
    qiQ                     f   d dl Zd dl mZ d dlZd dlmZ ddlmZ ddl	m
Z
 ddlmZ ddlmZmZmZ ddlmZmZ dd	lmZ dd
lmZmZmZ ddlmZmZ ddlmZ ddlm Z   G d dejB                        Z" G d dejB                        Z#	 	 d5dejB                  dejH                  dejH                  dejH                  dejH                  dz  de%dz  de%dee   fdZ& G d dejB                        Z' G d dejB                        Z( G d  d!ejB                        Z) G d" d#ejB                        Z* G d$ d%ejB                        Z+ G d& d'e      Z,e G d( d)e             Z- G d* d+ejB                        Z. G d, d-ejB                        Z/e G d. d/e-             Z0 ed01       G d2 d3e-             Z1g d4Z2y)6    N)Callable   )initialization)ACT2FN)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPoolingImageClassifierOutput)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstring	torch_int)can_return_tuplemerge_with_config_defaults)capture_outputs   )IJepaConfigc                   f     e Zd ZdZdef fdZddej                  dedej                  fdZ	 xZ
S )	IJepaPatchEmbeddingsz
    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
    `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
    Transformer.
    configc                    t         |           |j                  |j                  }}|j                  |j
                  }}t        |t        j                  j                        r|n||f}t        |t        j                  j                        r|n||f}|d   |d   z  |d   |d   z  z  }|| _        || _        || _        || _
        t        j                  ||||      | _        y )Nr   r   )kernel_sizestride)super__init__
image_size
patch_sizenum_channelshidden_size
isinstancecollectionsabcIterablenum_patchesnnConv2d
projection)selfr   r   r   r    r!   r&   	__class__s          Z/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/ijepa/modeling_ijepa.pyr   zIJepaPatchEmbeddings.__init__    s    !'!2!2F4E4EJ
$*$7$79K9Kk#-j+//:R:R#SZZdfpYq
#-j+//:R:R#SZZdfpYq
!!}
15*Q-:VW=:XY$$(&))L+:^hi    pixel_valuesinterpolate_pos_encodingreturnc                    |j                   \  }}}}|| j                  k7  rt        d| j                   d| d      |sV|| j                  d   k7  s|| j                  d   k7  r2t        d| d| d| j                  d    d| j                  d    d		      | j	                  |      j                  d
      j                  dd
      }|S )NzoMake sure that the channel dimension of the pixel values match with the one set in the configuration. Expected z	 but got .r   r   zInput image size (*z) doesn't match model (z).   )shaper    
ValueErrorr   r)   flatten	transpose)r*   r.   r/   
batch_sizer    heightwidth
embeddingss           r,   forwardzIJepaPatchEmbeddings.forward/   s    2>2D2D/
L&%4,,,!../yaI  (++u8J/J (% 9+,Adooa.@-AE  __\2::1=GG1M
r-   F)__name__
__module____qualname____doc__r   r   torchTensorboolr=   __classcell__r+   s   @r,   r   r      s;    j{ jELL D ]b]i]i r-   r   c            	            e Zd ZdZddededdf fdZdej                  de	d	e	dej                  fd
Z
	 	 ddej                  dej                  dz  dedej                  fdZ xZS )IJepaEmbeddingszb
    Construct the CLS token, position and patch embeddings. Optionally, also the mask token.
    r   use_mask_tokenr0   Nc                    t         |           |r4t        j                  t	        j
                  dd|j                              nd | _        t        |      | _	        | j                  j                  }t        j                  t	        j                  d||j                              | _        t        j                  |j                        | _        |j                   | _        || _        y )Nr   )r   r   r'   	ParameterrC   zerosr!   
mask_tokenr   patch_embeddingsr&   randnposition_embeddingsDropouthidden_dropout_probdropoutr   r   )r*   r   rJ   r&   r+   s       r,   r   zIJepaEmbeddings.__init__E   s    Q_",,u{{1a9K9K'LMei 4V <++77#%<<A{FL^L^0_#` zz&"<"<= ++r-   r<   r:   r;   c                 0   |j                   d   }| j                  j                   d   }t        j                  j	                         s||k(  r||k(  r| j                  S | j                  }|j                   d   }|| j
                  z  }|| j
                  z  }	t        |dz        }
|j                  d|
|
|      }|j                  dddd      }t        j                  j                  |||	fdd	      }|j                  dddd      j                  dd|      }|S )
a   
        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
        images. This method is also adapted to support torch.jit tracing.

