^jdZddlZddlmZmZmZmZddlZddlmZddl m Z ddl m Z ddl mZ dXd ed ed ed eej$fdZ dYd ededed eej$dej(f dZdddddej,fdeededededed eej$dej.dej(fdZdeedeefdZddddddd ddd dej,f deed!eej(d ed"eded ed#ed$ed%eeed&ed'ed eej$dej.deej(fd(ZGd)d*ej8Zd+Zd,ej(dej(fd-Z dZd,ej(d.ej(d/ej(d0edej(f d1Z dZd,eej(d.ej(d/ej(d0edeej(f d2Z! dZd,ej(d3ej(d0edej(fd4Z" dZd,ej(d5ej(d6ej(d7edej(f d8Z#dddddd ddd dej,f deed!eej(ded"eded ed$ed%eeed&ed'ed eej$dej.fd9Z$Gd:d;ej8Z%Gd<d=ej8Z& d[dej,d>d?ed@edAededBedej(f dCZ'ejPjRe d dej,fdDeed'ed eej$dej.deej(ej(ff dEZ*dFej(dGej(dHej(dej(fdIZ+GdJdKej8Z,GdLdMej8Z-ejPjRe dNd ddOej,fdPedQedRed'ed&ed ej$dej.dej(fdSZ.GdTdUej8Z/ d\dVededAedej8fdWZ0y)]zv Sin-cos, fourier, rotary position embedding modules and functions Hacked together by / Copyright 2022 Ross Wightman N)ListTupleOptionalUnion)nn)register_notrace_function)ndgrid)_assertT num_bandsmax_freq linear_bandsdevicec|r,tjd|dz |tj|}nBdtjdtj|ddz |tj|z}|tj zS)N?dtyperrr)torchlinspacefloat32mathlogpi)r r rrbandss g/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/timm/layers/pos_embed_sincos.pypixel_freq_bandsrsh sHqL)5==Y_`U^^Atxx!'>"  2 2D(QIkk599T?EIIdO<)LTTUVWG ::E: ""sA E seqcTt|dkr|S|d|dgt|ddzS)Nrrr)lenlist)rBs r swap_shape_xyrFPs4 3x!| FCF d3qr7m ++rijrmax_res include_grid in_pixelsref_feat_shape grid_offset grid_indexingc ||+|rt|t||| }n,t||d| }n| |j} | |j} | dk(rt |}| t |}|r6|D cgc]*} t jdd| | t j,}} nS|D cgc]H} t j| | t jjt j| zJ}} |(t|||Dcgc]\}}}||z |z}}}}t jt j|| d }|jd }||z}|j!j| |j#j| }}|r|||g}|S||g}|Scc} wcc} wcc}}}w) a Args: feat_shape: Feature shape for embedding. bands: Pre-calculated frequency bands. num_bands: Number of frequency bands (determines output dim). max_res: Maximum resolution for pixel based freq. temperature: Temperature for non-pixel freq. linear_bands: Linear band spacing for pixel based freq. include_grid: Include the spatial grid in output. in_pixels: Output in pixel freq. ref_feat_shape: Reference feature shape for resize / fine-tune. grid_offset: Constant offset to add to grid for non-pixel freq. grid_indexing: Indexing mode for meshgrid ('ij' or 'xy') dtype: Output dtype. device: Output device. Returns: )rrrr0xygr)stepsrrr2indexingr1r3r4)rfloatr(rrrFrrrr$r%r&zipr5meshgridr8r9r:)r*rr rJr rrKrLrMrNrOrrr<txfrr=pospos_sinpos_cosouts rbuild_fourier_pos_embedr`VsF } $g) E' E >\\F =KKE":.  %*>:N   NN3!F%-- P     LL6 = @ @ OR] ]  !&)!Z&H I I71aQUQY I I ;;u~~a-@b ID >>" D ,Cwwy||%|0#''),,U,2KWG&24' "C J:A'8JC J)   Js-/F-#A F2F7cTeZdZ ddedeffd Zd dZd dZd dZdZxZ S) FourierEmbedrJr ct|||_||_||_||_|j dtj|||d|jy)Nrr2F persistent) super__init__rJr concat_grid keep_spatialregister_bufferremptyreset_parameters)selfrJr rhrirr __class__s rrgzFourierEmbed.