RЦiSrSSKrSSKJrJrJrJrJrJrJ r J r SSK r SSK J r SSKJ s Jr SSKJrJr SSKJrJrJrJrJrJr SSKJr SSKJr SS KJ r SS K!J"r"J#r# S /r$"S S \ RJ5r&"SS\ RJ5r'"SS\ RJ5r("SS\ RJ5r)"SS\ RJ5r*"SS\ RJ5r+S\,S\S\ RJ4Sjr-SYS\\.S4S\/S\.S\\.\.\.\.44S jjr0"S!S"\ RJ5r1"S#S$\ RJ5r2SZS%\S&\.S'\.S(\\.S)\.S*\.S+\.S,\.S-\/S.\3S/\/S0\/S\S\ Rh4S1jjr5S[S%\S&\.S'\.S(\\.S)\.S-\/S.\3S/\/S0\/S\S\ Rh4S2jjr6"S3S \ RJ5r7S\S4\,S5\3S\S\74S6jjr8S]S7\,S\S\\,\44S8jjr9\#"\9"S9S:S;S<9\9"S9S=S>S?S@9\9"S9SAS;S<9\9"S9SBS>S?S@9\9"S9SCS;S<9\9"S9SDS>SES@9\9"S9SFS;S<9\9"S9SGSHSES@9\9"S9SIS;S<9\9"S9SJSHSES@9\9"S9SKSHSLS@9SM. 5r:\"S\S5\3S\S\74SNjj5r;\"S\S5\3S\S\74SOjj5r<\"S\S5\3S\S\74SPjj5r=\"S\S5\3S\S\74SQjj5r>\"S\S5\3S\S\74SRjj5r?\"S\S5\3S\S\74SSjj5r@\"S\S5\3S\S\74STjj5rA\"S\S5\3S\S\74SUjj5rB\"S\S5\3S\S\74SVjj5rC\"S\S5\3S\S\74SWjj5rD\"S\S5\3S\S\74SXjj5rEg)^a4Vision OutLOoker (VOLO) implementation Paper: `VOLO: Vision Outlooker for Visual Recognition` - https://arxiv.org/abs/2106.13112 Code adapted from official impl at https://github.com/sail-sg/volo, original copyright in comment below Modifications and additions for timm by / Copyright 2022, Ross Wightman N)AnyCallableDictListOptionalTupleUnionTypeIMAGENET_DEFAULT_MEANIMAGENET_DEFAULT_STD)DropPathMlp to_2tuple to_ntuple trunc_normal_use_fused_attn)build_model_with_cfg)feature_take_indices) checkpoint)register_modelgenerate_default_cfgsVOLOc^\rSrSrSrSS\S\S\S\S\S\S \S \4U4S jjjrS \ RS \ R4Sjr Sr U=r $)OutlookAttention'z,Outlook attention mechanism for VOLO models.dim num_heads kernel_sizepaddingstrideqkv_bias attn_drop proj_dropc >XS.n [T U]5 X-n X lX0lX@lXPlU S-Ul[R"X4SU0U D6Ul [R"XS-U-40U D6Ul [R"U5Ul [R"X40U D6Ul [R"U5Ul[R"X4US9Ul[R""XUSS9Ulg) aInitialize OutlookAttention. Args: dim: Input feature dimension. num_heads: Number of attention heads. kernel_size: Kernel size for attention computation. padding: Padding for attention computation. stride: Stride for attention computation. qkv_bias: Whether to use bias in linear layers. attn_drop: Attention dropout rate. proj_drop: Projection dropout rate. devicedtypebias)r r!r"T)r r" ceil_modeN)super__init__rr r!r"scalennLinearvattnDropoutr$projr%Unfoldunfold AvgPool2dpool)selfrrr r!r"r#r$r%r(r)ddhead_dim __class__s O/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/timm/models/volo.pyr/OutlookAttention.__init__*s2 / #"&  % 39(9b9IIc!#3i#?F2F I.IIc-"- I.iiKQWX LLVdS xreturncURup#pEURU5RSSSS5n[R"X0R - 5[R"X@R - 5pUR U5RX RXPR-URUR-Xx-5RSSSSS5nURURSSSS55RSSSS5n URU 5RX'U-URURUR-URUR-5RSSSSS5n XR-n U RSS9n URU 5n X-RSSSSS5RX%UR-UR-Xx-5n[R "XU4URUR"UR S9nUR%URSSSS55nUR'U5nU$) sForward pass. Args: x: Input tensor of shape (B, H, W, C). Returns: Output tensor of shape (B, H, W, C). rrr,r) output_sizer r!r")shaper3permutemathceilr"r8reshaperr r:r4r0softmaxr$Ffoldr!r6r%) r;rBBHWCr3hwr4s r?forwardOutlookAttention.forwardVsWW a FF1I  aAq )yy[[)499Q_+E1 KKN " " ~~qNN2   t// / 88?1aA8N yy1aA./771aCyy&& 1udnnd&6&69I9I&I   t// /118Aq!Q1G jj |||#~~d# X  q!Q1 - 5 5aT=M=M9MPTP`P`9`bcbg h FF1a&d6F6FPTP\P\eiepep q IIaii1a+ , NN1 rA) r4r$r rr!r:r6r%r0r"r8r3)rFrrFr[NN)__name__ __module__ __qualname____firstlineno____doc__intboolfloatr/torchTensorrY__static_attributes__ __classcell__r>s@r?rr's6 !"