        Adapted from:
        - https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174-L194, and
        - https://github.com/facebookresearch/dinov2/blob/e1277af2ba9496fbadf7aec6eba56e8d882d1e35/dinov2/models/vision_transformer.py#L179-L211
        r   g      ?r   r   r4   bicubicF)sizemodealign_corners)r5   rQ   rC   jit
is_tracingr   r   reshapepermuter'   
functionalinterpolateview)r*   r<   r:   r;   r&   num_positionspatch_pos_embeddim
new_height	new_widthsqrt_num_positionss              r,   r/   z(IJepaEmbeddings.interpolate_pos_encodingO   s#    !&&q)0066q9 yy##%+*F6UZ?+++22r"t.
T__,	&}c'9:)11!5GI[]`a)11!Q1=--33i(	 4 
 *11!Q1=BB1b#Nr-   r.   bool_masked_posr/   c                 x   |j                   \  }}}}| j                  ||      }|Z|j                   d   }	| j                  j                  ||	d      }
|j	                  d      j                  |
      }|d|z
  z  |
|z  z   }|r|| j                  |||      z   }n|| j                  z   }| j                  |      }|S )N)r/   r   rV   g      ?)	r5   rO   rN   expand	unsqueezetype_asr/   rQ   rT   )r*   r.   rh   r/   r9   _r:   r;   r<   
seq_lengthmask_tokensmasks               r,   r=   zIJepaEmbeddings.forwardv   s     (4'9'9$
Avu**<Rj*k
&#))!,J//00ZLK",,R088ED#sTz2[45GGJ $#d&C&CJPVX]&^^J#d&>&>>J\\*-
r-   r>   NF)r?   r@   rA   rB   r   rE   r   rC   rD   intr/   
BoolTensorr=   rF   rG   s   @r,   rI   rI   @   s    { D T %5<< % %UX %]b]i]i %T 48).	ll ))D0 #'	
 
