hYSSKJr SSKrSSKJs Jr SSKJr SSKJr SSK J r SSK J r SSK Jr SS KJrJr SS KJrJr SS KJrJr S r"S S\R25r"SS\RR25rg)) annotationsN)nn) trunc_normal_)MLP)LOGGER)SAM2TwoWayTransformer) MaskDecoderSAM2MaskDecoder)ImageEncoderViT PromptEncoder)get_1d_sine_peselect_closest_cond_framesgcd^\rSrSr%SrSrS\S'S S U4SjjjrSrSr U=r $) SAMModelaw Segment Anything Model (SAM) for object segmentation tasks. This class combines image encoders, prompt encoders, and mask decoders to predict object masks from images and input prompts. Attributes: mask_threshold (float): Threshold value for mask prediction. image_encoder (ImageEncoderViT): Backbone for encoding images into embeddings. prompt_encoder (PromptEncoder): Encoder for various types of input prompts. mask_decoder (MaskDecoder): Predicts object masks from image and prompt embeddings. pixel_mean (torch.Tensor): Mean values for normalizing pixels in the input image. pixel_std (torch.Tensor): Standard deviation values for normalizing pixels in the input image. Methods: set_imgsz: Set image size to make model compatible with different image sizes. Examples: >>> image_encoder = ImageEncoderViT(...) >>> prompt_encoder = PromptEncoder(...) >>> mask_decoder = MaskDecoder(...) >>> sam_model = SAMModel(image_encoder, prompt_encoder, mask_decoder) >>> # Further usage depends on SAMPredictor class Notes: All forward() operations are implemented in the SAMPredictor class. floatmask_thresholdc&>[TU]5 XlX lX0lUR S[ R"U5RSSS5S5 UR S[ R"U5RSSS5S5 g)aj Initialize the SAMModel class to predict object masks from an image and input prompts. Args: image_encoder (ImageEncoderViT): The backbone used to encode the image into image embeddings. prompt_encoder (PromptEncoder): Encodes various types of input prompts. mask_decoder (MaskDecoder): Predicts masks from the image embeddings and encoded prompts. pixel_mean (list[float]): Mean values for normalizing pixels in the input image. pixel_std (list[float]): Standard deviation values for normalizing pixels in the input image. Examples: >>> image_encoder = ImageEncoderViT(...) >>> prompt_encoder = PromptEncoder(...) >>> mask_decoder = MaskDecoder(...) >>> sam_model = SAMModel(image_encoder, prompt_encoder, mask_decoder) >>> # Further usage depends on SAMPredictor class Notes: All forward() operations moved to SAMPredictor. pixel_meanrF pixel_stdN) super__init__ image_encoderprompt_encoder mask_decoderregister_buffertorchTensorview)selfrrrrr __class__s \/home/james-whalen/.local/lib/python3.13/site-packages/ultralytics/models/sam/modules/sam.pyrSAMModel.__init__8sy8 *,( \5<< +C+H+HQPQ+RTYZ [%,,y*A*F*Fr1a*PRWXc [URS5(aURRU5 XRlUVs/sHo"S-PM snURlUSURlgs snf)CSet image size to make model compatible with different image sizes. set_imgszrN)hasattrrr*rinput_image_sizeimage_embedding_sizeimg_sizer#imgszxs r%r*SAMModel.set_imgsz[sl 4%%{ 3 3    ( ( //4,EJ3KUGU3K0&+Ah#4Ls B)rrr))g33333^@gR]@gRY@)g(\2M@g(\L@gL@) rr rr rr r list[float]rr4returnNone) __name__ __module__ __qualname____firstlineno____doc__r__annotations__rr*__static_attributes__ __classcell__r$s@r%rrss8 NE#^\rSrSr%SrSrS\S'SSU4Sjjjr\S5r Sr S r SS jr SS jr SS jrS rSSjrSrSrSrSSjrSr\S5rSSjrSrSrU=r$) SAM2Modelda