+ UV-j< Rt^RIt^RIHtHtHtHt^RIHt ^RI H t ^RI H t ^RIHtHtHtHt!RR] P&4t!RR] P&4t!R R ] P&4t!R R ] P&4t!R R] P&4t!RR] P&4t!RR] P&4t!RR] P&4t!RR] P&4t!RR] P&4tR#)ztSAM3 Tracker: SAM2-style memory-based tracker for video segmentation. Weight keys: tracker_model.*, tracker_neck.* N)DictListOptionalTuple TrackerConfig) LayerNorm2d RoPEAttentionSAMMaskDecoderSAMPromptEncodercTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #)SimpleMaskDownSamplerzoProgressive conv downsampling for mask encoding. Weight keys: tracker_model.memory_encoder.mask_downsampler.* c <V^8dQhRS[/#configr)format __classdict__s"l/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/mlx_vlm/models/sam3/tracker.py __annotate__"SimpleMaskDownSampler.__annotate__sTT}Tc t<\SV`4VPpVPpVPpVP p^^^^@V.p.Vn\^4F=pVP P\Wg,Wg^,,W4V44K? \P!W"^RR7Vn R#T) kernel_sizebiasN) super__init__mask_downsampler_embed_dimmask_downsampler_kernel_sizemask_downsampler_stridemask_downsampler_paddinglayersrangeappendDownsampleConvBlocknnConv2d final_conv) selfr embed_dimrstridepaddingchannelsi __class__s && rrSimpleMaskDownSampler.__init__s 55 99 //11q"b), qA KK  #Ka%+w ))IadSrcN<V^8dQhRS[PRS[P/#)rmasksreturnmxarray)rrs"rrr-s#  bhh 288 rc bTpVPF pV!V4pK VPV4pV#)zj Args: masks: (B, H, W, 1) binary masks Returns: (B, H', W', embed_dim) downsampled mask features )r#r))r*r3xlayers&& r__call__SimpleMaskDownSampler.__call__-s3 [[EaA! OOA r)r)r# __name__ __module__ __qualname____firstlineno____doc__rr;__static_attributes____classdictcell__ __classcell__r0rs@@rr r s% TT&   rr cTaa]tRt^;toRtV3RlV3RlltV3RlRltRtVtV;t #)r&z%Single conv + layernorm + GELU block.c 8<V^8dQhRS[RS[RS[RS[RS[/#)rin_chout_chrr,r-int)rrs"rr DownsampleConvBlock.__annotate__>s5..."%.47.AD.OR.rc<\SV`4\P!WW4VR7Vn\ V4VnR#))r,r-N)rrr'r(convr layer_norm)r*rIrJrr,r-r0s&&&&&&rrDownsampleConvBlock.__init__>s5 II ;w  &f-rcN<V^8dQhRS[PRS[P/#rr9r4r5)rrs"rrrMGs#"((rxxrcvVPV4pVPV4p\P!V4pV#N)rOrPr'gelu)r*r9s&&rr;DownsampleConvBlock.__call__Gs/ IIaL OOA  GGAJr)rOrPr=rFs@@rr&r&;s!/..rr&cTaa]tRt^NtoRtV3RlV3RlltV3RlRltRtVtV;t #)CXBlockzkConvNeXt-style block with depthwise conv. Weight keys: tracker_model.memory_encoder.memory_fuser.layers.* c <V^8dQhRS[/#rr)rrs"rrCXBlock.__annotate__Ts R R} Rrc<\SV`4VPpVPpVPp\ P !W"W4VR7Vn\V4Vn \ P!W!P4Vn \ P!VPV4Vn \P!V34VP ,VnR#))rr-groupsN)rrmemory_fuser_embed_dimmemory_fuser_kernel_sizememory_fuser_paddingr'r(depthwise_convrrPLinearmemory_fuser_intermediate_dimpointwise_conv1pointwise_conv2r6ones#memory_fuser_layer_scale_init_valuescale)r*rdimrr-r0s&& rrCXBlock.