
    bCiqo              	          S r SSKrSSKrSSKJr  SSKrSSKrSSKJrJ	r	  SSK
Jr  SSKJrJrJrJr  SSKJr  SS	KJrJr  SS
KJr  SSKJr  \R4                  " \5      rS6S\\\4   4S jjr " S S\	R@                  5      r! " S S\	RD                  5      r# " S S\	RH                  5      r% " S S\	RL                  5      r' " S S\	RH                  5      r(S7S\R                  S\)S\S\R                  4S jjr* " S S\	RH                  5      r+S8S jr, " S  S!\	RH                  5      r- " S" S#\	RH                  5      r. " S$ S%\	RH                  5      r/ " S& S'\	RH                  5      r0 " S( S)\	RH                  5      r1\ " S* S+\5      5       r2\ " S, S-\25      5       r3\" S.S/9 " S0 S1\25      5       r4\" S2S/9 " S3 S4\2\5      5       r5/ S5Qr6g)9z9PyTorch BiT model. Also supports backbone for ViT hybrid.    N)Optional)Tensornn   )ACT2FN)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)BackboneMixin   )	BitConfigreturnc                 "   SnU c  US-
  X1S-
  -  -   S-  n X4$ [        U [        5      (       a`  U R                  5       n U S:X  a/  US:X  a!  X1S-
  -  S-  S:X  a  US-
  X1S-
  -  -   S-  n X4$ Sn Sn X4$ U S:X  a  Sn X4$ US-
  X1S-
  -  -   S-  n X4$ )a<  
Utility function to get the tuple padding value given the kernel_size and padding.

Args:
    padding (Union[`str`, `int`], *optional*):
        Padding value, can be either `"same"`, `"valid"`. If a different value is provided the default padding from
        PyTorch is used.
    kernel_size (`int`, *optional*, defaults to 7):
        Kernel size of the convolution layers.
    stride (`int`, *optional*, defaults to 1):
        Stride value of the convolution layers.
    dilation (`int`, *optional*, defaults to 1):
        Dilation value of the convolution layers.
Fr      samer   Tvalid)
isinstancestrlower)paddingkernel_sizestridedilationdynamics        ^/home/james-whalen/.local/lib/python3.13/site-packages/transformers/models/bit/modeling_bit.pyget_padding_valuer    )   s     GQJ(Ao">>1D'3--/f{!O <AQF"QJ(Ao*FF1L    G  
h/&BBqHG    c                   B   ^  \ rS rSrSr      SU 4S jjrS rSrU =r$ )WeightStandardizedConv2dR   zConv2d with Weight Standardization. Includes TensorFlow compatible SAME padding. Used for ViT Hybrid model.

Paper: [Micro-Batch Training with Batch-Channel Normalization and Weight
Standardization](https://huggingface.co/papers/1903.10520v2)
c
                    > [        XSXFS9u  pZ[        TU ]	  UUUUUUUUS9  U
(       a  [        X4U5      U l        OS U l        Xl        g )N)r   r   )r   r   r   groupsbias)r    super__init__DynamicPad2dpadeps)self
in_channelout_channelsr   r   r   r   r&   r'   r,   
is_dynamic	__class__s              r   r)   !WeightStandardizedConv2d.__init__Y   s]     0Vg 	 		
 #KBDHDHr!   c           	         U R                   b  U R                  U5      n[        R                  R                  U R                  R                  SU R                  S5      S S SSU R                  S9R                  U R                  5      n[        R                  R                  XU R                  U R                  U R                  U R                  U R                  5      nU$ )Nr   T        )trainingmomentumr,   )r+   r   
functional
batch_normweightreshaper/   r,   
reshape_asconv2dr'   r   r   r   r&   )r-   hidden_stater:   s      r   forward WeightStandardizedConv2d.forwardv   s    8888L1L))KK4#4#4b94PT_bhlhphp * 