SAM2Model class for Segment Anything Model 2 with memory-based video object segmentation capabilities. This class extends the functionality of SAM to handle video sequences, incorporating memory mechanisms for temporal consistency and efficient tracking of objects across frames. Attributes: mask_threshold (float): Threshold value for mask prediction. image_encoder (ImageEncoderViT): Visual encoder for extracting image features. memory_attention (nn.Module): Module for attending to memory features. memory_encoder (nn.Module): Encoder for generating memory representations. num_maskmem (int): Number of accessible memory frames. image_size (int): Size of input images. backbone_stride (int): Stride of the backbone network output. sam_prompt_embed_dim (int): Dimension of SAM prompt embeddings. sam_image_embedding_size (int): Size of SAM image embeddings. sam_prompt_encoder (PromptEncoder): Encoder for processing input prompts. sam_mask_decoder (SAM2MaskDecoder): Decoder for generating object masks. obj_ptr_proj (nn.Module): Projection layer for object pointers. obj_ptr_tpos_proj (nn.Module): Projection for temporal positional encoding in object pointers. hidden_dim (int): Hidden dimension of the model. mem_dim (int): Memory dimension for encoding features. use_high_res_features_in_sam (bool): Whether to use high-resolution feature maps in the SAM mask decoder. use_obj_ptrs_in_encoder (bool): Whether to cross-attend to object pointers from other frames in the encoder. max_obj_ptrs_in_encoder (int): Maximum number of object pointers from other frames in encoder cross-attention. add_tpos_enc_to_obj_ptrs (bool): Whether to add temporal positional encoding to object pointers. proj_tpos_enc_in_obj_ptrs (bool): Whether to add an extra linear projection layer for temporal positional encoding in object pointers. use_signed_tpos_enc_to_obj_ptrs (bool): Whether to use signed distance in temporal positional encoding. only_obj_ptrs_in_the_past_for_eval (bool): Whether to only attend to object pointers in the past during evaluation. pred_obj_scores (bool): Whether to predict if there is an object in the frame. pred_obj_scores_mlp (bool): Whether to use an MLP to predict object scores. fixed_no_obj_ptr (bool): Whether to have a fixed no-object pointer when there is no object present. soft_no_obj_ptr (bool): Whether to mix in no-object pointer softly for easier recovery and error mitigation. use_mlp_for_obj_ptr_proj (bool): Whether to use MLP for object pointer projection. no_obj_embed_spatial (torch.Tensor | None): No-object embedding for spatial frames. max_cond_frames_in_attn (int): Maximum number of conditioning frames to participate in memory attention. directly_add_no_mem_embed (bool): Whether to directly add no-memory embedding to image feature on the first frame. multimask_output_in_sam (bool): Whether to output multiple masks for the first click on initial conditioning frames. multimask_min_pt_num (int): Minimum number of clicks to use multimask output in SAM. multimask_max_pt_num (int): Maximum number of clicks to use multimask output in SAM. multimask_output_for_tracking (bool): Whether to use multimask