__init__Ts ++55 -- ii +s &c*!yy.R.RS!yy)M)MsSWWcV_v'Q'QQ rcN<V^8dQhRS[PRS[P/#rSr5)rrs"rrr[bs#"((rxxrc pTpVPV4pVPV4pVPwr4rVVPW4,V,V4pVP V4p\ P !V4pVPV4pVPW4WV4pVPV,pW!,#)z Args: x: (B, H, W, C) ) rarPshapereshaperdr'rVrerh)r*r9residualBHWCs&& rr;CXBlock.__call__bs     " OOA WW a IIaeai #   # GGAJ   # IIaA ! JJN|r)rarPrdrerhr=rFs@@rrYrYNs% R RrrYcTaa]tRt^ttoRtV3RlV3RlltV3RlRltRtVtV;t #) MemoryFuserzpStack of CXBlocks for fusing mask and image features. Weight keys: tracker_model.memory_encoder.memory_fuser.* c <V^8dQhRS[/#rr)rrs"rrMemoryFuser.__annotate__zsWW}Wrc<\SV`4\VP4Uu.uFp\ V4NK upVnR#uupirU)rrr$memory_fuser_num_layersrYr#r*r_r0s&& rrMemoryFuser.__init__zs8 05f6T6T0UV0U1wv0UV VsAcN<V^8dQhRS[PRS[P/#rSr5)rrs"rrrx~s#"((rxxrc<VPF pV!V4pK V#rUr#r*r9r:s&& rr;MemoryFuser.__call__~s[[EaA!rrr=rFs@@rrvrvts% WWrrvcTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #) MemoryEncoderzcEncodes image features + mask into compressed memory. Weight keys: tracker_model.memory_encoder.* c <V^8dQhRS[/#rr)rrs"rrMemoryEncoder.__annotate__sLL}Lrc <\SV`4VPpVPp\ V4Vn\ V4Vn\P!W"^RR7Vn \P!W#^RR7Vn R#r) rrmemory_encoder_hidden_sizememory_encoder_output_channelsr mask_downsamplerrv memory_fuserr'r(feature_projection projection)r*rriout_dimr0s&& rrMemoryEncoder.__init__sg //77 5f ='/"$))C!$"O))CadKrch<V^8dQhRS[PRS[PRS[P/#)rfeaturesr3r4r5)rrs"rrrs5((xx  rc VPV4pVPV4pW,pVPV4pVPV4pV#)z Args: features: (B, H, W, D) backbone features masks: (B, H_mask, W_mask, 1) predicted masks Returns: (B, H', W', out_dim) compressed memory )rrrr)r*rr3 mask_featuresfusedmemorys&&& rr;MemoryEncoder.__call__sQ--e4 **84(!!%(' r)rrrrr=rFs@@rrrs% LLrrcTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #)MemoryAttentionLayeraQSingle layer of memory attention: self-attn (RoPE) + cross-attn (RoPE) + FFN. Weight keys per layer: self_attn.{q,k,v,o}_proj.{weight,bias} cross_attn_image.{q,k,v,o}_proj.{weight,bias} layer_norm1.{weight,bias} layer_norm2.{weight,bias} layer_norm3.{weight,bias} linear1.{weight,bias} linear2.{weight,bias} c <V^8dQhRS[/#rr)rrs"rr!MemoryAttentionLayer.