*T[[
! 	 }}++$))T[[$,,W[WbWb
 r!   )r,   r+   )r   SAMEr   r   Fgư>	__name__
__module____qualname____firstlineno____doc__r)   r?   __static_attributes____classcell__r1   s   @r   r#   r#   R   s+     :	 	r!   r#   c                   6   ^  \ rS rSrSrSU 4S jjrS rSrU =r$ )BitGroupNormActivation   zI
A module that combines group normalization with an activation function.
c                    > [         TU ]  UR                  X#US9  U(       a  [        UR                     U l        g [        R                  " 5       U l        g )N)r,   affine)r(   r)   
num_groupsr   
hidden_act
activationr   Identity)r-   confignum_channelsr,   rO   apply_activationr1   s         r   r)   BitGroupNormActivation.__init__   s?    **L&Q$V%6%67DO kkmDOr!   c                     [         R                  R                  XR                  U R                  U R
                  U R                  5      nU R                  U5      nU$ N)r   r8   
group_normrP   r:   r'   r,   rR   )r-   r>   s     r   r?   BitGroupNormActivation.forward   sF    }}//oot{{\`\e\egkgogop|4r!   )rR   )gh㈵>TTrB   rJ   s   @r   rL   rL      s    , r!   rL   c                   6   ^  \ rS rSrSrSU 4S jjrS rSrU =r$ )r*      z
A module that wraps dynamic padding of any input, given the parameters of the convolutional layer and the input
hidden states.
c                    > [         TU ]  5         [        U[        5      (       a  X4n[        U[        5      (       a  X"4n[        U[        5      (       a  X34nXl        X l        X0l        X@l        S nXPl        g )Nc                 p    [        [        R                  " X-  5      S-
  U-  US-
  U-  -   S-   U -
  S5      $ )Nr   r   )maxmathceil)xr   r   r   s       r   compute_padding.DynamicPad2d.__init__.<locals>.compute_padding   s@    		!*-1V;{QRZ>ZZ]^^abbdeffr!   )	r(   r)   r   intr   r   r   valuerd   )r-   r   r   r   rg   rd   r1   s         r   r)   DynamicPad2d.__init__   sp    k3''&4Kfc""%Fh$$ +H& 
	g  /r!   c           	         UR                  5       SS  u  p#U R                  X R                  S   U R                  S   U R                  S   5      nU R                  X0R                  S   U R                  S   U R                  S   5      nUS:  d  US:  a=  [
        R                  R                  UUS-  XUS-  -
  US-  XDS-  -
  /U R                  S9nU$ )Nr   r   r   )rg   )	sizerd   r   r   r   r   r8   r+   rg   )r-   inputinput_heightinput_widthpadding_heightpadding_widths         r   r?   DynamicPad2d.forward   s    $)JJL$5! --l<L<LQ<OQUQ\Q\]^Q_aeananopaqr,,[:J:J1:Mt{{[\~_c_l_lmn_op A!2MM%%!Q&!Q$66"a'"q%88	 jj & 	E r!   )rd   r   r   r   rg   )r   rB   rJ   s   @r   r*   r*      s    
/, r!   r*   c                   J   ^  \ rS rSrSr      SS\4U 4S jjjrS rSrU =r	$ )BitMaxPool2d   z1Tensorflow like 'SAME' wrapper for 2D max poolingr   c                   > [        U[        R                  R                  5      (       a  UOX4n[        U[        R                  R                  5      (       a  UOX"4n[        U[        R                  R                  5      (       a  UOX34n[        TU ]  XXSU5        U(       a  [        XX65      U l        g [        R                  " 5       U l        g rY   )
r   collectionsabcIterabler(   r)   r*   r+   r   rS   )	r-   r   r   r   	ceil_moder   padding_valueuse_dynamic_paddingr1   s	           r   r)   BitMaxPool2d.__init__   s     &0[__=U=U%V%Vk]h\v%fkoo.F.FGGfM])(KOO4L4LMM8T\SggK#KQDH{{}DHr!   c                     U R                  U5      n[        R                  R                  XR                  U R
                  U R                  U R                  U R                  5      $ rY   )	r+   r   r8   
max_pool2dr   r   r   r   ry   r-   hidden_statess     r   r?   BitMaxPool2d.forward   sK    /}}''++T[[$,,W[WeWe
 	