r-   rI   modulequerykeyvalueattention_maskscalingrT   kwargsc                    ||j                  d      dz  }t        j                  ||j                  dd            |z  }|||z   }t        j
                  j                  |d      }t        j
                  j                  ||| j                        }t        j                  ||      }	|	j                  dd      j                         }	|	|fS )NrV         r4   r   rd   )ptrainingr   )
rX   rC   matmulr8   r'   r_   softmaxrT   r   
contiguous)
rt   ru   rv   rw   rx   ry   rT   rz   attn_weightsattn_outputs
             r,   eager_attention_forwardr      s     **R.D( <<s}}Q':;gEL!#n4==((2(>L==((6??([L,,|U3K''1-88:K$$r-   c                   z     e Zd Zdef fdZdej                  deej                  ej                  f   fdZ xZ	S )IJepaSelfAttentionr   c                 2   t         |           |j                  |j                  z  dk7  r2t	        |d      s&t        d|j                   d|j                   d      || _        |j                  | _        t        |j                  |j                  z        | _        | j                  | j                  z  | _	        |j                  | _        | j                  dz  | _        d| _        t        j                  |j                  | j                  |j                         | _        t        j                  |j                  | j                  |j                         | _        t        j                  |j                  | j                  |j                         | _        y )	Nr   embedding_sizezThe hidden size z4 is not a multiple of the number of attention heads r2   r|   F)bias)r   r   r!   num_attention_headshasattrr6   r   rr   attention_head_sizeall_head_sizeattention_probs_dropout_probdropout_probry   	is_causalr'   Linearqkv_biasru   rv   rw   r*   r   r+   s     r,   r   zIJepaSelfAttention.__init__   sF    : ::a?PVXhHi"6#5#5"6 7334A7 
 #)#=#= #&v'9'9F<V<V'V#W !558P8PP"??//5YYv1143E3EFOO\
99V//1C1C&//ZYYv1143E3EFOO\
r-   hidden_statesr0   c           
         |j                   d   }|d| j                  | j                  f} | j                  |      j                  | j                  dd      } | j                  |      j                  | j                  dd      } | j                  |      j                  | j                  dd      }t        j                  | j                  j                  t              } || |||d | j                  | j                  | j                  sdn| j                         \  }}	|j#                         d d | j$                  fz   }
|j'                  |
      }||	fS )Nr   rV   r   r4           )r   ry   rT   )r5   r   r   rv   ra   r8   rw   ru   r   get_interfacer   _attn_implementationr   r   ry   r   r   rX   r   r]   )r*   r   r9   	new_shape	key_layervalue_layerquery_layerattention_interfacecontext_layerattention_probsnew_context_layer_shapes              r,   r=   zIJepaSelfAttention.forward   sF   "((+
D$<$<d>V>VV	0DHH]+00)<FFq!L	4djj/44i@JJ1aP4djj/44i@JJ1aP(?(M(MKK,,.E)
 *=nnLL#}}C$2C2C	*
& #0"4"4"6s";t?Q?Q>S"S%--.EFo--r-   )
r?   r@   rA   r   r   rC   rD   tupler=   rF   rG   s   @r,   r   r      s:    ]{ ](.U\\ .eELL%,,<V6W .r-   r   c                   x     e Zd ZdZdef fdZdej                  dej                  dej                  fdZ xZ	S )IJepaSelfOutputz
    The residual connection is defined in IJepaLayer instead of here (as is the case with other models), due to the
    layernorm applied before each block.
    r   c                     t         |           t        j                  |j                  |j                        | _        t        j                  |j                        | _        y N)	r   r   r'   r   r!   denserR   rS   rT   r   s     r,   r   zIJepaSelfOutput.__init__   sB    YYv1163E3EF
zz&"<"<=r-   r   input_tensorr0   c                 J    | j                  |      }| j                  |      }|S r   r   rT   r*   r   r   s      r,   r=   zIJepaSelfOutput.forward   s$    

=1]3r-   
r?   r@   rA   rB   r   r   rC   rD   r=   rF   rG   s   @r,   r   r      s=    
>{ >
U\\  RWR^R^ r-   r   c                   \     e Zd Zdef fdZdej                  dej                  fdZ xZS )IJepaAttentionr   c                 b    t         |           t        |      | _        t	        |      | _        y r   )r   r   r   	attentionr   outputr   s     r,   r   zIJepaAttention.__init__   s&    +F3%f-r-   r   r0   c                 R    | j                  |      \  }}| j                  ||      }|S r   )r   r   )r*   r   self_attn_outputrm   r   s        r,   r=   zIJepaAttention.forward   s,    "nn];!-}=r-   	r?   r@   rA   r   r   rC   rD   r=   rF   rG   s   @r,   r   r      s*    .{ .
U\\ ell r-   r   c                   \     e Zd Zdef fdZdej                  dej                  fdZ xZS )IJepaIntermediater   c                    t         |           t        j                  |j                  |j
                        | _        t        |j                  t              rt        |j                     | _        y |j                  | _        y r   )r   r   r'   r   r!   intermediate_sizer   r"   
hidden_actstrr   intermediate_act_fnr   s     r,   r   zIJepaIntermediate.__init__   s]    YYv1163K3KL
f''-'-f.?.?'@D$'-'8'8D$r-   r   r0   c                 J    | j                  |      }| j                  |      }|S r   )r   r   )r*   r   s     r,   r=   zIJepaIntermediate.forward  s&    

=100?r-   r   rG   s   @r,   r   r      s*    9{ 9U\\ ell r-   r   c                   t     e Zd Zdef fdZdej                  dej                  dej                  fdZ xZS )IJepaOutputr   c                     t         |           t        j                  |j                  |j
                        | _        t        j                  |j                        | _	        y r   )
r   r   r'   r   r   r!   r   rR   rS   rT   r   s     r,   r   zIJepaOutput.__init__  sB    YYv779K9KL
zz&"<"<=r-   r   r   r0   c                 T    | j                  |      }| j                  |      }||z   }|S r   r   r   s      r,   r=   zIJepaOutput.forward  s.    