output for tracking. use_multimask_token_for_obj_ptr (bool): Whether to use multimask tokens for object pointers. iou_prediction_use_sigmoid (bool): Whether to use sigmoid to restrict IoU prediction to [0-1]. memory_temporal_stride_for_eval (int): Memory bank's temporal stride during evaluation. non_overlap_masks_for_mem_enc (bool): Whether to apply non-overlapping constraints on object masks in memory encoder during evaluation. sigmoid_scale_for_mem_enc (float): Scale factor for mask sigmoid probability. sigmoid_bias_for_mem_enc (float): Bias factor for mask sigmoid probability. binarize_mask_from_pts_for_mem_enc (bool): Whether to binarize sigmoid mask logits on interacted frames with clicks during evaluation. use_mask_input_as_output_without_sam (bool): Whether to directly output the input mask without using SAM prompt encoder and mask decoder on frames with mask input. Methods: forward_image: Process image batch through encoder to extract multi-level features. track_step: Perform a single tracking step, updating object masks and memory features. set_binarize: Set binarize for VideoPredictor. set_imgsz: Set image size to make model compatible with different image sizes. Examples: >>> model = SAM2Model(image_encoder, memory_attention, memory_encoder) >>> image_batch = torch.rand(1, 3, 512, 512) >>> features = model.forward_image(image_batch) >>> track_results = model.track_step(0, True, features, None, None, None, {}) rrrc$>[T$U]5 XlXlU (aSOSUlUUlUUlU(a%[RRSSSSS9Ul UUl U(a U(deUUl UUl UUlX lUR UlX0lUR"Ul[)UR$S5(aW[)UR$R*S5(a2UR$R*R,R.SUlX@l[RR3[R4"USSUR&55UlR6SS 9 [RR3[R4"SSUR"55Ul[RR3[R4"SSUR"55Ul[9UR:SS 9 [9UR<SS 9 XlXpl Xl!Xl"UUl#UUl$Xl%Xl&Xl'UUl(UUl)UUl*UUl+XPl,X`l-U"Ul.UUl/UUl0UUl1UUl2URb(a&UR^(deUR (deUR^(ahUR (aW[RR3[R4"SUR"55Ul3[9URfSS 9 U Ul4S Ul5U!(aW[RR3[R4"SUR&55Ul5[9URjSS 9 URm5 Xl7U#(aQ[pRr"S 5 [Rt"URRvS S SS9URl;g g )a Initialize the SAM2Model for video object segmentation with memory-based tracking. Args: image_encoder (nn.Module): Visual encoder for extracting image features. memory_attention (nn.Module): Module for attending to memory features. memory_encoder (nn.Module): Encoder for generating memory representations. num_maskmem (int): Number of accessible memory frames. image_size (int): Size of input images. backbone_stride (int): Stride of the image backbone output. sigmoid_scale_for_mem_enc (float): Scale factor for mask sigmoid probability. sigmoid_bias_for_mem_enc (float): Bias factor for mask sigmoid probability. binarize_mask_from_pts_for_mem_enc (bool): Whether to binarize sigmoid mask logits on interacted frames with clicks during evaluation. use_mask_input_as_output_without_sam (bool): Whether to directly output the input mask without using SAM prompt encoder and mask decoder on frames with mask input. max_cond_frames_in_attn (int): Maximum number of conditioning frames to participate in memory attention. directly_add_no_mem_embed (bool): Whether to directly add no-memory embedding to image feature on the first frame. use_high_res_features_in_sam (bool): Whether to use high-resolution feature maps in the SAM mask decoder. multimask_output_in_sam (bool): Whether to