
    bCiB              	          S r SSKJr  SSKrSSK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  SSKJr  \R.                  " \5      rS+S\R4                  S\S\S\R4                  4S jjr " S S\R<                  5      r " S S\R@                  5      r! " S S\R<                  5      r" " S S\R<                  5      r# " S S\R<                  5      r$ " S S\R<                  5      r%\ " S S \5      5       r&\ " S! S"\&5      5       r'\" S#S$9 " S% S&\&5      5       r(\" S'S$9 " S( S)\&\5      5       r)/ S*Qr*g),zPyTorch ConvNext model.    )OptionalN)nn   )ACT2FN)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)BackboneMixin)can_return_tuple   )ConvNextConfiginput	drop_probtrainingreturnc                    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.
        r   r   )r   )dtypedevice)shapendimtorchrandr   r   floor_div)r   r   r   	keep_probr   random_tensoroutputs          h/home/james-whalen/.local/lib/python3.13/site-packages/transformers/models/convnext/modeling_convnext.py	drop_pathr$   (   s     CxII[[^

Q 77E

5ELL YYMYYy!M1FM    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$ )ConvNextDropPath=   zXDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).Nr   r   c                 .   > [         TU ]  5         Xl        g N)super__init__r   )selfr   	__class__s     r#   r,   ConvNextDropPath.__init__@   s    "r%   hidden_statesc                 B    [        XR                  U R                  5      $ r*   )r$   r   r   )r-   r0   s     r#   forwardConvNextDropPath.forwardD   s    FFr%   c                      SU R                    3$ )Nzp=r   )r-   s    r#   
extra_reprConvNextDropPath.extra_reprG   s    DNN#$$r%   r5   r*   )__name__
__module____qualname____firstlineno____doc__r   floatr,   r   Tensorr2   strr6   __static_attributes____classcell__r.   s   @r#   r'   r'   =   sQ    b#(5/ #T # #GU\\ Gell G%C % %r%   r'   c                   v   ^  \ rS rSrSrSSS.U 4S jjrS\R                  S\R                  4U 4S	 jjrS
r	U =r
$ )ConvNextLayerNormK   a5  LayerNorm that supports two data formats: channels_last (default) or channels_first.
The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shape (batch_size, height,
width, channels) while channels_first corresponds to inputs with shape (batch_size, channels, height, width).
ư>channels_lastepsdata_formatc                `   > [         TU ]  " U4SU0UD6  US;  a  [        SU 35      eX0l        g )NrI   )rG   channels_firstzUnsupported data format: )r+   r,   NotImplementedErrorrJ   )r-   normalized_shaperI   rJ   kwargsr.   s        r#   r,   ConvNextLayerNorm.__init__Q   s=    )=s=f=AA%(A+&OPP&r%   featuresr   c                    > U R                   S:X  a9  UR                  SSSS5      n[        TU ]  U5      nUR                  SSSS5      nU$ [        TU ]  U5      nU$ )zt
Args:
    features: Tensor of shape (batch_size, channels, height, width) OR (batch_size, height, width, channels)
rL   r      r   r   )rJ   permuter+   r2   )r-   rQ   r.   s     r#   r2   ConvNextLayerNorm.forwardW   sj    
 //''1a3Hwx0H''1a3H  wx0Hr%   rJ   r8   r9   r:   r;   r<   r,   r   r>   r2   r@   rA   rB   s   @r#   rD   rD   K   s9    
 15/ ' '   r%   rD   c                   f   ^  \ rS rSrSrU 4S jrS\R                  S\R                  4S jr	Sr
U =r$ )ConvNextEmbeddingse   zThis class is comparable to (and inspired by) the SwinEmbeddings class
found in src/transformers/models/swin/modeling_swin.py.
c                   > [         TU ]  5         [        R                  " UR                  UR
                  S   UR                  UR                  S9U l        [        UR
                  S   SSS9U l	        UR                  U l        g )Nr   kernel_sizestriderF   rL   rH   )
r+   r,   r   Conv2dnum_channelshidden_sizes
patch_sizepatch_embeddingsrD   	layernormr-   configr.   s     r#   r,   ConvNextEmbeddings.__init__j   sr     "		!4!4Q!7VEVEV_e_p_p!
 +6+>+>q+AtYij"//r%   pixel_valuesr   c                     UR                   S   nX R                  :w  a  [        S5      eU R                  U5      nU R	                  U5      nU$ )Nr   zeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)r   r`   
ValueErrorrc   rd   )r-   rh   r`   
embeddingss       r#   r2   ConvNextEmbeddings.forwardr   sT    #))!,,,,w  **<8
^^J/
r%   )rd   r`   rc   )r8   r9   r:   r;   r<   r,   r   FloatTensorr>   r2   r@   rA   rB   s   @r#   rY   rY   e   s/    0E$5$5 %,,  r%   rY   c                   j   ^  \ rS rSrSrSU 4S jjrS\R                  S\R                  4S jrSr	U =r
$ )	ConvNextLayer}   a  This corresponds to the `Block` class in the original implementation.