r!   )r+   )Nr   F)r   r   r   T)
rC   rD   rE   rF   rG   rf   r)   r?   rH   rI   rJ   s   @r   rs   rs      s6    ;
  %% %&
 
r!   rs   c                   F   ^  \ rS rSrSrS\4U 4S jjrS\S\4S jrSr	U =r
$ )	BitEmbeddings   zD
BiT Embeddings (stem) composed of a single aggressive convolution.
rT   c           	         > [         TU ]  5         [        UR                  UR                  SSSUR
                  S9U l        [        SSUR                  S9U l	        UR
                  b9  UR
                  R                  5       S:X  a  [        R                  " 5       U l        O[        R                  " SS	S
9U l        UR                  S:w  a  [!        XR                  S9U l        O[        R                  " 5       U l        UR                  U l        g )N   r   :0yE>)r   r   r,   r   r   )r   r   r{   rA   )r   r   r   r   r5   )r   rg   preactivationrU   )r(   r)   r#   rU   embedding_sizeglobal_paddingconvolutionrs   embedding_dynamic_paddingpoolerupperr   rS   r+   ConstantPad2d
layer_typerL   normr-   rT   r1   s     r   r)   BitEmbeddings.__init__   s    3!!))
 #qPVPpPpq   ,1F1F1L1L1NRX1X{{}DH''CHDH/.vDYDYZDIDI"//r!   pixel_valuesr   c                     UR                   S   nX R                  :w  a  [        S5      eU R                  U5      nU R	                  U5      nU R                  U5      nU R                  U5      nU$ )Nr   zeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)shaperU   
ValueErrorr   r+   r   r   )r-   r   rU   	embeddings       r   r?   BitEmbeddings.forward  sp    #))!,,,,w  $$\2	HHY'	IIi(	KK	*	r!   )r   r   rU   r+   r   )rC   rD   rE   rF   rG   r   r)   r   r?   rH   rI   rJ   s   @r   r   r      s,    0y 06F v  r!   r   rl   	drop_probr6   c                    US:X  d  U(       d  U $ SU-
  nU R                   S   4SU R                  S-
  -  -   nU[        R                  " X@R                  U R
                  S9-   nUR                  5         U R                  U5      U-  nU$ )a*  
Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
argument.
r5   r   r   )r   )dtypedevice)r   ndimtorchrandr   r   floor_div)rl   r   r6   	keep_probr   random_tensoroutputs          r   	drop_pathr     s     CxII[[^

Q 77E

5ELL YYMYYy!M1FMr!   c                      ^  \ rS rSrSrSS\\   SS4U 4S jjjrS\R                  S\R                  4S jr
S\4S	 jrS
rU =r$ )BitDropPathi,  zXDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).Nr   r   c                 .   > [         TU ]  5         Xl        g rY   )r(   r)   r   )r-   r   r1   s     r   r)   BitDropPath.__init__/  s    "r!   r   c                 B    [        XR                  U R                  5      $ rY   )r   r   r6   r   s     r   r?   BitDropPath.forward3  s    FFr!   c                      SU R                    3$ )Nzp=r   )r-   s    r   
extra_reprBitDropPath.extra_repr6  s    DNN#$$r!   r   rY   )rC   rD   rE   rF   rG   r   floatr)   r   r   r?   r   r   rH   rI   rJ   s   @r   r   r   ,  sQ    b#(5/ #T # #GU\\ Gell G%C % %r!   r   c                 d    Un[        U[        XS-  -   5      U-  U-  5      nUSU -  :  a  X1-  nU$ )Nr   g?)r`   rf   )rg   divisor	min_value	new_values       r   make_divr   :  sC    IIs5Q;#677BWLMI3;	r!   c                   F   ^  \ rS rSrSr        SU 4S jjrS rSrU =r$ )BitPreActivationBottleneckLayeriB  zPre-activation (v2) bottleneck block.
Follows the implementation of "Identity Mappings in Deep Residual Networks":
https://github.com/KaimingHe/resnet-1k-layers/blob/master/resnet-pre-act.lua