output multiple masks for the first click on initial conditioning frames. multimask_min_pt_num (int): Minimum number of clicks to use multimask output in SAM. multimask_max_pt_num (int): Maximum number of clicks to use multimask output in SAM. multimask_output_for_tracking (bool): Whether to use multimask output for tracking. use_multimask_token_for_obj_ptr (bool): Whether to use multimask tokens for object pointers. iou_prediction_use_sigmoid (bool): Whether to use sigmoid to restrict IoU prediction to [0-1]. memory_temporal_stride_for_eval (int): Memory bank's temporal stride during evaluation. non_overlap_masks_for_mem_enc (bool): Whether to apply non-overlapping constraints on object masks in memory encoder during evaluation. use_obj_ptrs_in_encoder (bool): Whether to cross-attend to object pointers from other frames in the encoder. max_obj_ptrs_in_encoder (int): Maximum number of object pointers from other frames in encoder cross-attention. add_tpos_enc_to_obj_ptrs (bool): Whether to add temporal positional encoding to object pointers in the encoder. proj_tpos_enc_in_obj_ptrs (bool): Whether to add an extra linear projection layer for temporal positional encoding in object pointers. use_signed_tpos_enc_to_obj_ptrs (bool): Whether to use signed distance in the temporal positional encoding in the object pointers. only_obj_ptrs_in_the_past_for_eval (bool): Whether to only attend to object pointers in the past during evaluation. pred_obj_scores (bool): Whether to predict if there is an object in the frame. pred_obj_scores_mlp (bool): Whether to use an MLP to predict object scores. fixed_no_obj_ptr (bool): Whether to have a fixed no-object pointer when there is no object present. soft_no_obj_ptr (bool): Whether to mix in no-object pointer softly for easier recovery and error mitigation. use_mlp_for_obj_ptr_proj (bool): Whether to use MLP for object pointer projection. no_obj_embed_spatial (bool): Whether add no obj embedding to spatial frames. sam_mask_decoder_extra_args (dict | None): Extra arguments for constructing the SAM mask decoder. compile_image_encoder (bool): Whether to compile the image encoder for faster inference. Examples: >>> image_encoder = ImageEncoderViT(...) >>> memory_attention = SAM2TwoWayTransformer(...) >>> memory_encoder = nn.Sequential(...) >>> model = SAM2Model(image_encoder, memory_attention, memory_encoder) >>> image_batch = torch.rand(1, 3, 512, 512) >>> features = model.forward_image(image_batch) >>> track_results = model.track_step(0, True, features, None, None, None, {}) r) kernel_sizestrideout_projweightrg{Gz?)stdNzFImage encoder compilation is enabled. First forward pass will be slow.z max-autotuneTF)mode fullgraphdynamic)$'>$ "$)88??1aQq?#QD (@% $+ ++)B&/N,2T/!1*22- 4&& 3 3@S@S@\@\^f8g8g..77>>DDQGDL& % 2 25;;{AqRVR^R^3_ `d++6!HH..u{{1a/QR#hh00Q4??1STd''T2d))t4)B&*C&(@%2T/-J*/N,5Y1'>$$8!$8!-J*/N,*D'%.+F(.#6 0.  '' ''// //   D$@$@#hh00Q1PQDO $//t 4(@%$(! (-(:(:5;;q$,,;W(XD % $33 > '>$ ! KK` a).""