!!*T*T*T *T  *T  *T*T*T*T*TX  %,,  rArc ^\rSrSrSrSSSSS\R \RSSS4 S\S \S \S \S \S \ S\ S\ S\ \RS\ \RS\ 4U4Sjjjr S\RS\R4SjrSrU=r$) Outlookeryz9Outlooker block that combines outlook attention with MLP.r@r[FNrr r!r"r mlp_ratior$ drop_path act_layer norm_layerr#c >XS.n[TU]5 U "U40UD6Ul[UU4UUUU US.UD6UlUS:a [ U5O[ R"5UlU "U40UD6Ul [SU[X-5U S.UD6Ul US:a[ U5Ul g[ R"5Ul g)aInitialize Outlooker block. Args: dim: Input feature dimension. kernel_size: Kernel size for outlook attention. padding: Padding for outlook attention. stride: Stride for outlook attention. num_heads: Number of attention heads. mlp_ratio: Ratio for MLP hidden dimension. attn_drop: Attention dropout rate. drop_path: Stochastic depth drop rate. act_layer: Activation layer type. norm_layer: Normalization layer type. qkv_bias: Whether to use bias in linear layers. r')r r!r"r#r$r[ in_featureshidden_featuresroN)r.r/norm1rr4rr1Identity drop_path1norm2rramlp drop_path2)r;rr r!r"rrmr$rnrorpr#r(r)r<r>s r?r/Outlooker.__init__|s> / *r* $    $     2;R(9-R[[]*r*  0    2;R(9-R[[]rArBrCcXRURURU555-nXRUR UR U555-nU$zGForward pass. Args: x: Input tensor. Returns: Output tensor. rxr4rvr{rzryr;rBs r?rYOutlooker.forwardM  $**Q- 89 9 A 78 8rAr4rxr{rzrvry)r\r]r^r_r`r1GELU LayerNormrarcr Modulerbr/rdrerYrfrgrhs@r?rjrjysC!!!)+*,,,"5S5S5S 5S  5S  5S5S5S5SBII5SRYY5S5S5Sn  %,,  rArjc ^\rSrSr%Sr\R R\\ S'SS\ S\ S\S\ S\ 4 U4S jjjr S \RS \R4S jrS rU=r$) Attentionz!Multi-head self-attention module. fused_attnrrr#r$r%cV>XgS.n[T U]5 X lX-n U S-Ul[ 5Ul[ R"XS-4SU0UD6Ul[ R"U5Ul [ R"X40UD6Ul [ R"U5Ul g)zInitialize Attention module. Args: dim: Input feature dimension. num_heads: Number of attention heads. qkv_bias: Whether to use bias in QKV projection. attn_drop: Attention dropout rate. proj_drop: Projection dropout rate. r'r*rFr+N) r.r/rr0rrr1r2qkvr5r$r6r%) r;rrr#r$r%r(r)r<r=r>s r?r/Attention.__init__s& / "#% (*99S'??B?I.IIc-"- I.rArBrCcURup#pEURU5RX#U-SURXPR-5R SSSSS5nUR S5upxn UR (a>[R"XxU UR(aURROSS9nOFXpR-nXxRSS 5-n U RS S 9n URU 5n X-nURSS5RX#XE5nURU5nUR!U5nU$) rErFrGrrr,r[) dropout_prHrI)rKrrOrrLunbindrrQscaled_dot_product_attentiontrainingr$pr0 transposerPr6r%) r;rBrSrTrUrVrqkr3r4s r?rYAttention.forwardsWW ahhqk!!!UAt~~qNN?RS[[\]_`bcefhij**Q-a ??..a.2mm$..**A JJA{{2r**D<>$'DA KK1  % %aA 1 IIaL NN1 rA)r$rrr6r%rr0)Fr[r[NN)r\r]r^r_r`rdjitFinalrb__annotations__rarcr/rerYrfrgrhs@r?rrs+ %% "!!/// /  /  //>%,,rArc^\rSrSrSrSSSS\R \RSS4S\S\S \ S \ S \ S \ S \ \RS\ \R4U4Sjjjr S\RS\R4SjrSrU=r$) Transformeriz9Transformer block with multi-head self-attention and MLP.@Fr[Nrrrmr#r$rnrorpc >XS.n [T U]5 U"U40U D6Ul[U4X$US.U D6UlUS:a [ U5O[ R"5UlU"U40U D6Ul [SU[X-5US.U D6Ul US:a[ U5Ul g[ R"5Ul g)awInitialize Transformer block. Args: dim: Input feature dimension. num_heads: Number of attention heads. mlp_ratio: Ratio for MLP hidden dimension. qkv_bias: Whether to use bias in QKV projection. attn_drop: Attention dropout rate. drop_path: Stochastic depth drop rate. act_layer: Activation layer type. norm_layer: Normalization layer type. r')rr#r$r[rrNru)r.r/rvrr4rr1rwrxryrrarzr{) r;rrrmr#r$rnrorpr(r)r<r>s r?r/Transformer.