+ jn}^RIt^RIHt^RIHt^RIHt^RIHtH t H t ^RI t ^RI t ^RIHuHt^RI HtHt^RIHtHtHt^RIHt^R IHtHt^R IHtHt^R IH t ^R I!H"t"H#t#^R I$H%t%.R1Ot&RRlt'RRlt(RRlt)!RR]PT4t+!RR]PT4t,!RR]PT4t-!RR]PT4t.!RR]PT4t/!R R!]PT4t0!R"R#]PT4t1!R$R%]PT4t2!R&R]PT4t3R2R'R(llt4!R)R]4t5]!4]!R*]5Pl3R+7R,RR-R./R/R0ll44t7R#)3N) OrderedDict)Sequence)partial)AnyCallableOptional)nnTensor)register_modelWeights WeightsEnum)_IMAGENET_CATEGORIES)_ovewrite_named_paramhandle_legacy_interface)Conv2dNormActivationSqueezeExcitation)StochasticDepth)ImageClassificationInterpolationMode)_log_api_usage_onceMaxVitMaxVit_T_Weightsc V^8dQhR\\\3,R\R\R\R\\\3,/#) input_size kernel_sizestridepaddingreturn)tupleint)formats"Z/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchvision/models/maxvit.py __annotate__r$sFuS#XSRU`chmnqsvnvhwcV^,V, ^V,,V,^,V^,V, ^V,,V,^,3#)r)rrrrs&&&&r#_get_conv_output_shaper(sJ A $q7{ 2v=A A $q7{ 2v=A r%c V^8dQhR\\\3,R\R\\\\3,,/#)rrn_blocksr)r r!list)r"s"r#r$r$!s;sCxCDQVWZ\_W_Q`Lar%c.p\V^^^4p\V4F"p\V^^^4pVPV4K$ V#)zQUtil function to check that the input size is correct for a MaxVit configuration.)r(rangeappend)rr*shapesblock_input_shape_s&& r#_make_block_input_shapesr2!sL F.z1aC 8_23DaAN '( Mr%cPV^8dQhR\R\R\P/#)rheightwidthr)r!torchr )r"s"r#r$r$+s%###S#U\\#r%c$\P!\P!\P!V4\P!V4.RR74p\P!V^4pVR,VR,, pVP ^^^4P 4pVR;;,V^, , uu&VR;;,V^, , uu&VR;;,^V,^, ,uu&VPR4#)ij)indexing)NNNr:N)r:Nr:)r:r:r)r:r:)r6stackmeshgridarangeflattenpermute contiguoussum)r4r5coords coords_flatrelative_coordss&& r#_get_relative_position_indexrG+s [[f)=u||E?R(S^bc dF--*K!*- J0GGO%--aA6AACOG *G )GE A -   r ""r%cXaa]tRt^6toRtRV3RlV3RllltV3RlRltRtVtV;t #)MBConvaMBConv: Mobile Inverted Residual Bottleneck. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. expansion_ratio (float): Expansion ratio in the bottleneck. squeeze_ratio (float): Squeeze ratio in the SE Layer. stride (int): Stride of the depthwise convolution. activation_layer (Callable[..., nn.Module]): Activation function. norm_layer (Callable[..., nn.Module]): Normalization function. p_stochastic_dropout (float): Probability of stochastic depth. c<V^8dQhRS[RS[RS[RS[RS[RS[RS[P3,RS[RS[P3,R S[R R / #) r in_channels out_channelsexpansion_ratio squeeze_ratioractivation_layer. norm_layerp_stochastic_dropoutrN)r!floatrr Module)r" __classdict__s"r#r$MBConv.__annotate__Ds;-;-;-;- ;-  ;-  ;-#3 >2;-S"))^,;-$;- ;-r%c $<\SV`4VV^8g;'gW8gp V 'd^\P!W^^RR7.p V^8Xd"\P!^V^R7.V ,p \P !V !VnM\P!4Vn\W#,4p \W$,4p V'd\VRR7Vn M\P!4Vn \4p V!V4V R&\VV ^^^VVRR7V R &\V V ^V^VVV RR 7 V R &\W\PR 7V R &\P!W^RR7V R&\P !V 4VnR#)r;T)rrbiasrrrrowmodepre_normN)rrrrOrPinplaceconv_a)rrrrOrPgroupsr]conv_b) activationsqueeze_excitation)rKrLrrWconv_c)super__init__r Conv2d AvgPool2d SequentialprojIdentityr!rstochastic_depthrrrSiLUlayers)selfrKrLrMrNrrOrPrQ should_projri mid_channels sqz_channels_layers __class__s&&&&&&&&& r#reMBConv.__init__Dsk  k@@[%@ IIkQqW[\]D{ 61MNQUU t,DI DI<9: <78 $34Hu$UD !$&KKMD !-(5 0  -!  1  -!  ):,acahah(i$%II,ghostmmG, r%c&<V^8dQhRS[RS[/#rxrr )r"rTs"r#r$rUs   F r%c tVPV4pVPVPV44pW!,#)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. Returns: Tensor: Output tensor with expected layout of [B, C, H / stride, W / stride]. )rirkrmrnrwress&& r#forwardMBConv.forwards0iil  ! !$++a. 1wr%)rmrirk) __name__ __module__ __qualname____firstlineno____doc__rer|__static_attributes____classdictcell__ __classcell__rsrTs@@r#rIrI6s$ ;-;-z   r%rIcfaa]tRt^toRtV3RlV3RlltV3RlRltV3RlRltRtVt V;t #) $RelativePositionalMultiHeadAttentionzRelative Positional Multi-Head Attention. Args: feat_dim (int): Number of input features. head_dim (int): Number of features per head. max_seq_len (int): Maximum sequence length. c0<V^8dQhRS[RS[RS[RR/#)rfeat_dimhead_dim max_seq_lenrNr!)r"rTs"r#r$1RelativePositionalMultiHeadAttention.__annotate__s9QQQQ Q  Qr%c<\SV`4W,^8wd\RV RV 24hW,VnW n\ \ P!V44VnW0n \P!WPVP,^,4Vn VR,Vn \P!VPVP,V4Vn\PP!\"P$!^VP,^, ^VP,^, ,VP3\"P&R74VnVP+R\-VPVP44\"PP.P1VP(RR7R#) rz feat_dim: z must be divisible by head_dim: )dtyperelative_position_index{Gz?stdNg)rdre ValueErrorn_headsrr!mathsqrtsizerr Linearto_qkv scale_factormerge parameter Parameterr6emptyfloat32relative_position_bias_tableregister_bufferrGinit trunc_normal_)rnrrrrss&&&&r#re-RelativePositionalMultiHeadAttention.__init__sK   ! #z(3ST\S]^_ _+   +./ &ii,,*F*JK $dNYYt}}t||;XF ,.LL,B,B KK!dii-!+DII 0ABDLLQY^YfYf g- ) 68TUYU^U^`d`i`i8jk  ##D$E$E4#Pr%c4<V^8dQhRS[P/#)rrr6r )r"rTs"r#r$rs**ell*r%c VPPR4pVPV,PVPVPR4pVP ^^^4P 4pVP ^4#)r;r<)rviewrrrArB unsqueeze)rn bias_index relative_biass& r#get_relative_positional_biasARelativePositionalMultiHeadAttention.get_relative_positional_biasss1166r: 99*EJJ4K[K[]a]m]moqr %--aA6AAC &&q))r%c&<V^8dQhRS[RS[/#rvrx)r"rTs"r#r$rsFr%c VPwr#rEVPVPrvVPV4p\P !V^RR7wrp V P W#WFV4P^^^^^4p V P W#WFV4P^^^^^4p V P W#WFV4P^^^^^4p WP,p \P!RW4p VP4p \P!W,RR7p \P!RW4pVP^^^^^4P W#WE4pVPV4pV#)z Args: x (Tensor): Input tensor with expected layout of [B, G, P, D]. Returns: Tensor: Output tensor with expected layout of [B, G, P, D]. )dimz!B G H I D, B G H J D -> B G H I Jz!B G H I J, B G H J D -> B G H I Dr<)shaperrrr6chunkreshaperAreinsumrFsoftmaxr)rnrwBGPDHDHqkvqkvdot_prodpos_biasouts&& r#r|,RelativePositionalMultiHeadAttention.forwards8WW a dmm2kk!n++c1"-a IIaA" % - -aAq! < IIaA" % - -aAq! < IIaA" % - -aAq! < !! !<< CQJ44699X0b9ll>Lkk!Q1a(00q<jjo r%)rrrrrrrr) rrrrrrerr|rrrrs@@r#rrs/QQ8** r%rcTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #)SwapAxeszPermute the axes of a tensor.c*<V^8dQhRS[RS[RR/#)rabrNr)r"rTs"r#r$SwapAxes.__annotate__s"##$r%c<<\SV`4WnW nR#N)rdrerr)rnrrrss&&&r#reSwapAxes.__init__s r%cN<V^8dQhRS[PRS[P/#rvr)r"rTs"r#r$rs#%,,r%c\\P!WPVP4pV#r)r6swapaxesrrrzs&& r#r|SwapAxes.forwardsnnQ/ r%)rrrrs@@r#rrs!' r%rcTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #)WindowPartitionz: Partition the input tensor into non-overlapping windows. c<V^8dQhRR/#rrNr')r"rTs"r#r$WindowPartition.__annotate__$r%c$<\SV`4R#rrdrernrss&r#reWindowPartition.