Except it puts the stride on 3x3 conv when available.
c           
        > [         TU ]  5         U=(       d    UnU=(       d    Un[        X4-  5      nU
(       a  [        UUUUSS9U l        OS U l        [        X5      U l        [        X+SSUR                  S9U l	        [        XS9U l
        [        XSXXSUR                  S9U l        [        X5      U l        [        XSSUR                  S9U l        U	S	:  a  [        U	5      U l        g [        R                   " 5       U l        g )
NTr   preactr   r   r,   r   r   r   )r   r&   r,   r   r   )r(   r)   r   BitDownsampleConv
downsamplerL   norm1r#   r   conv1norm2conv2norm3conv3r   r   rS   r   )r-   rT   in_channelsr/   bottle_ratior   r   first_dilationr&   drop_path_rateis_first_layermid_channelsr1   s               r   r)   (BitPreActivationBottleneckLayer.__init__J  s     	'38#2{ ;</DO #DO+F@
-kPT^d^s^st
+FN
-&T[a[p[p

 ,FA
-l!QU_e_t_tu
8F8J^4PRP[P[P]r!   c                 0   U R                  U5      nUnU R                  b  U R                  U5      nU R                  U5      nU R                  U R	                  U5      5      nU R                  U R                  U5      5      nU R                  U5      nX-   $ rY   )r   r   r   r   r   r   r   r   )r-   r   hidden_states_preactshortcuts       r   r?   'BitPreActivationBottleneckLayer.forwardv  s    #zz-8 !??&';<H 

#78

4::m#<=

4::m#<=}5''r!   )r   r   r   r   r   r   r   r   N      ?r   r   Nr   r5   FrB   rJ   s   @r   r   r   B  s3     *^X( (r!   r   c                   F   ^  \ rS rSrSr        SU 4S jjrS rSrU =r$ )BitBottleneckLayeri  z\Non Pre-activation bottleneck block, equivalent to V1.5/V1b bottleneck. Used for ViT Hybrid.c                 0  > [         TU ]  5         U=(       d    UnU=(       d    Un[        X4-  5      nU
(       a  [        UUUUSS9U l        OS U l        [        X+SSUR                  S9U l        [        XS9U l	        [        UUSUUUSUR                  S9U l
        [        XS9U l        [        XSSUR                  S9U l        [        XSS	9U l        U	S
:  a  [        U	5      O[        R                   " 5       U l        [$        UR&                     U l        g )NFr   r   r   r   r   r   )r   r   r&   r,   r   rU   rV   r   )r(   r)   r   r   r   r#   r   r   rL   r   r   r   r   r   r   r   rS   r   r   rQ   rR   )r-   rT   r   r/   r   r   r   r   r&   r   r   mid_chsr1   s               r   r)   BitBottleneckLayer.__init__  s    	'38#2{<67/DO #DO-kA4Y_YnYno
+FI
-#))	

 ,FI
-gQDZ`ZoZop
+F`ef
8F8J^4PRP[P[P] !2!23r!   c                 Z   UnU R                   b  U R                  U5      nU R                  U5      nU R                  U5      nU R                  U5      nU R	                  U5      nU R                  U5      nU R                  U5      nU R                  U5      nU R                  X-   5      nU$ rY   )	r   r   r   r   r   r   r   r   rR   )r-   r   r   s      r   r?   BitBottleneckLayer.forward  s     ??&}5H 

=1

=1

=1

=1

=1

=1}5(@Ar!   )	rR   r   r   r   r   r   r   r   r   r   rB   rJ   s   @r   r   r     s0    f /4b r!   r   c                   6   ^  \ rS rSr  SU 4S jjrS rSrU =r$ )r   i  c           	         > [         TU ]  5         [        X#SUSUR                  S9U l        U(       a  [
        R                  " 5       U l        g [        XSS9U l        g )Nr   r   )r   r,   r   Fr   )	r(   r)   r#   r   convr   rS   rL   r   )r-   rT   r   r/   r   r   r1   s         r   r)   BitDownsampleConv.__init__  sX     	,qT6K`K`
	
