdjdt|jjdjdt|j |j D]8\}}t|j|t|jdjdt|jdjdqVt |j dt |j dj|jrt |jjt |jjdjt |jjdj|jD]}t |djqdS)zjInitializes or resets the parameters of the model's various components with predefined weights and biases.rrr]rrrN)rr.rrrrZlayersrrrrrrrrrrr)r*Zbias_clsZcls_Zreg_layerrrr+rs    zRTDETRDecoder._reset_parameters)N)NN)rrrrrMr"ZReLUr2rPr6Zfloat32rrrrrrrrEr+rMs.O ' 'rcs&eZdZdZdZdfdd ZZS)ra v10 Detection head from https://arxiv.org/pdf/2405.14458. Args: nc (int): Number of classes. ch (tuple): Tuple of channel sizes. Attributes: max_det (int): Maximum number of detections. Methods: __init__(self, nc=80, ch=()): Initializes the v10Detect object. forward(self, x): Performs forward pass of the v10Detect module. bias_init(self): Initializes biases of the Detect module. TrrcsRt||t|dtjdtfdd|D_t j_ dS)zYInitializes the v10Detect object with the specified number of classes and input channels.rr c 3s`|]X}ttt||d|dt|dttddtdtjdVqdS)r!)rrN)r"r#r r$r.r&r/rr+r,Ms z%v10Detect.__init__..N) r1r2r8r9r.r"r:r<r?r@rBrCrEr/r+r2Hs  zv10Detect.__init__)rr)rrrrr>r2rrrrEr+r4sr)'rr?rr6Ztorch.nnr"Z torch.nn.initrrZultralytics.utils.talrrrrblockr r r r rr rZ transformerrrrutilsrr__all__Modulerrrrrrrrrrrr+s* (6h