=1]3%4r-   r   rG   s   @r,   r   r     s8    >{ >
U\\  RWR^R^ r-   r   c                   `     e Zd ZdZdef fdZdej                  dej                  fdZ xZ	S )
IJepaLayerz?This corresponds to the Block class in the timm implementation.r   c                 r   t         |           |j                  | _        d| _        t	        |      | _        t        |      | _        t        |      | _	        t        j                  |j                  |j                        | _        t        j                  |j                  |j                        | _        y )Nr   eps)r   r   chunk_size_feed_forwardseq_len_dimr   r   r   intermediater   r   r'   	LayerNormr!   layer_norm_epslayernorm_beforelayernorm_afterr   s     r,   r   zIJepaLayer.__init__  s    '-'E'E$'/-f5!&) "V-?-?VEZEZ [!||F,>,>FDYDYZr-   r   r0   c                     | j                  |      }| j                  |      }||z   }| j                  |      }| j                  |      }| j	                  ||      }|S r   )r   r   r   r   r   )r*   r   hidden_states_normattention_outputlayer_outputs        r,   r=   zIJepaLayer.forward%  si    !22=A>>*<= )=8 ++M:((6 {{<?r-   r   rG   s   @r,   r   r     s/    I[{ [U\\ ell r-   r   c                       e Zd ZU eed<   dZdZdZdZddgZ	dZ
dZdZdZeedZ ej$                         d	ej(                  ej*                  z  ej,                  z  d
dfd       Zy)IJepaPreTrainedModelr   ijepar.   )imageTrI   r   )r   
attentionsrt   r0   Nc                    t        |t        j                  t        j                  f      rct	        j
                  |j                  d| j                  j                         |j                   t	        j                  |j                         yyt        |t        j                        r?t	        j                  |j                         t	        j                  |j                         yt        |t              rct	        j
                  |j                  d| j                  j                         |j                   t	        j                  |j                         yyy)zInitialize the weightsr   )meanstdN)r"   r'   r   r(   inittrunc_normal_weightr   initializer_ranger   zeros_r   ones_rI   rQ   rN   )r*   rt   s     r,   _init_weightsz"IJepaPreTrainedModel._init_weightsG  s     fryy"))45v}}3DKK<Y<YZ{{&FKK( '-KK$JJv}}%0v99IfIfg  ,F--. - 1r-   )r?   r@   rA   r   __annotations__base_model_prefixmain_input_nameinput_modalitiessupports_gradient_checkpointing_no_split_modules_supports_sdpa_supports_flash_attn_supports_flex_attn_supports_attention_backendr   r   _can_record_outputsrC   no_gradr'   r   r(   r   r    r-   r,   r   r   6  s    $O!&*#*L9N"&#(
 U]]_/BII		$9BLL$H /T / /r-   r   c                   H     e Zd Zdef fdZdej                  defdZ xZ	S )IJepaEncoderr   c                     t         |           || _        t        j                  t        |j                        D cg c]  }t        |       c}      | _        d| _	        y c c}w rq   )
r   r   r   r'   
ModuleListrangenum_hidden_layersr   layergradient_checkpointing)r*   r   rm   r+   s      r,   r   zIJepaEncoder.__init__X  sN    ]]fF^F^@_#`1Jv$6#`a
&+# $as   A#r   r0   c                 d    t        | j                        D ]  \  }} ||      } t        |      S )N)last_hidden_state)	enumerater   r   )r*   r   ilayer_modules       r,   r=   zIJepaEncoder.forward^  s5    (4 	8OA|(7M	8 ??r-   )
r?   r@   rA   r   r   rC   rD   r   r=   rF   rG   s   @r,   r   r   W  s)    ,{ ,@U\\ @o @r-   r   c                   \     e Zd Zdef fdZdej                  dej                  fdZ xZS )IJepaPoolerr   c                     t         |           t        j                  |j                  |j
                        | _        t        |j                     | _	        y r   )
r   r   r'   r   r!   pooler_output_sizer   r   
pooler_act
activationr   s     r,   r   zIJepaPooler.__init__f  s>    YYv1163L3LM
 !2!23r-   r   r0   c                 \    |d d df   }| j                  |      }| j                  |      }|S )Nr   )r   r   )r*   r   first_token_tensorpooled_outputs       r,   r=   zIJepaPooler.forwardk  s6     +1a40