**# *D   & !r'cH[UR55R$)z=Return the device on which the model's parameters are stored.)next parametersdevicer#s r%rSAM2Model.devicelsDOO%&---r'c[S5e)zWProcess image and prompt inputs to generate object masks and scores in video sequences.zPlease use the corresponding methods in SAM2VideoPredictor for inference.See notebooks/video_predictor_example.ipynb for an example.)NotImplementedError)r#argskwargss r%rSAM2Model.forwardqs! J  r'c(URUlURUR-Ul[ URURUR4URUR4SS9Ul[S S[SURSSS9URSSURURURURURS . UR=(d 0D6UlUR (a|["R$R'URUR5UlUR*(a1[-URURURS5UlO#["R$R/5UlUR0(a:["R$R'URUR25Ulg ["R$R/5Ulg ) zMBuild SAM-style prompt encoder and mask decoder for image segmentation tasks.r+) embed_dimr.r- mask_in_chansrDi)depth embedding_dimmlp_dim num_heads) num_multimask_outputs transformertransformer_dimiou_head_depthiou_head_hidden_dimuse_high_res_featuresrprtruroN)rZsam_prompt_embed_dimrqrrsam_image_embedding_sizer sam_prompt_encoderr r rNrprtrurorssam_mask_decoderrPr rLinear obj_ptr_projryrIdentityrUr\obj_ptr_tpos_projrs r%r{SAM2Model._build_sam_headsxs$(OO!(,4;O;O(O%#0//----"#oot?# !0! "#-"77  !55 #"&"C"C'+'F'F 00 $ 8 8,0,P,P! //52!! $  ' ' % QD ,,$'$//[\$]! % 1 1 3D   ) )&+XX__T__dll%SD "%*XX%6%6%8D "r'c URSnURnURS5UR:XdeURS5UR:XdeURS5UR:XdeUb3USnUSn URSU:XaU RSU:XdeOG[ R "USSXqRS9n[ R"US[ RUS 9*n Ub[UR5S :XaURSSUS4:XdeURS SURR:wa;[R"UR5URRS S SS9n OUn OSn URX4SU S9upUR!UURR#5U U US US9upnnUR$(a)US:n[ R&"USS2SS4U [(5n [R"U UR*UR*4S S S9nUSS2S4nU(ar[ R,"USS9n[ R."XgS9nU UU4R1S5nUUU4R1S5nURS5S:aUUU4nOU UnnUR3U5nUR$(ahUR4(aUR75nOWR9UR5nUR:(aUU-nUSU- UR<--nU UUUUUU4$)aC Forward pass through SAM prompt encoders and mask heads. This method processes image features and optional point/mask inputs to generate object masks and scores. Args: backbone_features (torch.Tensor): Image features with shape (B, C, H, W). point_inputs (dict[str, torch.Tensor] | None): Dictionary containing point prompts. 'point_coords': Tensor of shape (B, P, 2) with float32 dtype, containing absolute pixel-unit coordinates in (x, y) format for P input points. 'point_labels': Tensor of shape (B, P) with int32 dtype, where 1 means positive clicks, 0 means negative clicks, and -1 means padding. mask_inputs (torch.Tensor | None): Mask of shape (B, 1, H*16, W*16), float or bool, with the same spatial size as the image. high_res_features (list[torch.Tensor] | None): List of two feature maps with shapes (B, C, 4*H, 4*W) and (B, C, 2*H, 2*W) respectively, used as high-resolution feature maps for SAM decoder. multimask_output (bool): If True, output 3 candidate masks and their IoU estimates; if False, output only 1 mask and its IoU estimate. Returns: low_res_multimasks (torch.Tensor): Tensor of shape (B, M, H*4, W*4) with SAM output mask logits. high_res_multimasks (torch.Tensor): Tensor of shape (B, M, H*16, W*16) with upsampled mask logits. ious (torch.Tensor): Tensor of shape (B, M) with estimated IoU for each output mask. low_res_masks (torch.Tensor): Tensor of shape (B, 1, H*4, W*4) with the best low-resolution mask. high_res_masks (torch.Tensor): Tensor of shape (B, 1, H*16, W*16) with the best high-resolution mask. obj_ptr (torch.Tensor): Tensor of shape (B, C) with object pointer vector for the output mask. object_score_logits (torch.Tensor): Tensor of shape (B) with object score logits. Examples: >>> backbone_features = torch.rand(1, 256, 32, 32) >>> point_inputs = {"point_coords": torch.rand(1, 2, 2), "point_labels": torch.tensor([[1, 0]])} >>> mask_inputs = torch.rand(1, 1, 512, 512) >>> results = model._forward_sam_heads(backbone_features, point_inputs, mask_inputs) >>> ( ... low_res_multimasks, ... high_res_multimasks, ... ious, ... low_res_masks, ... high_res_masks, ... obj_ptr, ... object_score_logits, ... ) = results rrrrDN point_coords point_labelsrdtype)rrrEFbilinearTsize align_cornersrK antialias)pointsboxesmasks)image_embeddingsimage_pesparse_prompt_embeddingsdense_prompt_embeddingsmultimask_output repeat_imagehigh_res_features)rrKrrdimr)r]rrrrr r`ronesint32lenrmask_input_sizeF interpolaterr get_dense_pertwhere NO_OBJ_SCORErqargmaxarange unsqueezerrwsigmoidtorvrx)r#backbone_features point_inputs mask_inputsrrBrsam_point_coordssam_point_labelssam_mask_promptsparse_embeddingsdense_embeddingslow_res_multimasksioussam_output_tokensobject_score_logitsis_obj_appearinghigh_res_multimaskssam_output_token best_iou_inds batch_inds low_res_maskshigh_res_masksobj_ptrlambda_is_obj_appearings r%_forward_sam_headsSAM2Model._forward_sam_headssh  # #A &")) %%a(D,E,EEEE %%a(D,I,IIII %%a(D,I,IIII  #+N; +N; #))!,16F6L6LQ6OST6T TT6T %{{1a6I`I`a  % 1au{{6 RR   "{(()Q.;3D3DRa3HQPQF3R RR  %)@)@)P)PP"#--%%'00@@"'#" ##.#O.2.E.E$7!/F/ + LPK`K`.,,99;%6$4-/LaL H"35H   2Q6 "'-=atm-LN`bn!o  mm //4??3  -QT2 !LL26Ma7J.z=/HISSTUVM0]1JKUUVWXN %%a(1,#4Z5N#O ,>@S>M##$45   ##*=*E*E*G'*:*=*=gmm*L'$$1G;%s* '--/*6A  %%b)Q.0C0CB0G10LM  ##K$5$5a$8!<BBD++/@/HL]Lekk+"3"3A"6P[PbPbcG)-(?(?"3 001BC"3)@) %Aq!QA!99[%8%8%;%A%A%Cc%IqQ+I6"2"8"8":'AHL   $$1;%n'[6R@"U#XSRBS9n([EU(U&- U'S9n(URGU(5n(U(RIS5RKSXR>5n(O%U%RM[ U#5XR>5n(UR>U :ajU%ROSXUR>-UR>5n%U%R SSSS5R#SS5n%U(RQXR>-SS9n(URU%5 URU(5 U%RRSnOSnOURT(a7USURV-n)U)R SSS5R XX5n)U)$URVRKSXR>5/nURXRKSXR>5/n[6RZ"USS9n*[6RZ"USS9n+UR]UUU*U+US9n)U)R SSS5R XX5n)U)$s snfs snnfs snnf)zePrepare memory-conditioned features by fusing current frame's visual features with previous memories.rrrrcond_frame_outputsnon_cond_frame_outputsNmaskmem_featurescuda)r non_blockingmaskmem_pos_encrrrrrD)currcurr_posmemory memory_posnum_obj_ptr_tokens)/rrZrr^rr"rrr|valuestrainingrirangegetappendrtyperrarPminrQrWitemsrVabszipr stackrTrUr\tensorrrrrexpand new_zerosreshaperepeat_interleaver]rdrbrccatrX),r# frame_idxis_init_cond_framecurrent_vision_featscurrent_vision_pos_embedsr output_dict num_framestrack_in_reverserCHWrr  tpos_sign_mul to_cat_memoryto_cat_memory_pos_embed cond_outputsselected_cond_outputsunselected_cond_outputsoutt_pos_and_prevsrt_post_relprev_frame_idxprevfeats maskmem_encrQtptr_cond_outputs pos_and_ptrst_diffpos_list ptrs_listobj_ptrs t_diff_maxtpos_dimobj_pospix_feat_with_memr memory_pos_embeds, r%$_prepare_memory_conditioned_features.SAM2Model._prepare_memory_conditioned_featuress ! $ ) )! , OO"~%b)00   q '+33Aq!<AA!M M.A !572M{#789A= ==&'; : !#:4I3O3O3QR3QC3x3QOR ]](L(LAq$"2"23((50A::JY%6PY\aPaN)(11}&:a%?N%3uqyAo%EN*3Q'71'<%=%AN%3uqyAo%EN!":;??PTU;255ndKC&&s|4-40 / t</0336P[P[_ePe3f$$U]]1%5%=%=aA%FG"#45b9<AN*AN3&C%($ (=$#3"8"8": #;3 $CC']m;!$Y]!3I #; $A'>?F.