__init__s`  "&( Wekk)FRW&Xejk rc$|jyz"Initialize parameters and buffers.N _init_buffersrms rrlzFourierEmbed.reset_parameters rcv|jjt|j|jy.Compute and fill non-persistent buffer values.N)rcopy_rr rJrss rrrzFourierEmbed._init_bufferss$ )$..$,,GHrc$|jyz"Initialize non-persistent buffers.Nrqrss rinit_non_persistent_buffersz(FourierEmbed.init_non_persistent_buffersrtrc |jdd\}}|jdd}t||j|j|j|j }t j|d}|jddjt|}|fd|jdz zz}|jrKt j||jdj|jdd ddgd}|St j|jddd d|jdj|gd}|j!||j#d}|S) Nr)rKrrr1r3r1rr)shaper`rrhrrrcatr7r6rDndimrir8expandpermutereshapenumel)rmrYBCr*emb batch_expands rforwardzFourierEmbed.forwardsCwwr{1WWQR[ %  JJ))''88  ii$mmB#++C O<teqvvz22     1cmmA.55lCKKAqRSUVWX^_`A  199Q1a0#--2B2I2I,2WX^`aA !Z--/4Ar)rGr)TFNNr"N) __name__ __module__ __qualname__intrgrlrrr{r __classcell__rns@rrbrbsC   &Irrbctj|ddddf |ddddfgdj|jS)N.rrr1)rr5rr)rYs rrotrsE ;;319 qcc{3R 8 @ @ IIrrYcb|jdd\}}tj| |gdS)Nrr1r3)chunkrr)rYx1x2s rrope_rotate_halfrs1WWQBW FB 99rc2YB ''rsin_embcos_embhalfcV|r||zt||zzS||zt||zzSN)rr)rYrrrs rapply_rot_embedrs= 7{-a07::: 7{SVg---rct|tjr|g}|r"|Dcgc]}||zt||zzc}S|Dcgc]}||zt ||zzc}Scc}wcc}wr) isinstancerTensorrr)rYrrrrXs rapply_rot_embed_listrsu !U\\" C FGGG .q1G;;GG 9::1G c!fw..:: H ;s A$A)rc|jdd\}}|r||zt||zzS||zt||zzS)Nrr1)rrr)rYrrrrs rapply_rot_embed_catrsR yyB'GW 7{-a07::: 7{SVg---r pos_embed keep_indicespos_embed_has_batchc|rt|jdk\dnYt|jdk\d|jdfd|jzz}|jdj |}|jdfd|jdz zz|jddfz}|j |}t |j}d|d <|j |}|jd |S) a Apply keep indices to different ROPE shapes Expected pos_embed shapes: * [seq_len, pos_embed_dim] --> output [batch_size, seq_len, pos_embed_dim] * [num_heads, seq_len, pos_embed_dim] --> output [batch_size, num_heads, seq_len, pos_embed_dim] * [depth, num_heads, seq_len, pos_embed_dim] --> output [batch_size, depth, num_heads, seq_len, pos_embed_dim] And all of the above with leading batch dimension already present if `pos_embed_has_batch == True` rzIncorrect number of dimensionsrrr~)rrr1r})r rrr8rviewrEgather)rYrrr expand_shape keep_shape keep_expands rapply_keep_indices_nlcr,s  !#%EF  !#%EF }uy~~'== ''*11,? $$Q')DINNQ4F,GG>$A dll#!..E A  ,,E xxvUx33rc t||j|j|j|j|j |j |j|j|| \}}||fS)N r*r+rJr rrLrMrNrOrr) rr+rJr rrLrMrNrO)rmr*rrrrs rrz%RotaryEmbedding._get_pos_embed_valuessi1!LL((**nn..((,,  rc$|jyrzrqrss rr{z+RotaryEmbedding.init_non_persistent_buffersrtrc|j}||jk7rm|jJ|jJ|j||jj|jj \|_|_||_yyyNr2)r*rrrrrrmr*s rupdate_feat_shapez!RotaryEmbedding.update_feat_shapes ?? &:+H%%1 11%%1 11595O5O))00((..6P6 2D  2 )DO,I &rrc|O|jCt||j|j|j|j|j S|j $|j|j |jfSJd)NrLrMrNrOzQget_embed() requires pre-computed pos embeds or valid shape w/ pre-computed bands)rrrLrMrNrOrr)rmrs r get_embedzRotaryEmbedding.get_embeds  !7) ..