There are two equivalent implementations: [DwConv, LayerNorm (channels_first), Conv, GELU,1x1 Conv]; all in (N, C,
H, W) (2) [DwConv, Permute to (N, H, W, C), LayerNorm (channels_last), Linear, GELU, Linear]; Permute back

The authors used (2) as they find it slightly faster in PyTorch.

Args:
    config ([`ConvNextConfig`]): Model configuration class.
    dim (`int`): Number of input channels.
    drop_path (`float`): Stochastic depth rate. Default: 0.0.
c                    > [         TU ]  5         [        R                  " X"SSUS9U l        [        USS9U l        [        R                  " USU-  5      U l        [        UR                     U l        [        R                  " SU-  U5      U l        UR                  S:  a6  [        R                  " UR                  [        R                   " U5      -  SS	9OS U l        US
:  a  [%        U5      U l        g [        R&                  " 5       U l        g )N   r   )r]   paddinggroupsrF   rI      r   T)requires_gradr   )r+   r,   r   r_   dwconvrD   rd   Linearpwconv1r   
hidden_actactpwconv2layer_scale_init_value	Parameterr   oneslayer_scale_parameterr'   Identityr$   )r-   rf   dimr$   r.   s       r#   r,   ConvNextLayer.__init__   s    iia3O*3D9yya#g.&++,yyS#. ,,q0 LL66CHX\] 	"
 9BC))4R[[]r%   rQ   r   c                 b   UnU R                  U5      nUR                  SSSS5      nU R                  U5      nU R                  U5      nU R	                  U5      nU R                  U5      nU R                  b  U R                  U-  nUR                  SSSS5      nX R                  U5      -   nU$ )Nr   rS   r   r   )rx   rT   rd   rz   r|   r}   r   r$   )r-   rQ   residuals      r#   r2   ConvNextLayer.forward   s    ;;x(##Aq!Q/>>(+<<)88H%<<)%%111H<H##Aq!Q/nnX66r%   )r|   r$   rx   r   rd   rz   r}   )r   rW   rB   s   @r#   ro   ro   }   s.    [   r%   ro   c                   j   ^  \ rS rSrSrSU 4S jjrS\R                  S\R                  4S jrSr	U =r
$ )	ConvNextStage   a}  ConvNeXT stage, consisting of an optional downsampling layer + multiple residual blocks.