__init__ r%c,<V^8dQhRS[RS[RS[/#)rrwprr r!)r"rTs"r#r$rs"CFr%c VPwr4rVTpVPW4WW,WvV,V4pVP^^^^^^4pVPW5V,Wg,,Ww,V4pV#)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. p (int): Number of partitions. Returns: Tensor: Output tensor with expected layout of [B, H/P, W/P, P*P, C]. rrrA)rnrwrrCrWrs&&& r#r|WindowPartition.forwardslWW a  IIaAFAAvq 1 IIaAq!Q ' IIaq&QV,aeQ 7r%r'rrs@@r#rrs#r%rcTaa]tRt^toRtV3RlV3RlltV3RlRltRtVtV;t #)WindowDepartitionzg Departition the input tensor of non-overlapping windows into a feature volume of layout [B, C, H, W]. c<V^8dQhRR/#rr')r"rTs"r#r$WindowDepartition.__annotate__rr%c$<\SV`4R#rrrs&r#reWindowDepartition.__init__rr%c 8<V^8dQhRS[RS[RS[RS[RS[/#)rrwr h_partitions w_partitionsrr)r"rTs"r#r$rs1Cs#RXr%c VPwrVrxTp Y4rVPWZWW4pVP^^^^^^4pVPWXW,W,4pV#)a2 Args: x (Tensor): Input tensor with expected layout of [B, (H/P * W/P), P*P, C]. p (int): Number of partitions. h_partitions (int): Number of vertical partitions. w_partitions (int): Number of horizontal partitions. Returns: Tensor: Output tensor with expected layout of [B, C, H, W]. r) rnrwrrrrrPPrrHPWPs &&&&& r#r|WindowDepartition.forwards`gg b B IIaRA ) IIaAq!Q ' IIaBFBF +r%r'rrs@@r#rrs#r%rcTaa]tRtRtoRtV3RlV3RlltV3RlRltRtVtV;t #)PartitionAttentionLayeriav Layer for partitioning the input tensor into non-overlapping windows and applying attention to each window. Args: in_channels (int): Number of input channels. head_dim (int): Dimension of each attention head. partition_size (int): Size of the partitions. partition_type (str): Type of partitioning to use. Can be either "grid" or "window". grid_size (Tuple[int, int]): Size of the grid to partition the input tensor into. mlp_ratio (int): Ratio of the feature size expansion in the MLP layer. activation_layer (Callable[..., nn.Module]): Activation function to use. norm_layer (Callable[..., nn.Module]): Normalization function to use. attention_dropout (float): Dropout probability for the attention layer. mlp_dropout (float): Dropout probability for the MLP layer. p_stochastic_dropout (float): Probability of dropping out a partition. c<V^8dQhRS[RS[RS[RS[RS[S[S[3,RS[RS[RS[P 3,R S[RS[P 3,R S[R S[R S[R R/ #)rrKrpartition_sizepartition_type grid_size mlp_ratiorO.rPattention_dropout mlp_dropoutrQrN)r!strr rr rSrR)r"rTs"r#r$$PartitionAttentionLayer.__annotate__-s8T8T8T8T  8T  8Tc?8T8T#3 >28TS"))^,8T!8T8T$8T !8Tr%c <\S V`4W,VnW nV^,V,VnW@nWPnVR9d \R4hVR8XdW0PuVnVn MVPVuVnVn \4Vn \4Vn VR8Xd \RR 4M\P !4VnVR8Xd \RR 4M\P !4Vn\P&!V!V4\)WV^,4\P*!V 44Vn\P&!\P.!V4\P0!WV,4V!4\P0!W,V4\P*!V 44Vn\5V RR7VnR#) rgridwindowz0partition_type must be either 'grid' or 'window'rYrZN)rr)rdrerr n_partitionsrrrrgr partition_oprdepartition_oprr rjpartition_swapdepartition_swaprhrDropout attn_layer LayerNormr mlp_layerrstochastic_dropout) rnrKrrrrrrOrPrrrQrss &&&&&&&&&&&&r#re PartitionAttentionLayer.__init__-su" ".  %aLN:," !3 3OP P X %+->->NDFDF!..NDFDF+-/12@F2Jhr2.PRP[P[P]4Bf4LR 0RTR]R]R_-- { # 1XYHY Z JJ( )   LL % IIk#: ;   IIk-{ ; JJ{ #  #22FU"Sr%c&<V^8dQhRS[RS[/#rvrx)r"rTs"r#r$rgsFr%c VP^,VP,VP^,VP,r2\P!VP^,VP,^8H;'d(VP^,VP,^8HRP VPVP44VP WP4pVP V4pWPVPV44,pWPVPV44,pVPV4pVPWPW#4pV#)z Args: x (Tensor): Input tensor with expected layout of [B, C, H, W]. Returns: Tensor: Output tensor with expected layout of [B, C, H, W]. z[Grid size must be divisible by partition size. Got grid size of {} and partition size of {}) rrr6_assertr"r r rrrr r )rnrwghgws&& r#r|PartitionAttentionLayer.forwardgs"dff,dnnQ.?466.IB NN1  &! + O Oq0ADFF0Ja0O i p p     a (    " ''(:; ; ''q(9: :  ! !! $   662 2r%)rr r rrrrrrrr r rrrrs@@r#rrs&"8T8Ttr%rcTaa]tRtRtoRtV3RlV3RlltV3RlRltRtVtV;t #) MaxVitLayeria MaxVit layer consisting of a MBConv layer followed by a PartitionAttentionLayer with `window` and a PartitionAttentionLayer with `grid`. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. expansion_ratio (float): Expansion ratio in the bottleneck. squeeze_ratio (float): Squeeze ratio in the SE Layer. stride (int): Stride of the depthwise convolution. activation_layer (Callable[..., nn.Module]): Activation function. norm_layer (Callable[..., nn.Module]): Normalization function. head_dim (int): Dimension of the attention heads. mlp_ratio (int): Ratio of the MLP layer. mlp_dropout (float): Dropout probability for the MLP layer. attention_dropout (float): Dropout probability for the attention layer. p_stochastic_dropout (float): Probability of stochastic depth. partition_size (int): Size of the partitions. grid_size (Tuple[int, int]): Size of the input feature grid. c<V^8dQhRS[RS[RS[RS[RS[RS[RS[P3,RS[RS[P3,R S[R S[R S[R S[R S[RS[RS[S[S[3,RR/#)rrKrLrNrMrrP.rOrrrrrQrrrN)r!rRrr rSr )r"rTs"r#r$MaxVitLayer.__annotate__s?,?,?, ?,  ?,  ?,?,S"))^,?,#3 >2?,?,?,?,!?, $!?,$%?,&c?'?,( )?,r%c6<\SV`4\4p\VVVVVVVV R7VR&\ VVV RVV V\ P V V V R7 VR&\ VVV RVV V\ P V V V R7 VR&\ P!V4VnR#))rKrLrMrNrrOrPrQMBconvr) rKrrrrrrOrPrrrQwindow_attentionrgrid_attentionN) rdrerrIrr rrhrm)rnrKrLrNrMrrPrOrrrrrQrrrmrss&&&&&&&&&&&&&&& r#reMaxVitLayer.__init__s* )m"#%+'-!!5  x&=$)#-||/#!5 & !"$;$)!-||/#!5 $  mmF+ r%c&<V^8dQhRS[RS[/#rvrx)r"rTs"r#r$rsFr%c (VPV4pV#zu Args: x (Tensor): Input tensor of shape (B, C, H, W). Returns: Tensor: Output tensor of shape (B, C, H, W). rm)rnrws&&r#r|MaxVitLayer.forwards KKNr%r%rrs@@r#rrs$(?,?,Br%rcTaa]tRtRtoRtV3RlV3RlltV3RlRltRtVtV;t #) MaxVitBlockia A MaxVit block consisting of `n_layers` MaxVit layers. Args: in_channels (int): Number of input channels. out_channels (int): Number of output channels. expansion_ratio (float): Expansion ratio in the bottleneck. squeeze_ratio (float): Squeeze ratio in the SE Layer. activation_layer (Callable[..., nn.Module]): Activation function. norm_layer (Callable[..., nn.Module]): Normalization function. head_dim (int): Dimension of the attention heads. mlp_ratio (int): Ratio of the MLP layer. mlp_dropout (float): Dropout probability for the MLP layer. attention_dropout (float): Dropout probability for the attention layer. p_stochastic_dropout (float): Probability of stochastic depth. partition_size (int): Size of the partitions. input_grid_size (Tuple[int, int]): Size of the input feature grid. n_layers (int): Number of layers in the block. p_stochastic (List[float]): List of probabilities for stochastic depth for each layer. c<V^8dQhRS[RS[RS[RS[RS[RS[P3,RS[RS[P3,RS[R S[R S[R S[R S[R S[S[S[3,RS[RS[S[,RR/#)rrKrLrNrMrP.rOrrrrrinput_grid_sizen_layers p_stochasticrN)r!rRrr rSr r+)r"rTs"r#r$MaxVitBlock.__annotate__s111 1  1  1S"))^,1#3 >21111!1 !1"sCx#1&'1(5k)1* +1r%c<\SV`4\V4V 8Xg\RV RV R24h\P !4Vn\V ^^^R7Vn\V4FUwppV^8Xd^M^pT;P \V^8XdTMTVVVVVVVVV V V VPVR7., unKW R#)z'p_stochastic must have length n_layers=z, got p_stochastic=.rX)rKrLrNrMrrPrOrrrrrrrQN) rdrelenrr ModuleListrmr(r enumerater)rnrKrLrNrMrPrOrrrrrr*r+r,idxrrrss&&&&&&&&&&&&&&& r#reMaxVitBlock.__init__s, < H,FxjPcdpcqqrst tmmo /QWXbcd -FC(QF KK/2ax \!-"/$3!)%5%' +&7#1"nn)* K.r%c&<V^8dQhRS[RS[/#rvrx)r"rTs"r#r$r--s   F r%c <VPF pV!V4pK V#r$r%)rnrwlayers&& r#r|MaxVitBlock.forward-s [[EaA!r%)rrmrrs@@r#r(r(s$*11f   r%r(caa]tRtRtoRtR]P R^^RRR3V3RlV3RllltV3R lR ltR t R t Vt V;t #) ri9a1 Implements MaxVit Transformer from the `MaxViT: Multi-Axis Vision Transformer `_ paper. Args: input_size (Tuple[int, int]): Size of the input image. stem_channels (int): Number of channels in the stem. partition_size (int): Size of the partitions. block_channels (List[int]): Number of channels in each block. block_layers (List[int]): Number of layers in each block. stochastic_depth_prob (float): Probability of stochastic depth. Expands to a list of probabilities for each layer that scales linearly to the specified value. squeeze_ratio (float): Squeeze ratio in the SE Layer. Default: 0.25. expansion_ratio (float): Expansion ratio in the MBConv bottleneck. Default: 4. norm_layer (Callable[..., nn.Module]): Normalization function. Default: None (setting to None will produce a `BatchNorm2d(eps=1e-3, momentum=0.01)`). activation_layer (Callable[..., nn.Module]): Activation function Default: nn.GELU. head_dim (int): Dimension of the attention heads. mlp_ratio (int): Expansion ratio of the MLP layer. Default: 4. mlp_dropout (float): Dropout probability for the MLP layer. Default: 0.0. attention_dropout (float): Dropout probability for the attention layer. Default: 0.0. num_classes (int): Number of classes. Default: 1000. Ng?r~ic <V^8dQhRS[S[S[3,RS[RS[RS[S[,RS[S[,RS[RS[RS[S[R S[P3,,R S[R S[P3,R S[R S[R S[RS[RS[RS[RR/#)rr stem_channelsrblock_channels block_layersrstochastic_depth_probrP.rOrNrMrrr num_classesrN)r r!r+rRrrr rS)r"rTs"r#r$MaxVit.__annotate__Nstt#s(Ot  ttS t3itt %t"Xc299n56#t$#3 >2%t()t*+t./t01t2!3t67t8 9tr%c,<\SV`4\V4^pVf\\P RRR7p\ V\V44p\V4FHwppV^,V,^8wgV^,V,^8wgK2\RV RV RV RV R 2 4h \P!\VV^^VV R RR 7\W"^^RRR R 74Vn \V^^^R7pW0n\P!4VnV.VRR,pTp\"P$!^V\'V44P)4p^p\+VVV4FiwpppVP P-\/VVV V VV VV V VVVVVVVV,R74VP R,P0pVV, pKk \P!\P2!