#22 ,,"00     +0B0B0N%%t'9'99 9 mm m5rc^|j|jdd\}}t|||SNr)rrr)rmrYrrs rrzRotaryEmbedding.forwards.>>!''!"+6q'733r rGi'TFNNrHrINNrNNr)rrr__doc__boolrrrrUstrrgrlrrrrrrr{rrrrrs@rrrs!&.226!#!%-  - !c+- %T#Y/- - - ^.4 CGemm S   )DI )nxS 2n 4rrcZeZdZdZ ddedededededeeed eeed ed e ffd Z ddZ ddZ ddZ dejfdeefdZddZdeefdZddeeefdZ ddeeeefdeed eej,eej,ffdZdZxZS)RotaryEmbeddingCata Rotary position embedding w/ concatenatd sin & cos The following impl/resources were referenced for this impl: * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py * https://blog.eleuther.ai/rotary-embeddings/ Nr+rJr rLrr*rMrNrOc t|||_||_||_||_||_||_||_||_ | |_ |du|_ |8|dzf} |jdtj| | | dd|_nFd|_d} |D]}| |z} | |dzf}|jdtj|| | d|j#y) Nr/rr2Frdrrr)rfrgr+rJr rLrr*rMrNrOrrjrrkrrrlrs rrgzRotaryEmbeddingCat.__init__s  &"($,&*",4!7  !8+K  %++k&X]*^kp q!DNDJG 1   #'*I  ekk)FZ_.`mr s rr"c$|jyrprqrss rrlz#RotaryEmbeddingCat.reset_parametersKrtrc|js*|jj|jy|jj|j |j yrv)rrrxrrrr*rss rrrz RotaryEmbeddingCat._init_buffersOsG%% JJ  T002 3 NN !;!;DOO!L Mrc|jr9t|jdzt|j|j }n%t |jdz|jd}|j||Srrrs rrz!RotaryEmbeddingCat._compute_bandsVrrc t||j|j|j|j|j |j |j|j|| }tj|dS)Nrr1) rr+rJr rrLrMrNrOrr)rmr*rrembedss rrz(RotaryEmbeddingCat._get_pos_embed_valuesfsj'!LL((**nn..((,,  yy$$rc$|jyrzrqrss rr{z.RotaryEmbeddingCat.init_non_persistent_buffersvrtrc|jg||jk7rW|jJ|j||jj|jj|_||_yyyr)r*rrrrrs rrz$RotaryEmbeddingCat.update_feat_shapezsm ?? &:+H>>- --!77~~,,nn**8DN )DO,I &rrc|e|jYt||j|j|j|j|j }t j|dS|j |jSJd)Nrr1zPget_embed() requires pre-computed pos embed or valid shape w/ pre-computed bands) rrrLrMrNrOrrr)rmrrs rrzRotaryEmbeddingCat.get_embedsy  !7+ ..#22 ,,"00 F99VR( ( ^^ '>> ! ll l5rshapesseq_lenc|sgS|j tdtd|D}td|D}t||f|j|j|j |j |j\}}tj||gdj||d}|{tjt|||jdj|}t|D]/\} \} } | | z} |d| d| fj!| d|| d| f<1|S|D cgc]#\} } |d| d| fj!| | zd%} } } | Scc} } w)aGenerate ROPE embeddings for multiple grid shapes efficiently. Computes embeddings for the maximum grid size once, then extracts and flattens the relevant portions for each requested shape. Args: shapes: List of (H, W) tuples representing different grid sizes Returns: List of concatenated sin/cos embeddings for each shape, where each tensor has shape (H*W, dim) NzMBatch embedding generation requires cached bands, not pre-computed embeddingsc3&K|] \}}| ywr.0hws r z6RotaryEmbeddingCat.get_batch_embeds..)$!QA)c3&K|] \}}| ywrrrs rrz6RotaryEmbeddingCat.get_batch_embeds..rr)r*rrLrMrNrOr1r3)r RuntimeErrormaxrrLrMrNrOrrrzerosrDrtype_as enumerater)rmrrmax_hmax_wrr rope_embed_2d flat_embedsirrsrc_lenflat_embeds_lists rget_batch_embedsz#RotaryEmbeddingCat.get_batch_embedss"I :: no o)&)))&))2u~**nn..