__init__s2 / *r* ceYU^ebde 1:R(9-R[[]*r* h3CO8LXahegh1:R(9-R[[]rArBrCcXRURURU555-nXRUR UR U555-nU$r~rrs r?rYTransformer.forward+rrAr)r\r]r^r_r`r1rrrarcrbr rr/rdrerYrfrgrhs@r?rrsC ""!!)+*,,,!S!S!S !S  !S  !S!SBII!SRYY!S!SF  %,,  rArc ^\rSrSrSrSS\S\S\\S\S\S\4 U4S jjjr S \ RS \ R4S jr S r U=r$)ClassAttentioni9z6Class attention mechanism for class token interaction.rrr=r#r$r%c  >XxS.n [T U]5 X lUbX0lO X-nX0lUS-Ul[ R "XRUR-S-4SU0U D6Ul[ R "XRUR-4SU0U D6Ul[ R"U5Ul [ R "URUR-U40U D6Ul [ R"U5Ul g)a;Initialize ClassAttention. Args: dim: Input feature dimension. num_heads: Number of attention heads. head_dim: Dimension per head. If None, computed as dim // num_heads. qkv_bias: Whether to use bias in QKV projection. attn_drop: Attention dropout rate. proj_drop: Projection dropout rate. r'Nr*rGr+) r.r/rr=r0r1r2kvrr5r$r6r%) r;rrr=r#r$r%r(r)r<r>s r?r/ClassAttention.__init__<s* / "  $M'H$M% ))C!?!!CY(YVXY3  >TXTQSTI.IIdmmdnn>1dmm Ltzz Y;;r2&&|||#~~d#X((A.66q!T]]T^^=[\ IIi( NN9- rA)r$r=rrr6r%rr0)rNFr[r[NN)r\r]r^r_r`rarrbrcr/rdrerYrfrgrhs@r?rr9s@ &*"!!#/#/#/sm #/  #/  #/#/#/J%,,rArc ^\rSrSrSrSSSSSS\R \RSS4 S\S\S \ \S \ S \ S \ S \ S\ S\ \RS\ \R4U4SjjjrS\R S\R 4SjrSrU=r$) ClassBlockizz3Class block that combines class attention with MLP.NrFr[rrr=rmr#dropr$rnrorpc >XS.n [TU]5 U "U40U D6Ul[U4UUUUUS.U D6UlUS:a [ U5O[ R"5UlU "U40U D6Ul [SU[X-5U US.U D6Ul US:a[ U5Ul g[ R"5Ul g)aInitialize ClassBlock. Args: dim: Input feature dimension. num_heads: Number of attention heads. head_dim: Dimension per head. If None, computed as dim // num_heads. mlp_ratio: Ratio for MLP hidden dimension. qkv_bias: Whether to use bias in QKV projection. drop: Dropout rate. attn_drop: Attention dropout rate. drop_path: Stochastic depth drop rate. act_layer: Activation layer type. norm_layer: Normalization layer type. r')rr=r#r$r%r[)rsrtrorNru)r.r/rvrr4rr1rwrxryrrarzr{)r;rrr=rmr#rr$rnrorpr(r)r<r>s r?r/ClassBlock.__init__}s: / *r* "      2;R(9-R[[]*r*  0    2;R(9-R[[]rArBrCcUSS2SS24nX RURURU555-nX RUR UR U555-n[ R"X!SS2SS24/SS9$)zForward pass. Args: x: Input tensor of shape (B, N, C) where first token is class token. Returns: Output tensor with updated class token. NrrI)rxr4rvr{rzryrdcat)r;rBrs r?rYClassBlock.forwardsxa!eH  $**Q-0H II I9N0O PP yy)q!"uX.A66rAr)r\r]r^r_r`r1rrrarrcrbr rr/rdrerYrfrgrhs@r?rrzs= '+!"!!)+*,,,3S3S3Ssm 3S  3S  3S3S3S3SBII3SRYY3S3Sj 7 7%,, 7 7rAr block_typekwargsrCc 8US:Xa [S0UD6$SU35e)zGet block based on type. Args: block_type: Type of block ('ca' for ClassBlock). **kwargs: Additional keyword arguments for block. Returns: The requested block module. cazInvalid block type: ru)r)rrs r? get_blockrs,T#F##9,ZL99urAsize.lamr0cUSU-nUSU-n[R"U[RS9n[R"U[RS9n[R"SU- 5nXW-R 5nXg-R 5n [R "SUS5n [R "SUS5n [R "XS-- SU5n [R "XS-- SU5n [R "XS--SU5n[R "XS--SU5nU R5U R5UR5UR54$)aGet random bounding box for token labeling. Reference: https://github.com/zihangJiang/TokenLabeling Args: size: Input tensor size tuple. lam: Lambda parameter for cutmix. scale: Scaling factor. Returns: Bounding box coordinates (bbx1, bby1, bbx2, bby2). rrG)r)?rr)rdtensorfloat32sqrtrarandintclampitem)rrr0rUrTW_tH_tcut_ratcut_wcut_hcxcybbx1bby1bbx2bby2s r? rand_bboxrs' Q5A Q5A ,,q .C ,,q .Cjjc"G ]   !E ]   !E q!T "B q!T "B ;;rQJ1 -D ;;rQJ1 -D ;;rQJ1 -D ;;rQJ1 -D 99; TYY[$))+ ==rAc^\rSrSrSrSS\S\S\S\S\S\S \4U4S jjjrS \RS \R4S jr Sr U=r $) PatchEmbediz6Image to patch embedding with multi-layer convolution.img_size stem_conv stem_stride patch_sizein_chans hidden_dim embed_dimc >XS.n [T U]5 US;deU(a[R"[R"XV4SUSSS.U D6[R "U40U D6[R "SS9[R"Xf4SS S SS.U D6[R "U40U D6[R "SS9[R"Xf4SS S SS.U D6[R "U40U D6[R "SS95 UlOS Ul[R"UU4XC-XC-S .U D6UlX-X--Ul g ) aInitialize PatchEmbed. Different from ViT which uses 1 conv layer, VOLO uses multiple conv layers for patch embedding. Args: img_size: Input image size. stem_conv: Whether to use stem convolution layers. stem_stride: Stride for stem convolution. patch_size: Patch size (must be 4, 8, or 16). in_chans: Number of input channels. hidden_dim: Hidden dimension for stem convolution. embed_dim: Output embedding dimension. r')r,rrFF)r r"r!r+T)inplacerNr r") r.r/r1 SequentialConv2d BatchNorm2dReLUconvr6 num_patches) r;rrrrrrrr(r)r<r>s r?r/PatchEmbed.__init__s>2 / Z'''   (oAk[\cholnoz0R0% *gaST[`gdfgz0R0% *gaST[`gdfgz0R0% DIDIII   #1,     %2x7MNrArBrCcdURbURU5nURU5nU$)z}Forward pass. Args: x: Input tensor of shape (B, C, H, W). Returns: Output tensor of shape (B, embed_dim, H', W'). )rr6rs r?rYPatchEmbed.forward)s- 99  ! A IIaLrA)rrr6) FrrrF@NN) r\r]r^r_r`rarbr/rdrerYrfrgrhs@r?rrs@ #   2O2O2O 2O  2O  2O2O2O2Oh  %,,  rArc^\rSrSrSrS S\S\S\4U4SjjjrS\RS\R4S jr S r U=r $) Downsamplei8z#Downsampling module between stages. in_embed_dim out_embed_dimrcf>[TU]5 XES.n[R"X4X3S.UD6Ulg)zInitialize Downsample. Args: in_embed_dim: Input embedding dimension. out_embed_dim: Output embedding dimension. patch_size: Patch size for downsampling. r'rN)r.r/r1rr6)r;rrrr(r)r<r>s r?r/Downsample.__init__;s3  /IIlkzkhjk rArBrCcxURSSSS5nURU5nURSSSS5nU$)zvForward pass. Args: x: Input tensor of shape (B, H, W, C). Returns: Output tensor of shape (B, H', W', C'). rrFrrG)rLr6rs r?rYDownsample.forwardNs> IIaAq ! IIaL IIaAq !rA)r6)rGNN) r\r]r^r_r`rar/rdrerYrfrgrhs@r?rr8s\-   lll ll&  %,,  rArblock_fnindexrlayersrr r!r"rmr#r$drop_path_ratec /n[X15HInU U[USU5--[U5S- - nURU"U4UUUUUU U UU U S. UD65 MK [R"U6nU$)aGenerate outlooker layers for stage 1. Args: block_fn: Block function to use (typically Outlooker). index: Index of current stage. dim: Feature dimension. layers: List of layer counts for each stage. num_heads: Number of attention heads. kernel_size: Kernel size for outlook attention. padding: Padding for outlook attention. stride: Stride for outlook attention. mlp_ratio: Ratio for MLP hidden dimension. qkv_bias: Whether to use bias in QKV projection. attn_drop: Attention dropout rate. drop_path_rate: Stochastic depth drop rate. **kwargs: Additional keyword arguments. Returns: Sequential module containing outlooker blocks. Nr) r r!r"rrmr#r$rnr(r)rangesumappendr1r)rrrrrr r!r"rmr#r$rr(r)rblocks block_idx block_dprs r?outlooker_blocksr]sJF6=) "i#fVen2E&EF#f+XY/Z  h   #    * ]]F #F MrAc /n [X15HCn X[USU5--[U5S- - n U RU"UU4UUUU S.U D65 ME [R"U 6n U $)a-Generate transformer layers for stage 2. Args: block_fn: Block function to use (typically Transformer). index: Index of current stage. dim: Feature dimension. layers: List of layer counts for each stage. num_heads: Number of attention heads. mlp_ratio: Ratio for MLP hidden dimension. qkv_bias: Whether to use bias in QKV projection. attn_drop: Attention dropout rate. drop_path_rate: Stochastic depth drop rate. **kwargs: Additional keyword arguments. Returns: Sequential module containing transformer blocks. Nr)rmr#r$rnr) rrrrrrmr#r$rrrrrs r?transformer_blocksrs:F6=) "#fVen2E&EF#f+XY/Z  h        *]]F #F MrAc-B^\rSrSrSrSSSSSSS S S S S S S S S S \R SSS SS S 4S\\S\S\S\S\ S\S\S\ \\S\ \\S\ \ S4S\ \ S4S\ S\ S\ S \ S!\ S"\ S#\\RS$\ \ \ S4S%\ S&\ S'\4,U4S(jjjrS)\RS*S 4S+jr\R&R(S*\4S,j5r\R&R(SES-\ S*\\ \44S.jj5r\R&R(SFS/\ S*S 4S0jj5r\R&R(S*\R4S1j5rSGS\S\ \ S*S 4S2jjrS3\R:S*\R:4S4jrS3\R:S*\R:4S5jrS3\R:S*\ \R:\ \R:\R:\ \\\\4444S6jr!SHS3\R:S7\ \ \\\4S8\ S9\ S:\ S;\ S*\ \\R:\ \R:\\R:444S<jjr"SIS7\ \\\4S=\ S>\ S*\\4S?jjr#S3\R:S*\R:4S@jr$SES3\R:SA\ S*\R:4SBjjr%S3\R:S*\R:4SCjr&SDr'U=r($)JrizVision Outlooker (VOLO) model.rrFtokenrrN)TFFFrlFr[)rrTrGrrr num_classes global_poolrstem_hidden_dim embed_dimsr downsamples.outlook_attentionrmr# drop_rate pos_drop_rateattn_drop_raterrp post_layers use_aux_head use_mix_token pooling_scalec>[T!U]5 UUS.n[U5n[U5"U 5n [ U5nX@lX0lXPlUUlUUl US=Ul Ul U(aSUl US:XdS5eSUl [SSSUUUUS S .UD6UlUnUS U-U-US U-U-4n[ R""[$R&"S US US US40UD65Ul[ R*"US 9Ul/Ul/Ul/nS n[3[U55HnU U(a#[5[6UUUUU U4U UU UUS .UD6n O#[9[:UUUUU U4U UU UUUS.UD6n UR=U 5 UR.R=U5 UR0R=[?UUUSU3S95 US - nU U(dMUR=[AUUUUS -S40UD65 US-nUS - nM [ RB"U5Ul"SUl#Ub[ RB"[3[U55Vs/sH#n[IUU4USU SU SU USUS.UD6PM% sn5Ul#[ R""[$R&"S S US40UD65Ul%[MURJSS9 U(aCUS :a"[ RN"URU40UD6O[ RP"5Ul)OSUl)U"UR40UD6Ul*[ R*"U5Ul+US :a"[ RN"URU40UD6O[ RP"5Ul,[MUR(SS9 UR[UR\5 gs snf)aoInitialize VOLO model. Args: layers: Number of blocks in each stage. img_size: Input image size. in_chans: Number of input channels. num_classes: Number of classes for classification. global_pool: Global pooling type ('token', 'avg', or ''). patch_size: Patch size for patch embedding. stem_hidden_dim: Hidden dimension for stem convolution. embed_dims: List of embedding dimensions for each stage. num_heads: List of number of attention heads for each stage. downsamples: Whether to downsample between stages. outlook_attention: Whether to use outlook attention in each stage. mlp_ratio: Ratio for MLP hidden dimension. qkv_bias: Whether to use bias in QKV projection. drop_rate: Dropout rate. pos_drop_rate: Position embedding dropout rate. attn_drop_rate: Attention dropout rate. drop_path_rate: Stochastic depth drop rate. norm_layer: Normalization layer type. post_layers: Post-processing layer types. use_aux_head: Whether to use auxiliary head. use_mix_token: Whether to use token mixing for training. pooling_scale: Pooling scale factor. r'rHrrz)return all tokens if mix_token is enabledFTrGr)rrrrrrr)r)rmr#r$rp)rmr#rr$rpznetwork.)num_chs reductionmoduleNr[)rrrmr#r$rnrp{Gz?stdru)/r.r/lenrrrrr mix_tokenr num_featureshead_hidden_sizebetagrad_checkpointingr patch_embedr1 Parameterrdzeros pos_embedr5pos_drop stage_ends feature_inforrrjrrrdictr ModuleListnetwork post_networkr cls_tokenrr2rwaux_headnorm head_dropheadapply _init_weights)"r;rrrrrrrrrr r rmr#r r r rrprrrrr(r)r< num_layersr patch_gridr(ristager>s" r?r/ VOLO.