Args:
    config ([`ConvNextConfig`]): Model configuration class.
    in_channels (`int`): Number of input channels.
    out_channels (`int`): Number of output channels.
    depth (`int`): Number of residual blocks.
    drop_path_rates(`list[float]`): Stochastic depth rates for each layer.
c                   > [         T	U ]  5         X#:w  d  US:  a:  [        R                  " [	        USSS9[        R
                  " X#XES9/5      U l        O[        R                  " 5       U l        U=(       d    S/U-  n[        R                  " [        U5       Vs/ s H  n[        XXx   S9PM     sn5      U l	        g s  snf )Nr   rF   rL   rH   r\   r   )r   r$   )
r+   r,   r   
ModuleListrD   r_   downsampling_layerrangero   layers)
r-   rf   in_channelsout_channelsr]   r^   depthdrop_path_ratesjr.   s
            r#   r,   ConvNextStage.__init__   s    &&1*&(mm%ktIYZIIk[`'D# ')mmoD#):cUU]mm\abg\hi\hWX]6?QR\hi
is   B>rQ   r   c                 r    U R                    H  nU" U5      nM     U R                   H  nU" U5      nM     U$ r*   r   r   )r-   rQ   layers      r#   r2   ConvNextStage.forward   s7    ,,EXH -[[EXH !r%   r   )rS   rS   rS   NrW   rB   s   @r#   r   r      s-    
"   r%   r   c                   ^   ^  \ rS rSrU 4S jr SS\R                  S\\   S\	4S jjr
SrU =r$ )	ConvNextEncoder   c           
      .  > [         TU ]  5         [        R                  " 5       U l        [
        R                  " SUR                  [        UR                  5      SS9R                  UR                  5       Vs/ s H  nUR                  5       PM     nnUR                  S   n[        UR                  5       HT  nUR                  U   n[        UUUUS:  a  SOSUR                  U   X5   S9nU R                  R!                  U5        UnMV     g s  snf )Nr   cpu)r   rS   r   )r   r   r^   r   r   )r+   r,   r   r   stagesr   linspacedrop_path_ratesumdepthssplittolistra   r   
num_stagesr   append)	r-   rf   xr   prev_chsiout_chsstager.   s	           r#   r,   ConvNextEncoder.__init__   s    mmo ^^Av'<'<c&-->PY^_eeflfsfst
t HHJt 	 
 &&q)v(()A))!,G!$$EqqmmA& / 2E KKu%H *
s   9Dr0   output_hidden_statesr   c                     U(       a  U/OS nU R                    H!  nU" U5      nUc  M  UR                  U5        M#     [        XS9$ )N)last_hidden_stater0   )r   r   r   )r-   r0   r   all_hidden_stateslayer_modules        r#   r2   ConvNextEncoder.forward   sJ     0D]O KKL(7M ,!((7 (
 .oor%   )r   )F)r8   r9   r:   r;   r,   r   r>   r   boolr   r2   r@   rA   rB   s   @r#   r   r      s>    , SX
p"\\
pAI$
p	'
p 
pr%   r   c                   8    \ rS rSr% \\S'   SrSrS/r0 r	S r
Srg)	ConvNextPreTrainedModel   rf   convnextrh   ro   c                    [        U[        R                  [        R                  45      (       ak  UR                  R
                  R                  SU R                  R                  S9  UR                  b%  UR                  R
                  R                  5         gg[        U[        R                  [        45      (       aJ  UR                  R
                  R                  5         UR                  R
                  R                  S5        g[        U[        5      (       aH  UR                  b:  UR                  R
                  R                  U R                  R                   5        ggg)zInitialize the weightsr   )meanstdNg      ?)
isinstancer   ry   r_   weightdatanormal_rf   initializer_rangebiaszero_	LayerNormrD   fill_ro   r   r~   )r-   modules     r#   _init_weights%ConvNextPreTrainedModel._init_weights   s    fryy"))455 MM&&CT[[5R5R&S{{&  &&( '/@ ABBKK""$MM$$S)..++7,,11778Z8Z[ 8 /r%    N)r8   r9   r:   r;   r   __annotations__base_model_prefixmain_input_name_no_split_modules_can_record_outputsr   r@   r   r%   r#   r   r      s(    "$O()\r%   r   c            	       x   ^  \ rS rSrU 4S jr\\ SS\\R                     S\\
   S\4S jj5       5       rSrU =r$ )	ConvNextModeli  c                    > [         TU ]  U5        Xl        [        U5      U l        [        U5      U l        [        R                  " UR                  S   UR                  S9U l        U R                  5         g )Nru   )r+   r,   rf   rY   rk   r   encoderr   r   ra   layer_norm_epsrd   	post_initre   s     r#   r,   ConvNextModel.__init__	  s^     ,V4&v. f&9&9"&=6CXCXY 	r%   rh   r   r   c                    Uc  U R                   R                  nUc  [        S5      eU R                  U5      nU R	                  X2S9nUR
                  nU R                  UR                  SS/5      5      n[        UUUR                  S9$ )Nz You have to specify pixel_valuesr   r   )r   pooler_outputr0   )
rf   r   rj   rk   r   r   rd   r   r	   r0   )r-   rh   r   embedding_outputencoder_outputsr   pooled_outputs          r#   r2   ConvNextModel.forward  s    
  '#';;#C#C ?@@??<8:>,, ;G ;
 ,== '8'='=r2h'GH7/')77
 	
r%   )rf   rk   r   rd   NN)r8   r9   r:   r;   r,   r   r   r   r   rm   r   r	   r2   r@   rA   rB   s   @r#   r   r     sP     gk
$U%6%67
V^_cVd
	1
  