((,,   7G"4"=BB5%QST  ++c&k7M> U1XfEMMB U1XfEMMBLE5 ((5/ ! ! #C ((5/ ! ! #C 8Orfreqsrrc|j}|j}|jd|djdz}|jd|djdz}||z}tj|j dd}tj |j dd}tj||gd} | j|S)z&Compute mixed (learnable) frequencies.r1rr}rrr3) rrUr8rr9rr:rr&) rrrrfreqs_xfreqs_ycombinedrr rope_embedss rget_mixed_freqsrs KKE KKME}}R 58#5#5b#99G}}R 58#5#5b#99G Hii!33Ar:Gii!33Ar:G))Wg.B7K >>%  rc veZdZdZ d dedeeeefdedeffd Zde ede jfd Z xZ S) RotaryEmbeddingMRopeaaInterleaved multimodal RoPE (Qwen2-VL style) for vision, matching the reference GenLIP layout. Drop-in sibling of ``RotaryEmbeddingCat``: ``get_embed(shape) -> [N, 2*dim]``, consumed by ``apply_rot_embed_cat(..., half=True)`` (no new apply path / no separate sin/cos tensors). The ``dim // 2`` frequency channels are assigned to height/width/temporal axes in a strided ``T,H,W,T,H,W,...`` interleave (the reference ``apply_interleaved_mrope``): channels ``1,4,7,...`` -> height, ``2,5,8,...`` -> width, and the remainder -> temporal. ``mrope_section`` sets the strided extent per axis; the actual per-axis channel *counts* equal ``mrope_section`` only for the standard equal-section configs that tile exactly (``3*section == dim // 2``, e.g. ``(12,12,12)`` -> 36 channels = 12/12/12), and are otherwise the clamped interleave (e.g. ``dim=64, (8,12,12)`` -> 11/11/10) -- this matches the reference, which also clamps. For an image encoder there is no text, so the temporal channels sit at position 0 (inert) and this reduces to a 2-axis ``(h, w)`` rope -- numerically identical to a checkpoint trained with the reference MRoPE. Only ``grid_indexing='ij'`` is supported (GenLIP / NaFlex ``(y, x)`` row-major patch order); ``'xy'`` would require mirroring the timm axial shape-swap and is intentionally not implemented here. r+ mrope_sectionr rOct ||dzdk(sJdtd|Ds Jd||dk(sJd||_||_||_||_d|tjd|d| j|z zz }|jd |d |\}} } tj|dztj| } d| d| dzd<d| d| dzd<|jd| d y)Nrrzdim (head_dim) must be evenc3&K|] }|dk\ yw)rNr)rr<s rrz0RotaryEmbeddingMRope.__init__..9s1a161rz0mrope_section entries must be non-negative, got rIzXRotaryEmbeddingMRope supports grid_indexing='ij' only (GenLIP/NaFlex (y,x) patch order).rrinv_freqFrdrrraxis) rfrgallr+rr rOrr$rUrjrlong) rmr+rr rOrrr!_sec_tsec_hsec_wr"rns rrgzRotaryEmbeddingMRope.__init__,s1 Qw!|:::|1=11 O>}o N O1$ g f g$*&*+%,,q#q*P*V*V*X[^*^_` ZeD -u{{3!85::fEQuqy]Quqy] VTe!'!%!=!=!c3/!= !=  !