__annotate__sVV}Vrc <\SV`4VPp\VVPVP \ VP4VPR7Vn \VVPVP \ VP4VPVPRR7Vn \P!V4Vn\P!V4Vn\P!V4Vn\P"!W!P$4Vn\P"!VP$V4VnR#)) hidden_size num_headsdownsample_rate feat_sizes rope_thetaT)rrrrrkv_dim rope_k_repeatN)rrmemory_attention_hidden_sizer $memory_attention_num_attention_heads memory_attention_downsample_ratetuple memory_attention_rope_feat_sizesmemory_attention_rope_theta self_attnrcross_attn_imager' LayerNorm layer_norm1 layer_norm2 layer_norm3rb)memory_attention_feed_forward_hidden_sizelinear1linear2)r*rdr0s&& rrMemoryAttentionLayer.__init__s   / /&AA"CCVDDE99  !.AA"CCVDDE9988! <<?<<?<<?yy$T$TU yy!Q!QSTU rch<V^8dQhRS[PRS[PRS[P/#rsrcrr4r5)rrs"rrrs5 XX  rc LVPWV4pW,pVPV4pVPWV4pW,pVPV4pVP \ P !VPV444pW,pVPV4pV#)zb Args: src: (B, HW, D) current frame features memory: (B, N_mem, mem_dim) memory features ) rrrrrr'relurr)r*rrsrc2s&&& rr;MemoryAttentionLayer.__call__s~~c,js#$$S&9js#||BGGDLL$567js# r)rrrrrrrr=rFs@@rrrs% VV8rrcTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #)MemoryAttentionz]Memory attention module with multiple layers. Weight keys: tracker_model.memory_attention.* c <V^8dQhRS[/#rr)rrs"rrMemoryAttention.__annotate__sLL}Lrc<\SV`4\VP4Uu.uFp\ V4NK upVn\ P!VP4Vn R#uupirU) rrr$memory_attention_num_layersrr#r'rrrPr{s&& rrMemoryAttention.__init__s\ 6==> > ! (>  ,,v'J'JK  sA*ch<V^8dQhRS[PRS[PRS[P/#rr5)rrs"rrrs5 XX  rc ^VPF pV!W4pK VPV4pV#)z Args: src: (B, HW, D) current frame features memory: (B, N_mem, mem_dim) memory bank Returns: (B, HW, D) attended features )r#rP)r*rrr:s&&& rr;MemoryAttention.__call__s/[[E$C!ooc" r)rPr#r=rFs@@rrrs% LLrrcTaa]tRtRtoRtV3RlV3RlltV3RlRltRtVtV;t #)ObjectPointerMLPizaProjects SAM output tokens to object pointers. Weight keys: tracker_model.object_pointer_proj.* c <V^8dQhRS[/#)rrrK)rrs"rrObjectPointerMLP.__annotate__!s<!W#4Vn \CV^,4Vn"VPF'd \&P(!^V34Vn$R#R#)rT)rr,rN)%rrrrr image_size vision_configbackbone_config patch_sizer prompt_encoder_configprompt_encoderr mask_decoder_config mask_decoderrmemory_attentionrmemory_encoderr6zerosno_memory_embeddingno_memory_positional_encodingno_object_pointer num_maskmem#memory_temporal_positional_encodingrobject_pointer_projr'r(mask_downsamplerb-temporal_positional_encoding_projection_layerSharedImageEmbeddingshared_image_embedding"enable_occlusion_spatial_embedding%occlusion_spatial_embedding_parameter)r*rrmem_dimr feat_sizer0s&& rrTrackerModel.__init__9su   / /77&& "6"6"F"F"Q"QQ /v/K/KL*6+E+EF!0 7+F3$&88Q1I#6 -/XXq!Qi-@*!