  KKM 		 (\ab 		r!   c                 B    U R                  U R                  U5      5      $ rY   )r   r   )r-   rc   s     r   r?   BitDownsampleConv.forward  s    yy1&&r!   )r   r   )r   T)rC   rD   rE   rF   r)   r?   rH   rI   rJ   s   @r   r   r     s     
$' 'r!   r   c                   L   ^  \ rS rSrSr  S	U 4S jjrS rS\S\4S jrSr	U =r
$ )
BitStagei  z/
A ResNet v2 stage composed by stacked layers.
c	                 `  > [         TU ]  5         US;   a  SOSn	UR                  S:X  a  [        n
O[        n
Un[
        R                  " 5       U l        [        U5       HM  nU R                  XU5      u  pMnU R                  R                  [        U5      U
" UUUUUUU	UUS9	5        UnUn	MO     g )N)r   r   r   r   
bottleneck)r   r   r   r   r   r   )r(   r)   r   r   r   r   
Sequentiallayersrange_get_updated_hyperparameters
add_moduler   )r-   rT   r   r/   r   r   depthr   layer_dropoutr   	layer_clsprev_chs	layer_idxr   r   r1   s                  r   r)   BitStage.__init__  s     	&&0a ,*I7ImmouI595V5V=62FN KK""I !%!-#1#1#1
 $H%N+ &r!   c                 @    U(       a  X1   nOSnUS:w  a  SnUS:H  nX$U4$ )zd
Get the new hyper-parameters with respect to the previous ones and the index of the current layer.
r5   r   r    )r-   r   r   r   r   r   s         r   r   %BitStage._get_updated_hyperparameters  s4     *5N N>F"a~55r!   rl   r   c                 V    Un[        U R                  5       H  u  p4U" U5      nM     U$ rY   )	enumerater   )r-   rl   r>   _layers        r   r?   BitStage.forward)  s,    !$++.HA .L /r!   )r   )r   N)rC   rD   rE   rF   rG   r)   r   r   r?   rH   rI   rJ   s   @r   r   r     s3     ,&\6 V   r!   r   c            	       V   ^  \ rS rSrS\4U 4S jjrS r SS\S\S\S\	4S	 jjr
S
rU =r$ )
BitEncoderi0  rT   c                   > [         TU ]  5         [        R                  " / 5      U l        UR
                  nSnSn[        R                  " [        R                  " SUR                  [        UR                  5      5      5      R                  UR                  5       Vs/ s H  nUR                  5       PM     nn[        [!        UR                  UR"                  U5      5       HY  u  nu  pn
U R%                  XsXU5      u  pn['        UUUUUUU
S9nUnX<-  nU R                  R)                  [+        U5      U5        M[     g s  snf )N   r   r   )r   r   r   r   )r(   r)   r   
ModuleListstagesr   r   r   nplinspacer   sumdepthssplittolistr   ziphidden_sizesr   r   r   r   )r-   rT   r   current_strider   rc   layer_dropouts	stage_idxcurrent_depthcurrent_hidden_sizer   r/   r   stager1   s                 r   r)   BitEncoder.__init__1  s5   mmB'((  \\"++a1F1FFMMHZ"[\bbcicpcpq
q HHJq 	 

 OXv22NCO
JIJM .2-N-N+>&.*L( !#+E $H$NKK""3y>59+O

s   Ec                 v    [        X5R                  -  5      nUS:X  a  SOSnX%R                  :  a  XG-  nSnXgU4$ )Nr   r   r   )r   width_factoroutput_stride)r-   r  r  r
  r   rT   r/   r   s           r   r   'BitEncoder._get_updated_hyperparametersW  sG     36I6I IJ1n!111HFX--r!   r>   output_hidden_statesreturn_dictr   c                     U(       a  SOS nU R                    H  nU(       a  XA4-   nU" U5      nM     U(       a  XA4-   nU(       d  [        S X4 5       5      $ [        UUS9$ )Nr   c              3   .   #    U  H  oc  M  Uv   M     g 7frY   r   ).0vs     r   	<genexpr>%BitEncoder.forward.<locals>.<genexpr>n  s     S$Aq$As   	)last_hidden_stater   )r   tupler	   )r-   r>   r  r  r   stage_modules         r   r?   BitEncoder.forward_  sk     3 KKL# - ?'5L	 (  )O;MS\$ASSS-*'
 	