=F9tCy9U\\9rrc @eZdZdZ ddededededeeedef fd Z dd Z dd Z deeefd Z deeefd Z ddZddeeed ej fdZ ddeeeefdeed eej eej ffdZdZdZxZS)RotaryEmbeddingMixeda~Rotary position embedding with depth-dependent learnable frequencies. This implementation supports mixed (learnable) ROPE. In mixed mode, each transformer block has its own set of learnable frequency parameters. Based on 'Rotary Position Embedding for Vision: https://arxiv.org/abs/2403.13298)' Compatible with original at https://github.com/naver-ai/rope-vit r+rrr r*rOc t |||_||_||_||_||_||_||z} | dzdk(s Jd| t| |||d||} tj| |_ |sd} |D]} | | z} |jdtj| || d |jd tj| || d |jydx|_|_y) aInitialize rotary embeddings. Args: dim: Embedding dimension (should be divisible by 4) depth: Number of transformer blocks num_heads: Number of attention heads temperature: Base for frequency computation feat_shape: Spatial dimensions [H, W] if known in advance grid_indexing: How to index grid positions ('xy' or 'ij') r/rz%head_dim must be divisible by 4, got T)r rrrNrrr2Frdr)rfrgr+rrr r*rOr r Parameterrrjrrkrrrr)rmr+rrr r*rOrrrrrr<rns rrgzRotaryEmbeddingMixed.__init__rs*  "&$*)#!|q T$I("TT $   # \\%(  !G 1     GFRW(Xej k   GFRW(Xej k    "& &DHtxrr"c|jU|j|j\}}|jj||jj|yyrv)r*_get_grid_valuesrrxr)rmrrs rrrz"RotaryEmbeddingMixed._init_bufferssI ?? &,,T__=HC HHNN3  HHNN3  'rc$|jyrprqrss rrlz%RotaryEmbeddingMixed.reset_parametersrtrcht||j|jj\}}||fS)NrOr)rrOrrrmr*rrs rr2z%RotaryEmbeddingMixed._get_grid_valuess4! ,,::$$ S Cxrc|j||jk7r|jJ|jJ|j|\}}|j |jj |jj |_|j |jj |jj |_||_yyyr)r*rrr2r&rrr6s rrz&RotaryEmbeddingMixed.update_feat_shapes ?? &:+H88' ''88' '',,Z8HCvvdhhootxx~~>DHvvdhhootxx~~>DH(DO ,I &rc$|jyrzrqrss rr{z0RotaryEmbeddingMixed.init_non_persistent_buffersrtrrc|0t||j|jj\}}n8|j%|j |j|j }}nJdt |j||S)zGenerate rotary embeddings for the given spatial shape. Args: shape: Spatial dimensions [H, W] Returns: Tensor of shape (depth, H*W, dim) containing concatenated sin/cos embeddings r5z8get_embed() requires pre-computed t_x/t_y or valid shape)rrOrrrrr)rmrrrs rrzRotaryEmbeddingMixed.get_embedsu  %"00zz((HC XX !dhh&:xxC TT T5tzz344rrrc |sgStd|D}td|D}t||g|j|jj\}}t |j||}|j \}} } } |j|| ||| } |t|} tj| || || |jj|jj}t|D]B\}\}}| ddddd|d|fj|| ||z| }||z}|||ddddd|f<D|Sg}|D]<\}}| ddddd|d|fj|| ||z| }|j|>|S)aiGenerate ROPE embeddings for multiple grid shapes efficiently. Computes embeddings for the maximum grid size once, then extracts and flattens the relevant portions for each requested shape. Args: shapes: List of (H, W) tuples representing different grid sizes seq_len: If provided, return padded tensor of this length. Otherwise return list. Returns: If seq_len is provided: Padded tensor of shape (len(shapes), depth, num_heads, seq_len, dim) Otherwise: List of tensors with shape (depth, num_heads, H*W, dim) for each shape c3&K|] \}}| ywrrrs rrz8RotaryEmbeddingMixed.get_batch_embeds..rrc3&K|] \}}| ywrrrs rrz8RotaryEmbeddingMixed.get_batch_embeds..rrr5Nr2)rrrOrrrrrrDrrrrrappend)rmrrrrrr max_embedrr_r+ max_embed_2drpaddedrrr embed_slice actual_lenresultss rrz%RotaryEmbeddingMixed.get_batch_embedss$I)&)))&))" EN,,::$$ S $DJJS9 $-?? y!S ~~eYucJ  F A[[E9gs4::K\K\dhdndndtdtuF&v. ; 6Aq*1a!RaR<8@@ STWXSXZ]^ U /:q!Q  +,  ; MG ,1*1a!RaR<8@@ STWXSXZ]^ {+ ,NrcV|j|jdd}t||Srrrs rrzRotaryEmbeddingMixed.forwardrrcdhS)z/Exclude frequency parameters from weight decay.rrrss rno_weight_decayz$RotaryEmbeddingMixed.no_weight_decays yr)$@NrQNNrr)rrrrrrUrrrrgrrrlr2rr{rrrrrrrrGrrs@rr.r.is"&.2!