__init__s j  /[ j))4 X&& &&*4>rNBD1 DI') V+V V)"'% !& m    qkZ/=@(1+Q[B[_lBlm ekk!Z]JqMS]^`Sa&heg&hi ]3  s6{#A #(qMaL  (l%,)  +qMaL  (l%#1,)   NN5 ! OO " "9 -    $ $T*Q-1W_`i_jUk%l m NI1~~z*Q-AE9JATQSTUQQ M$P}}W- !  " " s;/0 /1AN "2'm'm%, )  1 / !D  \\%++aJrN*Qb*QRDN $..c 2 OZ]^BIId&7&7KKdfdododqDM DMt007B7 I.GRUVBIId//CC\^\g\g\i dnn#. 4%%&; /s5*O*mrCc[U[R5(am[URSS9 [U[R5(a9UR b+[R RUR S5 gggg)zDInitialize weights for modules. Args: m: Module to initialize. rrNr) isinstancer1r2rweightr+init constant_)r;r7s r?r0VOLO._init_weightsnsb a # # !(( ,!RYY''AFF,>!!!&&!,-?' $rAc SS1$)z_Get set of parameters that should not have weight decay. Returns: Set of parameter names. r"r*rur;s r?no_weight_decayVOLO.no_weight_decayys[))rAcoarsec [SSS//SQS9$)zGet parameter grouping for optimizer. Args: coarse: Whether to use coarse grouping. Returns: Parameter grouping dictionary. z ^cls_token|pos_embed|patch_embed)z^network\.(\d+)\.(\d+)N)z^network\.(\d+)r))z ^cls_tokenrD)z^post_network\.(\d+)N)z^norm)i)stemrblocks2)r&)r;rBs r? group_matcherVOLO.group_matchers$41*  rAenablecXlg)zYSet gradient checkpointing. Args: enable: Whether to enable gradient checkpointing. N)r)r;rIs r?set_grad_checkpointingVOLO.set_grad_checkpointings #)rAcUR$)zAGet classifier module. Returns: The classifier head module. )r.r?s r?get_classifierVOLO.get_classifiersyyrAc2XlUbX l[URS5(a URRR OSn[URS5(a URRR OSnUS:a [R"URXUS9O[R"5UlURbAUS:a [R"URXUS9O[R"5Ul gg)z{Reset classifier head. Args: num_classes: Number of classes for new classifier. global_pool: Global pooling type. Nr:rr') rrhasattrr.r:r(r)r1r2rrwr+)r;rrr(r)s r?reset_classifierVOLO.reset_classifiers'  "* ,3DIIx,H,H!!((d*1$))X*F*F   &&DKVYZ?II   {H`b`k`k`m == $OZ]^II!!;ULdfdododq M %rArBcb[UR5Hsup#US:XaXR-nURU5nUR(a0[ R R5(d [X15nMkU"U5nMu URupEpgURUSU5nU$)zForward pass through token processing stages. Args: x: Input tensor of shape (B, H, W, C). Returns: Token tensor of shape (B, N, C). rGrH) enumerater(r"r#rrdr is_scriptingrrKrO)r;rBidxblockrSrTrUrVs r?forward_tokensVOLO.forward_tokenss$DLL1JCax&MM!$&&uyy/E/E/G/Gu(!H2WW a IIaQ rAcBURup#nURRUSS5n[R"XQ/SS9nUR HLnUR (a0[RR5(d [Xa5nMDU"U5nMN U$)zForward pass through class attention blocks. Args: x: Input token tensor of shape (B, N, C). Returns: Output tensor with class token of shape (B, N+1, C). rHrrI) rKr*expandrdrr)rrrVr)r;rBrSrrV cls_tokensrXs r? forward_clsVOLO.forward_clss''a^^**1b"5 IIzo1 -&&E&&uyy/E/E/G/Gu(!H ' rAcURU5nURSSSS5nUR(Ga(UR(Ga[R R URUR5R5nURSUR-URSUR-pC[UR5X RS9upVpxUR5n URU-URU-pURU-URU-pURS5SS2X2X2SS24U SS2X2X2SS24'U nOSupVpxURU5nUR bUR#U5nUR%U5nUR&S:XaUR)SS 9nOUR&S :Xa USS2S4nOUnUR*cU$UR+USS2SS245nUR(dUS UR-S5S--$UR(aUR(aUR/URSWWURS 5nUR5n URS5SS2XW2Xh2SS24U SS2XW2Xh2SS24'U nUR/URSX4-URS 5nXXVXx44$) zForward pass for training with mix token support. Args: x: Input tensor of shape (B, C, H, W). Returns: If training with mix_token: tuple of (class_token, aux_tokens, bbox). Otherwise: class_token tensor. rrGrFr)r0N)rrrravgrIr?rH)rrLrrrd distributionsBetarsamplerKrrrclonefliprYr)r^r,rmeanr+maxrO)r;rBrpatch_hpatch_wrrrrtemp_xsbbx1sbby1sbbx2sbby2x_clsx_auxs r? forward_trainVOLO.forward_trains    Q  IIaAq ! >>>dmmm%%**499dii@GGIC wwqzT-?