h~#zSSKrSSKJr SSKJr SSKJrJr SSKJ r SSK J r SSK J r SS KJr "S S \5rg) N)Image)SegmentationPredictor) DEFAULT_CFGchecks)box_iou) scale_masks) TORCH_1_10)adjust_bboxes_to_image_bordercX^\rSrSrSr\SS4U4SjjrU4SjrS SjrSr Sr S r U=r $) FastSAMPredictora FastSAMPredictor is specialized for fast SAM (Segment Anything Model) segmentation prediction tasks. This class extends the SegmentationPredictor, customizing the prediction pipeline specifically for fast SAM. It adjusts post-processing steps to incorporate mask prediction and non-maximum suppression while optimizing for single-class segmentation. Attributes: prompts (dict): Dictionary containing prompt information for segmentation (bboxes, points, labels, texts). device (torch.device): Device on which model and tensors are processed. clip_model (Any, optional): CLIP model for text-based prompting, loaded on demand. clip_preprocess (Any, optional): CLIP preprocessing function for images, loaded on demand. Methods: postprocess: Apply postprocessing to FastSAM predictions and handle prompts. prompt: Perform image segmentation inference based on various prompt types. set_prompts: Set prompts to be used during inference. Nc4>[TU]XU5 0Ulg)a Initialize the FastSAMPredictor with configuration and callbacks. This initializes a predictor specialized for Fast SAM (Segment Anything Model) segmentation tasks. The predictor extends SegmentationPredictor with custom post-processing for mask prediction and non-maximum suppression optimized for single-class segmentation. Args: cfg (dict): Configuration for the predictor. overrides (dict, optional): Configuration overrides. _callbacks (list, optional): List of callback functions. N)super__init__prompts)selfcfg overrides _callbacks __class__s \/home/james-whalen/.local/lib/python3.13/site-packages/ultralytics/models/fastsam/predict.pyrFastSAMPredictor.__init__#s 4 c>URRSS5nURRSS5nURRSS5nURRSS5n[T U] XU5nUHn [R "SSU R SU R S/USR[RS9n [U RRU R 5n [R"[U SU 5S :5R5n U R5S:wdMXRRU 'M UR!XXVUS 9$) a] Apply postprocessing to FastSAM predictions and handle prompts. Args: preds (list[torch.Tensor]): Raw predictions from the model. img (torch.Tensor): Input image tensor that was fed to the model. orig_imgs (list[np.ndarray]): Original images before preprocessing. Returns: (list[Results]): Processed results with prompts applied. bboxesNpointslabelstextsrr )devicedtypeg?)rrrr)rpopr postprocesstorchtensor orig_shaper float32r boxesxyxynonzerorflattennumelprompt)rpredsimg orig_imgsrrrrresultsresultfull_boxr(idxrs rr#FastSAMPredictor.postprocess3s9!!(D1!!(D1!!(D1   $/'%e)<F||Av((+V->->q-AB5QR8??bgboboH2&,,2C2CVEVEVWE-- > DEMMOCyy{a)1 !!#&{{7&W\{]]rc  UcUcUcU$/n[U[5(dU/nUGHn[U5S:XaURU5 M&URR nUR SSUR:wa[USUR5Sn[R"[U5[RURS9n Ub[R"U[RURS9nURS:XaUSOUnUSS2S4USS2S4- USS2S4USS2S4- -n [R "UV s/sH)oSS2U SU S2U SU S24R#SS9PM+ sn 5n [R""USS9n U SS2S4U -U - nS U [R$"X- SS9'UGb[R"U[RURS9nURS:XaUSOUnUc#[R&"UR S5n[R"U[RURS9n[U5[U5:XdS [U5S [U535eUR#5S:Xa7[R&"[U5[RURS9O6[R"[U5[RURS9n[)X45H8unn[U5U[R*"USS2USUS4S S 9S'M: X-n UGbp[U[,5(aU/n//nn[/UR0R2R555Hunn S U 5unnnn[6(aUUR#5O!UUR#S5R#5S::aURU5 MkUR[8R:"UR<UU2UU2SSS2455 M UR?UU5n[R$"USS9n[U5(a@U[R@"UURS9S[CU5:*R#S5- nS U U'URXy5 GM U$s sn f)a Perform image segmentation inference based on cues like bounding boxes, points, and text prompts. Args: results (Results | list[Results]): Original inference results from FastSAM models without any prompts. bboxes (np.ndarray | list, optional): Bounding boxes with shape (N, 4), in XYXY format. points (np.ndarray | list, optional): Points indicating object locations with shape (N, 2), in pixels. labels (np.ndarray | list, optional): Labels for point prompts, shape (N, ). 1 = foreground, 0 = background. texts (str | list[str], optional): Textual prompts, a list containing string objects. Returns: (list[Results]): Output results filtered and determined by the provided prompts. Nrr )r!r )r r8)dimTz5Expected `labels` with same size as `point`, but got z and )as_tuplec38# UHn[U5v M g7f)N)int).0xs r *FastSAMPredictor.prompt..s%8ac!ffasdr )" isinstancelistlenappendmasksdatashaper&rr$zerosboolr as_tensorint32ndimstacksumargmaxoneszipr*str enumerater(r)tolistr r fromarrayorig_img_clip_inferencer%r<)rr1rrrrprompt_resultsr2rHr4 bbox_areasb mask_areasfull_mask_areasunion point_idxpointlabelcrop_ims filter_idxix1y1x2y2 similaritytext_idxs rr-FastSAMPredictor.promptOsY >fnN'4((iGF6{a%%f-LL%%E{{12&"3"33#E$K1B1BCAF++c&kDKKPC!u{{4;;W)/)9v$QTlVAqD\9fQTlVTUWXTX\>YZ "[[fl)mflab1Q4!A$;!qt 0K*L*P*PU[*P*\fl)mn "'))Ev">"1d7+o= J?CELL!3;<!u{{4;;W)/)9v>"ZZ Q8Fu{{4;;W6{c&k1KCPVK=X]^abh^i]jk1 zz|q(JJs6{%**T[[QS[ 4;;W %($7LE5_cdi_jIemmE!U1XuQx2G,HSWXYZ[\%8  eS))"GE')2*%fll&7&7&>&>&@ADAq%8a%8NBB*4*a %(,,q/:M:M:OTWW"))!, OOEOOFOOBrE2b5RVTVRVDV4W$XY B "11(EB  << ;z??j!Md!SWZ[cWd!d i ijk llH $H  ! !&+ .gjQ*ns0S; cSSKn[ US5(a[ US5(d'UR SUR S9uUlUl[R"UVs/sH,o@RU5RUR 5PM. sn5nURU5RUR 5nURRU5nURRU5nXfRSS S 9-nXwRSS S 9-nXgSS2S4-R!S5$![a [R"S5 SSKnGN\f=fs snf) am Perform CLIP inference to calculate similarity between images and text prompts. Args: images (list[PIL.Image]): List of source images, each should be PIL.Image with RGB channel order. texts (list[str]): List of prompt texts, each should be a string object. Returns: (torch.Tensor): Similarity matrix between given images and texts with shape (M, N). rNz+git+https://github.com/ultralytics/CLIP.git clip_modelclip_preprocesszViT-B/32rCrBT)r9keepdim)clip ImportErrorrcheck_requirementshasattrloadr rorpr$rPtotokenize encode_image encode_textnormrQ)rimagesrrrimagetokenized_textimage_features text_featuress rrZ FastSAMPredictor._clip_inferences;  l++WTCT5U5U48IIjQUQ\Q\I4] 1DOT1W]^W]e2259<rs. A1--40f,fr