%5'5'5' 5'  5' !c+ 5'5'n 8DI+>)HT#Y,?)5xS 25ell52&*6sCx)6c]6 u||T%,,// 0 6p1 rr.separatecpuheightwidthnormalize_coordsctjd||tj|z}tjd||tj|z}|dk(rtt ||} | } | } nI|dk(rtt ||} | } | } n*|dk(rt|} t|} nt d||| z }|| z }|j|}|j|}|dk(r5tj||d\} } tj| | gd }n-tjtj||d d }|jd d }d|zdz }|S)z{Make coordinate grid matching offset and normalization of original. Returns: coords with shape (HW, 2) in [-1, 1]. g?r2rminrIzUnknown normalize_coords: rQrSr1r3rIrr@r) rr$rrUrrO ValueErrorr&rWr5r6)rKrLrMrOrNrrcoords_hcoords_wdenomh_denomw_denomgrid_wgrid_hcoordss rmake_coords_dinov3rZ!s]||CemmL{ZH||CvU]]KkYH5 c&%() U "c&%() Z '-,56F5GHII'!H'!H{{5!H{{5!H(TJff-26U^^HhNTVW ^^Aq !F 6\C F MrcpeZdZdZ d"dedeedeedeedeeededed ed e d eed eed eeffd Z d#dZ d#dZ de jfde jde j de j"fdZde j"de j"fdZde j"dee j"e j"ffdZ d$deede de j"fdZdeefdZdeefdZd#dZd%deeede j"fdZd e j"de j"fd!ZxZS)&RotaryEmbeddingDinoV3aRoPE for timm DinoV3 port, numerically matching original. Math is aligned to original DinoV3 RopePositionEmbedding at https://github.com/facebookresearch/dinov3: - 0.5-centered coords normalized by H/W (or min/max), mapped to [-1,1] - training-time augmentations (shift/jitter/rescale) - periods schedule equals Rope's temperature (base) or min/max period r+r min_period max_periodr*rMrNrO rotate_half shift_coords jitter_coordsrescale_coordsctt|||_| |_t ||_||_||_||_| |_ | |_ | |_ t|j|j|jfDcgc]}|duc}|_ ||_||_||_|dzf}|j#dt%j&|| |d|=|d|dz}||dzf}|j#d t%j&|| |dnd|_|j+ycc}w) Nr/periodsr2Frdrrrpos_embed_cached)rfrgr+r_rUr r]r^rMr`rarbany aug_activer*rNrOrjrrkrerl)rmr+r r]r^r*rMrNrOr_r`rarbrra periods_shaperrrns rrgzRotaryEmbeddingDinoV3.__init__\sG" &!-$$!1(*,t7H7H$J\J\^b^q^q6rsq}st%&*  Y M&X](^kpq  ! mjm3G #'*I  !3U[[SYaf5gty z$(D ! )ts D5r"c$|jyrprqrss rrlz&RotaryEmbeddingDinoV3.reset_parametersrtrc|jj|j|jF|j9|j |jd}|jj|yyy)rwNTno_aug)rdrx_compute_periodsr*re _create_embed)rm rope_embeds rrrz#RotaryEmbeddingDinoV3._init_bufferssh 40023 ?? &4+@+@+L++DOOD+IJ  ! ! ' ' 3,M &rrJrrc|jdz}|j^|jRtjdd|dtj }|j|j|jz |zz}n_|j tddtj|dtj z|jdzz }|j |z}|j||S) z5Construct periods from either min/max or temperature.r/rrrJr2z0Provide either min/max periods or `temperature`.rPr) r+r]r^rrrr rQr$r&)rmrrr+ exponentsrds rrnz&RotaryEmbeddingDinoV3._compute_periodsshh!m ?? &4??