r!   )r   )FT)rC   rD   rE   rF   r   r)   r   r   boolr	   r?   rH   rI   rJ   s   @r   r   r   0  sF    $:y $:L. ]a
"
:>
UY
	'
 
r!   r   c                   4    \ rS rSr% \\S'   SrSrS/rS r	Sr
g)	BitPreTrainedModeliv  rT   bitr   r   c                 b   [        U[        R                  5      (       a*  [        R                  R	                  UR
                  SSS9  g [        U[        R                  5      (       a  [        R                  R                  UR
                  [        R                  " S5      S9  UR                  by  [        R                  R                  UR
                  5      u  p#US:  a  S[        R                  " U5      -  OSn[        R                  R                  UR                  U* U5        g g [        U[        R                  [        R                  45      (       aU  [        R                  R                  UR
                  S5        [        R                  R                  UR                  S5        g g )Nfan_outrelu)modenonlinearity   )ar   r   )r   r   Conv2dinitkaiming_normal_r:   Linearkaiming_uniform_ra   sqrtr'   _calculate_fan_in_and_fan_outuniform_BatchNorm2d	GroupNorm	constant_)r-   modulefan_inr   bounds        r   _init_weights BitPreTrainedModel._init_weights}  s   fbii((GG##FMM	PV#W		**GG$$V]]diil$C{{&GGAA&--P	17!DIIf--  ufe< '  >??GGfmmQ/GGfkk1- @r!   r   N)rC   rD   rE   rF   r   __annotations__base_model_prefixmain_input_name_no_split_modulesr6  rH   r   r!   r   r  r  v  s!    $O().r!   r  c            
       ^   ^  \ rS rSrU 4S jr\ S	S\S\\   S\\   S\	4S jj5       r
SrU =r$ )
BitModeli  c                 F  > [         TU ]  U5        Xl        [        U5      U l        [        U5      U l        UR                  S:X  a  [        XR                  S   S9O[        R                  " 5       U l        [        R                  " S5      U l        U R                  5         g )Nr   r4   r   )r   r   )r(   r)   rT   r   embedderr   encoderr   rL   r  r   rS   r   AdaptiveAvgPool2dr   	post_initr   s     r   r)   BitModel.__init__  s     %f-!&)   O3 #68K8KB8OP 		 **62r!   r   r  r  r   c                 H   Ub  UOU R                   R                  nUb  UOU R                   R                  nU R                  U5      nU R	                  XBUS9nUS   nU R                  U5      nU R                  U5      nU(       d	  Xg4USS  -   $ [        UUUR                  S9$ )Nr  r  r   r   )r  pooler_outputr   )	rT   r  use_return_dictr?  r@  r   r   r
   r   )r-   r   r  r  embedding_outputencoder_outputsr  pooled_outputs           r   r?   BitModel.forward  s    
 %9$D $++JjJj 	 &1%<k$++B]B]==6,,U` ' 
 ,A. II&78$56%58KKK7/')77
 	
r!   )rT   r?  r@  r   r   NN)rC   rD   rE   rF   r)   r   r   r   r  r
   r?   rH   rI   rJ   s   @r   r=  r=    sI    " os
"
:B4.
^fgk^l
	1
 
r!   r=  z
    BiT Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
    ImageNet.
    )custom_introc                      ^  \ rS rSrU 4S jr\    S
S\\R                     S\\R                     S\\
   S\\
   S\4
S jj5       rS	rU =r$ )BitForImageClassificationi  c                   > [         TU ]  U5        UR                  U l        [        U5      U l        [
        R                  " [
        R                  " 5       UR                  S:  a.  [
        R                  " UR                  S   UR                  5      O[
        R                  " 5       5      U l        U R                  5         g )Nr   r4   )r(   r)   
num_labelsr=  r   r   r   Flattenr+  r  rS   
classifierrB  r   s     r   r)   "BitForImageClassification.__init__  s      ++F#--JJLEKEVEVYZEZBIIf))"-v/@/@A`b`k`k`m