r%   r   z
    ConvNext Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
    ImageNet.
    )custom_introc            	          ^  \ rS rSrSrU 4S jr\\ S	S\\	R                     S\\	R                     S\4S jj5       5       rSrU =r$ )
ConvNextForImageClassificationi1  Fc                 B  > [         TU ]  U5        UR                  U l        [        U5      U l        UR                  S:  a4  [
        R                  " UR                  S   UR                  5      U l        O[
        R                  " 5       U l        U R                  5         g )Nr   r   )r+   r,   
num_labelsr   r   r   ry   ra   
classifierr   r   re   s     r#   r,   'ConvNextForImageClassification.__init__:  su      ++%f- q  ii(;(;B(?ARARSDO kkmDO 	r%   rh   labelsr   c                     U R                   " U40 UD6nUR                  nU R                  U5      nSnUb  U R                  X&U R                  S9n[        UUUR                  S9$ )ab  
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 regression loss is computed (Mean-Square loss), If
    `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
N)r   pooled_logitsrf   )losslogitsr0   )r   r   r   loss_functionrf   r
   r0   )r-   rh   r   rO   outputsr   r   r   s           r#   r2   &ConvNextForImageClassification.forwardI  su     =AMM,<aZ`<a--/%%VRVR]R]%^D3!//
 	
r%   )r   r   r   r   )r8   r9   r:   r;   accepts_loss_kwargsr,   r   r   r   r   rm   
LongTensorr
   r2   r@   rA   rB   s   @r#   r   r   1  s^       ei
$U%6%67
HPQVQaQaHb
	-
  
r%   r   zQ
    ConvNeXt backbone, to be used with frameworks like DETR and MaskFormer.
    c            	       v   ^  \ rS rSrSrU 4S jr\\ S	S\R                  S\
\   S\4S jj5       5       rSrU =r$ )
ConvNextBackboneic  Fc                   > [         TU ]  U5        [         TU ]	  U5        [        U5      U l        [        U5      U l        UR                  S   /UR                  -   U l        0 n[        U R                  U R                  5       H  u  p4[        USS9X#'   M     [        R                  " U5      U l        U R!                  5         g )Nr   rL   rV   )r+   r,   _init_backbonerY   rk   r   r   ra   num_featureszip_out_featureschannelsrD   r   
ModuleDicthidden_states_normsr   )r-   rf   r   r   r`   r.   s        r#   r,   ConvNextBackbone.__init__k  s     v&,V4&v.#0034v7J7JJ !#&t'9'94==#IE):<Ue)f& $J#%==1D#E  	r%   rh   r   r   c                 z   Uc  U R                   R                  nU R                  U5      nU R                  USS9nUR                  n/ n[        U R                  U5       H<  u  pxXpR                  ;   d  M  U R                  U   " U5      nUR                  U5        M>     [        [        U5      U(       a  US9$ 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("facebook/convnext-tiny-224")
>>> model = AutoBackbone.from_pretrained("facebook/convnext-tiny-224")

>>> inputs = processor(image, return_tensors="pt")
>>> outputs = model(**inputs)
```NTr   )feature_mapsr0   )rf   r   rk   r   r0   r   stage_namesout_featuresr   r   r   tuple)	r-   rh   r   r   r   r0   r   r   hidden_states	            r#   r2   ConvNextBackbone.forward|  s    2  '#';;#C#C ??<8<@LLIYptL<u--#&t'7'7#GE)))#77>|L##L1 $H
 |,+?-
 	
EI
 	
r%   )rk   r   r   r   r*   )r8   r9   r:   r;   has_attentionsr,   r   r   r   r>   r   r   r   r2   r@   rA   rB   s   @r#   r   r   c  sT     N"  04'
ll'
 'tn'
 
	'
  '
r%   r   )r   r   r   r   )r   F)+r<   typingr   r   r   activationsr   modeling_outputsr   r   r	   r
   modeling_utilsr   utilsr   r   utils.backbone_utilsr   utils.genericr   configuration_convnextr   
get_loggerr8   loggerr>   r=   r   r$   Moduler'   r   rD   rY   ro   r   r   r   r   r   r   __all__r   r%   r#   <module>r     sg       !  . , 1 - 2 
		H	%U\\ e T V[VbVb *%ryy % 4 0(BII (V!BII !H pbii  pF \o \ \. &
+ &
 &
R )
%< )
)
X 
=
. =

=
@ mr%   