+Fq!SemmTIoo$//DOO*KPY)YZG' !STTell3uEMMRRVZV^V^bcVcdI&&)3Gzzuz55rrYc,|jr |js|S|j}|j}|jJt |j}t jd||j| |}||dddfz}|j}t |j}|dkr tdtj|}t jd||j| |j}||dddfz}|jvt |j} | dkr tdtj| } t jd||j| | j} || z}|S)z4Apply shift/jitter/rescale train time augmentations.Nrr2rzAjitter_coords must be > 0 (interpreted as multiplicative factor).zBrescale_coords must be > 0 (interpreted as multiplicative factor).r)trainingrgrrr`rUrrkuniform_rarQrrr'rb) rmrYrrshiftshift_hw jitter_factor jitter_max jitter_hwrescale_factor rescale_maxrescales r_apply_coord_augsz'RotaryEmbeddingDinoV3._apply_coord_augssq}}DOOM     ($++,E{{1V5AJJE6SXYHhtQw//F    )!$"4"45M! !dee-0J AfEBKKZKYcdhhjIia00F    *"4#6#67N" !eff((>2Kkk!F%@II;,XcdhhjGg%F rc&|jdz}|jj}|jj}|jj |k(sJ|dddddfj ||}dt jz|z|jddddfz }|jd}|jr|jd}n|jdd}tj|}tj|}||fS)zEReturn sin/cos embeddings with either 'half' or 'interleaved' layout.r/Nr2rrr1r3)r+rdrrrr&rrr6r_tilerrr9r:)rmrYr+rrr r9r:s r_get_pos_embed_from_coordsz0RotaryEmbeddingDinoV3._get_pos_embed_from_coordsshh!m$$ ""||!!#s***1d #&&fE&BTWWv% T4](CC"   [[^F--aR-8FiiiiCxrrmc|\}}t|||j|j|j}|s|j |}|j |\}}t j||gd}|S)N)rMrOrNr1r3)rZrMrOrNr~rrr) rmr*rmHWrYr9r:rps rroz#RotaryEmbeddingDinoV3._create_embedsz 1# q!22,,((   ++F3F226:SYYSzr2 rc`|j|d}|jd|d||_y)NTrlreFrd)rorjr*)rmr*rps r _cache_embedz"RotaryEmbeddingDinoV3._cache_embeds4'' 4'@  /N$rc`|j"||jk7r|j|yyyr)r*rrs rrz'RotaryEmbeddingDinoV3.update_feat_shapes. ?? &:+H   j ),I &rc$|jyrzrqrss rr{z1RotaryEmbeddingDinoV3.init_non_persistent_buffersrtrrc||j|}|S|jduxs|jxr |j}|r0|jJd|j|j}|S|jJ|j}|S)zGenerate rope_embed matching DINOv3 RopePositionEmbedding numerics. Returns: (HW, num_heads, 2 * head_dim) with last dim = [sin, cos] cat. Nz&feature shape must be cached on create)rorertrgr*)rmrrp need_creates rrzRotaryEmbeddingDinoV3.get_embed s  ++E2J//47^DMM<]dooK2\4\\2!//@ ,,888!22 rrYcn|j|jdd}t|||jS)z$Get and apply rotary embeddings to xrN)r)rrrr_rs rrzRotaryEmbeddingDinoV3.forwards1NN17712;/ "1id6F6FGGr) gY@NNNrIrHrITNNNNNrFr)rrrrrrrUrrrrgrlrrrrrrrrnr~rrrorrr{rrrrs@rr\r\Ss,1*.*..2$.!$!% $,0-1.25 5 "%5 ! 5 ! 5 !c+ 5 "5 5 5 5 #5/5 $E?5 %UO5 n473)#3>v>> e  =$'!AcY&6A&AA g  ;% #T*#KyKFKK h  ;% #T*$D )9DVDD g ? A JJq$  #Cy(8CFCC.yk:;;r)gl@TN)rrNr)rHT)rir,)1rrtypingrrrrrr_fxr r=r trace_utilsr rrUrrrrr(rrrArFrr`Modulerbrrrrrrrrrr r wraprrrr.rZr\rrrrrs // * !)-    & $)- &   \\ ###()-"]]&#I&# &#&# &# ! &# & &#{{&# \\&#R,tCy,T#Y,)-#"".2!)-"]]RIR %RR R  R  RRR!c+RRR&R{{R %,,Rj62996rJ ( (( . <<... .  \\ .. ;  ;;; ;  %,, ;0. <<. \\.. \\ ..%* ". <<".<<".ll"." ".  \\ ".N)-#".2!)-"]]5I5 %55 5  5  55!c+555&5{{5pJ4biiJ4Zx1x1~" (mm(((( (  ( \\(D")-"]] Cy&{{   5<< %& $!||! \\!\\! \\ !$M9299M9`u299up!+!$"]]---- -  -  -{{- \\--`MHBIIMHb*<*< *<*< YY *