-??tOaOaAaW%.qvvxDVDV%W "DWWYF--4d6H6H46O5--4d6H6H46O556VVAYq%+u{\]?]5^F1ek5;1 2A%/ "D    "    (  #A IIaL   u $FFqFME    (adGEE == L a12h'}}31a00 0 >>dmmMM%++a.'7EKKPROTE[[]F16Aq$)TYXY?Y1ZF1diA- .EMM%++a.'2CU[[QS_UEd$555rAindicesr, stop_early output_fmtintermediates_onlyc US;dS5e/n[[UR5U5upUV s/sHoRU PM nn URU n URuppUR U5R SSSS5n[ RR5(dU(d URnOURSU S-n[U5HunnUS:XaXR-nURU5nUR(a0[ RR5(d [UU5nOU"U5nUU;dM|U(aUS:aURU5nOUnUR!UR SSSS55 M U(aU$URun nnnUR#U SU5nUR$bUR'U5nURU5nX4$s sn f) aForward features that returns intermediates. Args: x: Input image tensor indices: Take last n blocks if int, all if None, select matching indices if sequence norm: Apply norm layer to all intermediates stop_early: Stop iterating over blocks when last desired intermediate hit output_fmt: Shape of intermediate feature outputs intermediates_only: Only return intermediate features Returns: )NCHWzOutput format must be NCHW.rrGrFrNrH)rrr$rKrrLrdrrVr(rUr"r#rrr,rrOr)r^)r;rBrur,rvrwrx intermediates take_indices max_indexr4rS_heightwidthr(rWrXx_interrTrUrVs r?forward_intermediatesVOLO.forward_intermediates$s*Y&E(EE& "6s4??7KW"U 4@ALq*L AOOI.  ggf   Q  ' '1a 3 99 ! ! # #:llGll>IM2G#G,JCax&MM!$&&uyy/E/E/G/Gua(!Hl"C1H"iilGG$$W__Q1a%@A-  WW 1a IIaQ     (  #A IIaLOBsH prune_norm prune_headc>[[UR5U5upEURUnURSUS-UlU(a[R "5UlU(a,[R"5UlURSS5 U$)aPrune layers not required for specified intermediates. Args: indices: Indices of intermediate layers to keep. prune_norm: Whether to prune normalization layer. prune_head: Whether to prune classification head. Returns: List of kept intermediate indices. Nrr) rrr$r(r1rwr,r'r)rR)r;rurrr|r}s r?prune_intermediate_layersVOLO.prune_intermediate_layersesz  "6s4??7KW"U OOI. ||NY]3  DI  " D   ! !!R (rAcURU5RSSSS5nURU5nURbUR U5nUR U5nU$)zyForward pass through feature extraction. Args: x: Input tensor of shape (B, C, H, W). Returns: Feature tensor. rrGrFr)rrLrYr)r^r,rs r?forward_featuresVOLO.forward_featuressg   Q  ' '1a 3    "    (  #A IIaLrA pre_logitscVURS:XaURSS9nOURS:Xa USS2S4nOUnURU5nU(aU$URU5nURb4UR USS2SS245nUSUR S5S--nU$)zForward pass through classification head. Args: x: Input feature tensor. pre_logits: Whether to return pre-logits features. Returns: Classification logits or pre-logits features. rarrIrNrrb)rrhr-r.r+ri)r;rBroutauxs r? forward_headVOLO.forward_heads   u $&&Q&-C    (AqD'CC NN1  Jiin == $--!QR%)Ccggajm++C rAcLURU5nURU5nU$)zForward pass (simplified, without mix token training). Args: x: Input tensor of shape (B, C, H, W). Returns: Classification logits. )rrrs r?rY VOLO.forwards- @  ! !! $   a rA)r+rr*r%rrr.r-rrrr(r,rrrrr#r"r)r$F)T)N)NFFrzF)rFT))r\r]r^r_r`r1rrrastrrrrbrcr rr/r0rdrignoresetr@rrrGrKrNrRrerYr^r rsrrrrrYrfrgrhs@r?rrsB(  #&#%.2-1,G2M""!#%$&$&*,,,5A!%"'!"3e'Ie'e' e'  e'  e'e'!e'!c+e' S *e'tSy)e' %T3Y/e'e'e'e' !!e'""#e'$"%e'&RYY'e'("%S/2)e'*+e', -e'./e'e'N -ryy -T - YY*** YY D T#s(^  , YY)T)T)) YY rCrhsmrW[r$ .U\\ell&;6u||;6ellE%,,X]XdXdfkloqtvy{~l~fJEA7A1B;6@8<$$',? ||? eCcN34?  ?  ?  ? !%?  