#1a&!14688   Aw /4 0 $4A#6  "yyA1QTR>@YYq=R:';16&B#  4 4 49;1g,9OD 6 5rcN<V^8dQhRS[PRS[P/#)rbackbone_featuresr4r5)rrs"rrrcs#66bhh62886rc RVPwr#rEVPW#V,V4#)z/Flatten backbone features for memory attention.)rmrn)r*rrprqrrrss&& r encode_imageTrackerModel.encode_imagecs(&,, a ((E155rc<V^8dQhRS[PRS[S[S[P,,RS[S[S[P,,RS[S[S[PS[P3,,RS[S[P,RS[S[P,RS[RS[S[S[P,,R S[S[S[P3,/ #) rcurrent_features memory_bank memory_pos prompt_points prompt_boxes prompt_masksmultimask_outputhigh_res_featuresr4)r6r7rrrboolrstr)rrs"rrrhsQ Q ((Q d288n-Q T"((^, Q  bhh&8 9: Q rxx( Q rxx(Q Q $DN3Q  c288m Q rc  VPwrrVPWV ,V 4p V'd:\V4^8d*\P!V^R7pVP W4p VP P4p\P!WW,V 34pVP VVVR7wppVPV VVVVVR7wppppVPVR ,4pVR ,pVP^^^^4pRpVP^,V8wdS\P!VVP^,, VVP^,, 3RR7pV!V4pVPVV4pVPwppppVPVVV,V4pRVRVR VR VR V/#)a[Run one tracking step. Args: current_features: (B, H, W, D) backbone features for current frame memory_bank: list of (B, HW, mem_dim) past memory features memory_pos: list of temporal positional encodings prompt_*: optional prompt inputs for this frame Returns: dict with masks, iou_scores, obj_scores, object_pointer, memory axis)pointsboxesr3)image_embeddingsimage_pesparse_prompt_embeddingsdense_prompt_embeddingsrrinearest) scale_factormode pred_masks iou_scores obj_scoresobject_pointerr)NNN)r:rrN)rmrnlenr6 concatenaterr get_dense_pe broadcast_torr transposer'Upsampler)r*rrrrrrrrrprqrrDrrr sparse_emb dense_embr3iou_pred sam_tokens obj_scoreobj_ptr mask_for_memtargetupB_mH_mW_mC_ms&&&&&&&&& r track_stepTrackerModel.track_stephs*&++ a&&qa%3 3{+a/^^Ka8F''4C&&335??8];!% 3 3 !4!  I261B1B %/$--/ 2C2 .xY**:d+;<V} #--aAq9    a F *\//22\//22 Bl+L$$%5|D#\\S#sS3Y4 % ( ) g f   r)rrrrrrrrrrrrrr)NNNNNFN) r>r?r@rArBrrr%rCrDrErFs@@rrr3s3 (P(PT66 Q Q Q rrcXaa]tRtRtoRtRV3RlV3RllltV3RlRltRtVtV;t #) rizqShared positional embedding for tracker. Weight keys: tracker_model.shared_image_embedding.positional_embedding c <V^8dQhRS[/#)r num_pos_featsrK)rrs"rr!SharedImageEmbedding.__annotate__sAAcArc^<\SV`4\P!^V34VnR#)rN)rrr6rpositional_embedding)r*r)r0s&&rrSharedImageEmbedding.__init__s$ $&HHa-?$@!rcP<V^8dQhRS[S[S[3,RS[P/#)rsizer4)rrLr6r7)rrs"rrr*s+ I IU38_ I IrcVwr#\P!V4P\P4V, p\P!V4P\P4V, p\P!WERR7wrg\P !VP R4VP R4.RR7p^V,^, pWP,p^\P,V,p\P!\P!V4\P!V4.RR7#)ij)indexingr) r6arangeastypefloat32meshgridstackrnr,mathpirsincos) r*r/rqrrgrid_ygrid_xgygxcoordss && rr;SharedImageEmbedding.__call__s1$$RZZ0141$$RZZ014Vd;2::b>2::b>:DVa333TWWv%~~rvvf~rvvf~>RHHr)r,)r=rFs@@rrrs( AA I I Irr)rBr9typingrrrrmlx.corecorer6mlx.nnr'rrsam_componentsrr r r Moduler r&rYrvrrrrrrrrrKs  ..!XX$BII$N"))&#bii#L "))  %BII%ZB299BJbiiJ ryy 0F 299F RI299Ir