 	r!   r   labelsr  r  r   c                 J   Ub  UOU R                   R                  nU R                  XUS9nU(       a  UR                  OUS   nU R	                  U5      nSnUb  U R                  X'U R                   5      nU(       d  U4USS -   n	Ub  U4U	-   $ U	$ [        XUR                  S9$ )a  
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
    Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
    config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
NrE  r   r   )losslogitsr   )rT   rG  r   rF  rS  loss_functionr   r   )
r-   r   rU  r  r  outputsrJ  rX  rW  r   s
             r   r?   !BitForImageClassification.forward  s     &1%<k$++B]B]((<`k(l1<--'!*/%%fdkkBDY,F'+'7D7V#CVC3\c\q\qrrr!   )r   rS  rQ  )NNNN)rC   rD   rE   rF   r)   r   r   r   FloatTensor
LongTensorr  r   r?   rH   rI   rJ   s   @r   rO  rO    s    
  59-1/3&*su001s ))*s 'tn	s
 d^s 
.s sr!   rO  zL
    BiT backbone, to be used with frameworks like DETR and MaskFormer.
    c            
       b   ^  \ rS rSrSrU 4S jr\ S
S\S\\	   S\\	   S\
4S jj5       rS	rU =r$ )BitBackbonei  Fc                    > [         TU ]  U5        [         TU ]	  U5        [        U5      U l        UR
                  /UR                  -   U l        U R                  5         g rY   )	r(   r)   _init_backboner=  r   r   r  num_featuresrB  r   s     r   r)   BitBackbone.__init__  sQ     v&F##223f6I6II 	r!   r   r  r  r   c                    Ub  UOU R                   R                  nUb  UOU R                   R                  nU R                  USSS9nUR                  nSn[        U R                  5       H  u  pxXR                  ;   d  M  XeU   4-  nM      U(       d  U4n	U(       a  XR                  4-  n	U	$ [        UU(       a  UR                  SS9$ SSS9$ )a  
Examples:

```python
>>> from transformers import AutoImageProcessor, AutoBackbone
>>> import torch
>>> from PIL import Image
>>> import requests

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> processor = AutoImageProcessor.from_pretrained("google/bit-50")
>>> model = AutoBackbone.from_pretrained("google/bit-50")

>>> inputs = processor(image, return_tensors="pt")
>>> outputs = model(**inputs)
```NTrE  r   )feature_mapsr   
attentions)	rT   rG  r  r   r   r   stage_namesout_featuresr   )
r-   r   r  r  rZ  r   re  idxr  r   s
             r   r?   BitBackbone.forward  s    , &1%<k$++B]B]$8$D $++JjJj 	 ((<dPT(U--#D$4$45JC)))s!3 55 6 "_F#0022M%3G'//
 	
MQ
 	
r!   )r   rb  rL  )rC   rD   rE   rF   has_attentionsr)   r   r   r   r  r   r?   rH   rI   rJ   s   @r   r_  r_    sP     N os-
"-
:B4.-
^fgk^l-
	-
 -
r!   r_  )rO  r=  r  r_  )Nr   r   r   )r5   F)   )7rG   rv   ra   typingr   numpyr   r   r   r   activationsr   modeling_outputsr   r	   r
   r   modeling_utilsr   utilsr   r   utils.backbone_utilsr   configuration_bitr   
get_loggerrC   loggerr  r  r    r(  r#   r1  rL   Moduler*   	MaxPool2drs   r   r   r   r   r   r   r   r   r   r   r  r=  rO  r_  __all__r   r!   r   <module>rz     s   @       !  . , 1 ( 
		H	%&ERWY]R]L^ &R-ryy -`R\\ $0299 0f
2<< 
:/BII /fU\\ e T V[VbVb *%")) %A(bii A(HF FR'		 '.Gryy GTC
 C
L . . .* .
! .
 .
b +s 2 +s+s\ 
;
$m ;

;
| Yr!   