tELL!5tELL7I)I#JJ K? F./$# 3S >*  c 8%,,5<<(ell4  %,,  rAvariant pretrainedc ^URSS5n[[UU4S[USS90UD6$)zCreate VOLO model. Args: variant: Model variant name. pretrained: Whether to load pretrained weights. **kwargs: Additional model arguments. Returns: VOLO model instance. out_indicesrF feature_cfggetter)r feature_cls)poprrr&)rrrrs r? _create_volorsE**]A.K   [hG     rAurlc 4USSSSSS[[SSS S . UE$) zCreate model configuration. Args: url: URL for pretrained weights. **kwargs: Additional configuration options. Returns: Model configuration dictionary. r)rFrrNQ?bicubicTzpatch_embed.conv.0)r.r+z apache-2.0) rr input_size pool_sizecrop_pct interpolationfixed_input_sizerhr first_conv classifierlicenser )rrs r?_cfgrs8=t)%.B*:N    rAztimm/zLhttps://github.com/sail-sg/volo/releases/download/volo_1/d1_224_84.2.pth.tarr) hf_hub_idrrzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d1_384_85.2.pth.tarr)rFrr)rrrrzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d2_224_85.2.pth.tarzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d2_384_86.0.pth.tarzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d3_224_85.4.pth.tarzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d3_448_86.3.pth.tar)rFrzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d4_224_85.7.pth.tarzMhttps://github.com/sail-sg/volo/releases/download/volo_1/d4_448_86.79.pth.targffffff?zMhttps://github.com/sail-sg/volo/releases/download/volo_1/d5_224_86.10.pth.tarzLhttps://github.com/sail-sg/volo/releases/download/volo_1/d5_448_87.0.pth.tarzMhttps://github.com/sail-sg/volo/releases/download/volo_1/d5_512_87.07.pth.tar)rFr) zvolo_d1_224.sail_in1kzvolo_d1_384.sail_in1kzvolo_d2_224.sail_in1kzvolo_d2_384.sail_in1kzvolo_d3_224.sail_in1kzvolo_d3_448.sail_in1kzvolo_d4_224.sail_in1kzvolo_d4_448.sail_in1kzvolo_d5_224.sail_in1kzvolo_d5_448.sail_in1kzvolo_d5_512.sail_in1kc >[SSSSS.UD6n[SSU0UD6nU$)VOLO-D1 model, Params: 27M.r,r,rrGrrr rrrrrrru) volo_d1_224r&rrr model_argsmodels r?rr4p\6JVepiopJ  L: L LE LrAc >[SSSSS.UD6n[SSU0UD6nU$)rrrrrrru) volo_d1_384rrs r?rr rrAc >[SSSSS.UD6n[SSU0UD6nU$)VOLO-D2 model, Params: 59M.rr, r,rrrrrrrrrru) volo_d2_224rrs r?rr(4q]7KWfqjpqJ  L: L LE LrAc >[SSSSS.UD6n[SSU0UD6nU$)rrrrrrru) volo_d2_384rrs r?rr0rrAc >[SSSSS.UD6n[SSU0UD6nU$)VOLO-D3 model, Params: 86M.rrrr,rrrrru) volo_d3_224rrs r?rr8rrAc >[SSSSS.UD6n[SSU0UD6nU$)rrrrrrru) volo_d3_448rrs r?rr@rrAc >[SSSSS.UD6n[SSU0UD6nU$)VOLO-D4 model, Params: 193M.rrrrrrrrrrru) volo_d4_224rrs r?rrH4r]7KWgrkqrJ  L: L LE LrAc >[SSSSS.UD6n[SSU0UD6nU$)rrrrrrru) volo_d4_448rrs r?rrPrrAc B[SSSSSSS.UD6n[S SU0UD6nU$) bVOLO-D5 model, Params: 296M. stem_hidden_dim=128, the dim in patch embedding is 128 for VOLO-D5. rrr,rrr,rrrrmrrru) volo_d5_224rrs r?rrXB 4+?K[S4,24J  L: L LE LrAc B[SSSSSSS.UD6n[S SU0UD6nU$) rrrrr,rrrru) volo_d5_448rrs r?rrerrAc B[SSSSSSS.UD6n[S SU0UD6nU$) rrrrr,rrrru) volo_d5_512rrs r?rrrrrAr) rrFrrGrlFrr[NN)rlFrr[r)r)Fr`rMtypingrrrrrrr r rdtorch.nnr1torch.nn.functional functionalrQ timm.datar r timm.layersrrrrrr_builderr _featuresr _manipulater _registryrr__all__rrrjrrrrrrrarcrrrrbrrrrrr default_cfgsrrrrrrrrrrrrurAr?rsm* JJJ AZZ*+#< (OryyOdE EPA AH1"))1h>RYY>BD7D7N :# : : : >E#s(O>%>>E#sTWY\J\D]>BCCL""T "7777S 7  7  77777777 ]]!7@ "****S *  *  *****]]*Zw299wt#434*c#$sCx.*%! Z" Z0" Z" Z0" Z" Z0" Z" [-1" [" Z-1" [-1S-&- `DCDDCDDCDDCDDCDDCDDCDDCD D C D   D C D   D C D  rA