> F*IPVDVA1u!7AO%&>?CCAG^GbGbcdfjGklC$++VS^,DE @ *-|*<'Hi${{9!CCA!O  ASd( s YY$0Y$9Y$9Y*cUSRS5nURnUSupUSRSSS5RXgX5n UR(a"UR (dUR U5nUR=(a Un U (a0UR (dUS:RU R5n O[R"U5n URS:waXR-n URS:waXR-n URXSS9n U S nU S nURbEUS:R!5nUSUS - URS R""UR$6-- nX4$) zXEncode frame features and masks into a new memory representation for video segmentation.rrrr?rT)skip_mask_sigmoidvision_featuresr).NN)rrZrr"rhr"_apply_non_overlapping_constraintsrgrrr rrerfr[rzrrr])r#r rpred_masks_high_resris_mask_from_ptsrr%r&r'pix_featbinarize mask_for_mem maskmem_outrrrs r%_encode_new_memorySAM2Model._encode_new_memory*s ! $ ) )! , OO"~'+33Aq!<AA!M  - -dmm#'"I"IJ]"^ ::O?O DMM/!377GL!==)<=L  ) )S 0'*H*HHL  ( (C /'*G*GGL))(TX)Y &'89%&67  $ $ 0 3a 7>>@  %5o%F!F$JcJcKfK&,,K.!. .  00r'c d[U5S:ah[USSUSS5V V s/sHHupU RSSS5R"U R S5U R S5/U Q76PMJ nn n OSnUb[UR (aJUSRSSS5nUR"SUR /USQ76nURXU5nOPURUUUSSUSSUSSUU U S9nU b UbUbeU nURX&5nURUUUUUS9nUX4$s sn n f)hPerform a single tracking step, updating object masks and memory features based on current frame inputs.rNrrr)rrr r!rr"r#r$)rrrrr) rrrr"rrjrZrrC_use_multimaskr)r#rrr r!rrrr"r#r$prev_sam_mask_logitsr2srrL sam_outputsrs r% _track_stepSAM2Model._track_stepVs # $q ( 4Sb 9:cr?K!KDA !Q"''q 166!9AqAK !  !%   "t'P'P,B/771a@H}}RJ:b>JH22;J[\K@@##5%9"#%>*CBC*H%bc?'%!1A H$/#/K4GGG2 #223ET 11"*)'"3!1 2K-77O!sAD,cU(a0URS:a URUUUUUSLS9upXS'XS'gSUS'SUS'g)z^Run memory encoder on predicted mask to encode it into a new memory feature for future frames.rN)r rrJrrKrr)r^rP) r#r rrrun_mem_encoderrr current_outrrs r%_encode_memory_in_output"SAM2Model._encode_memory_in_outputsl t//!3040G0G%9%$2$7".d": 1H1 - /?* +-<) *.2K* +-1K) *r'c URUUUUUUUUU U U 5 un nU u pnnnUUUS.nUR(dUUS'URUUUU UUU5 U$)rS) pred_masksrJrr)rXrr])r#rrr r!rrrr"r#r$r[rUrWrrrrrr\s r% track_stepSAM2Model.track_steps,!,,   %          QP[L1a9L(#1 }}2EK- . %%       r'cUcSOUSRS5nUR=(aE U=(d UR=(a) URUs=:*=(a UR:*$s $)zaDetermine whether to use multiple mask outputs in the SAM head based on configuration and inputs.rrr)rrkrnrlrm)r#rrnum_ptss r%rTSAM2Model._use_multimasksn#+!n1M1R1RST1U  ( ( T#It'I'I T**gRR9R9RR S r'c URSnUS:XaU$URn[R"USSS9n[R"XS9SS2SSS4nX4:Hn[R "XP[R "USS95nU$) z\Apply non-overlapping constraints to masks, keeping the highest scoring object per location.rrT)rkeepdimrNr)max)r]rr rrrclamp)r` batch_sizer max_obj_indsbatch_obj_indskeeps r%rI,SAM2Model._apply_non_overlapping_constraintss %%a( ? ""||JAtD j@D$PTATU-[[5;;zu3UV r'cXlg)z Set binarize for VideoPredictor.N)rg)r#rMs r% set_binarizeSAM2Model.set_binarizes2:/r'cUSUlXRlUVs/sHo"S-PM snURlURUR-Ulgs snf)r)rr+N)rqrr-r.rrrr0s r%r*SAM2Model.set_imgszsU(380IN7OAR7O4(,4;O;O(O%8PsA)0rTrrrgrdrvrZrrqrprSrar|rQr\rXr[rirmrlrnrkrbrcrzrxrhrOr^rrrWrtrurUrrrsrrrfrerwrNrjryrorPrV) ir+rFrFFrFFFrrFFFrFFr+TFFFFFFFFFNF)roboolrtrurururvrurwruryrurzrurru)NNNF)NN)rz torch.Tensor)F)FTN)r7r8r9r:r;rr<rpropertyrrr{rrrrrCrPrXr]rarT staticmethodrIrpr*r=r>r?s@r%rArAdsDL NE"%!$+0-2 ""'%* %&+05#(()&+ % "!%"'(-+0 %$)!& %).%*$(&+I}&*.'}:;}<"=}>?}@A}B#'C}D#E}H $I}}~.. -9d U n- ^ I0b!H*1X88t2H !':x ";PPr'rA) __future__rr torch.nn.functionalr functionalr torch.nn.initrultralytics.nn.modulesrultralytics.utilsrblocksr decodersr r encodersr r utilsrrrModulerrArr'r%rs] # '&$)24= H/ryyH/VePePr'