^4\P4!4\P6!VR,4\P8!VR,VR,4\P:!4\P8!VR,VR R74VnVP?4R#)NgMbP?g{Gz?)epsmomentumz Input size z of block z$ is not divisible by partition size zx. Consider changing the partition size or the input size. Current configuration yields the following block input sizes: r/F)rrPrOrWr]T)rrPrOrWrX)rKrLrNrMrPrOrrrrrr*r+r,)rWr<) rdrerrr BatchNorm2dr2r0r2rrhrstemr(rr1blocksnplinspacerCtolistzipr.r(rAdaptiveAvgPool2dFlattenrrTanh classifier _init_weights)rnrr;rr<r=rr>rPrOrNrMrrrr?input_channelsblock_input_sizesr3block_input_sizerKrLr,p_idx in_channel out_channel num_layersrss&&&&&&&&&&&&&&&& r#reMaxVit.__init__Ns: D!   TDIJ 5Z^ATU%./@%A !C!"^3q857WW#!" #&7ttl + +r%cV^8dQhR\R\\,R\\,R\R\R\R\\,R\ R \ R \/ #) rr;r<r=r>rrweightsprogresskwargsr)r!r+rRrr boolrr)r"s"r#r$r$s{'''I ' s) ' ! '''k "'''  !'r%c Ve\VR\VPR,44VPR,^,VPR,^,8XgQh\VRVPR,4VPRR 4p \ RRVRVRVRVR VR VRV /VBp Ve#V P VP VR R 74V #)Nr? categoriesmin_sizerr;r<r=r>rrT)ri check_hashrqr')rr0metapoprload_state_dictget_state_dict) r;r<r=r>rrrhrirjrmodels &&&&&&&&, r#_maxvitrws$fmSl9S5TU||J'*gll:.Fq.IIIIflGLL4LML*5J  # % " 4    &    E g44hSW4XY Lr%c]tRtRt]!R]!]^^]PR7R] RRRRRR R R R R RR//RRRRRR/R7t ] t Rt R#)riz9https://download.pytorch.org/models/maxvit_t-bc5ab103.pth) crop_size resize_size interpolationrm num_paramsirnrecipezLhttps://github.com/pytorch/vision/tree/main/references/classification#maxvit_metricsz ImageNet-1Kzacc@1gT@zacc@5g|?5.X@_opsgZd;@ _file_sizegK7]@_docszThese weights reproduce closely the results of the paper using a similar training recipe. They were trained with a BatchNorm2D momentum of 0.99 instead of the more correct 0.01.)url transformsrrr'Nrp) rrrrr rrrBICUBICr IMAGENET1K_V1DEFAULTrr'r%r#rrsz G 3CO`OhOh  . (  d VV  E ' g M.Gr% pretrained)rhrhriTc^V^8dQhR\\,R\R\R\/#)rrhrirjr)rrrkrr)r"s"r#r$r$!s0!!"23!d!]`!ek!r%c j\PV4p\R R^@R.R OR.R OR^ RRR^RVR V/VB#) aF Constructs a maxvit_t architecture from `MaxViT: Multi-Axis Vision Transformer `_. Args: weights (:class:`~torchvision.models.MaxVit_T_Weights`, optional): The pretrained weights to use. See :class:`~torchvision.models.MaxVit_T_Weights` below for more details, and possible values. By default, no pre-trained weights are used. progress (bool, optional): If True, displays a progress bar of the download to stderr. Default is True. **kwargs: parameters passed to the ``torchvision.models.maxvit.MaxVit`` base class. Please refer to the `source code `_ for more details about this class. .. autoclass:: torchvision.models.MaxVit_T_Weights :members: r;r<r=rr>g?rrhrir')@i)rrr)rverifyrw)rhrirjs$$,r#maxvit_trsw.%%g.G   * "   "        r%)rrr)NF)8r collectionsrcollections.abcr functoolsrtypingrrrnumpyrHr6torch.nn.functionalr functionalrr torchvision.models._apir r r torchvision.models._metartorchvision.models._utilsrrtorchvision.ops.miscrr torchvision.ops.stochastic_depthrtorchvision.transforms._presetsrrtorchvision.utilsr__all__r(r2rGrSrIrrrrrrr(rrwrrrr'r%r#rsG #$** HH9TH<R1 #TRYYTnF299FR ryy bii4