#566r-   r   rG   s   @r,   r   r   e  s*    4{ 4
U\\ ell r-   r   c                        e Zd Zddededef fdZdefdZe e	d      e
	 	 	 dd
ej                  d	z  dej                  d	z  ded	z  dee   def
d                     Z xZS )
IJepaModelFr   add_pooling_layerrJ   c                    t         |   |       || _        t        ||      | _        t        |      | _        t        j                  |j                  |j                        | _        |rt        |      nd| _        | j                          y)z
        add_pooling_layer (bool, *optional*, defaults to `True`):
            Whether to add a pooling layer
        use_mask_token (`bool`, *optional*, defaults to `False`):
            Whether to use a mask token for masked image modeling.
        )rJ   r   N)r   r   r   rI   r<   r   encoderr'   r   r!   r   	layernormr   pooler	post_init)r*   r   r  rJ   r+   s       r,   r   zIJepaModel.__init__v  sm     	 )&P#F+f&8&8f>S>ST->k&)D 	r-   r0   c                 .    | j                   j                  S r   )r<   rO   )r*   s    r,   get_input_embeddingszIJepaModel.get_input_embeddings  s    ///r-   )tie_last_hidden_statesNr.   rh   r/   rz   c                    |t        d      | j                  j                  j                  j                  j
                  }|j
                  |k7  r|j                  |      }| j                  |||      }| j                  |      }|j                  }| j                  |      }| j                  | j                  |      nd}	t        ||	      S )z
        bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`, *optional*):
            Boolean masked positions. Indicates which patches are masked (1) and which aren't (0).
        Nz You have to specify pixel_values)rh   r/   )r   pooler_output)r6   r<   rO   r)   r   dtypetor  r   r	  r
  r	   )
r*   r.   rh   r/   rz   expected_dtypeembedding_outputencoder_outputssequence_outputr  s
             r,   r=   zIJepaModel.forward  s     ?@@ 99DDKKQQ/'??>:L??/Tl + 
 ,0<<8H+I);;..98<8OO4UY)O[hiir-   )FFNNN)r?   r@   rA   r   rE   r   r   r  r   r   r   rC   rD   rs   r   r   r	   r=   rF   rG   s   @r,   r  r  t  s    { t ]a $0&: 0  E2 -13704	jllT)j ))D0j #'+	j
 +,j 
$j  3  jr-   r  a  
    IJepa Model transformer with an image classification head on top (a linear layer on top of the final hidden states)
    e.g. for ImageNet.

    <Tip>

        Note that it's possible to fine-tune IJepa on higher resolution images than the ones it has been trained on, by
        setting `interpolate_pos_encoding` to `True` in the forward of the model. This will interpolate the pre-trained
        position embeddings to the higher resolution.

    </Tip>
    )custom_introc                        e Zd Zdef fdZee	 	 	 d
dej                  dz  dej                  dz  de	dz  de
e   def
d	              Z xZS )IJepaForImageClassificationr   c                 .   t         |   |       |j                  | _        t        |d      | _        |j                  dkD  r*t        j                  |j                  |j                        nt        j                         | _	        | j                          y )NF)r  r   )r   r   
num_labelsr  r   r'   r   r!   Identity
classifierr  r   s     r,   r   z$IJepaForImageClassification.__init__  ss      ++%@
 OUN_N_bcNc"))F$6$68I8IJikititiv 	r-   Nr.   labelsr/   rz   r0   c                     | j                   |fd|i|}|j                  }| j                  |j                  d            }d}| | j                  ||| j
                  fi |}t        |||j                  |j                        S )a  
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        r/   r   r}   N)losslogitsr   r   )	r   r   r  r   loss_functionr   r
   r   r   )	r*   r.   r  r/   rz   outputsr  r"  r!  s	            r,   r=   z#IJepaForImageClassification.forward  s      /9djj/
%=/
 /

 "33!5!5!!5!<=%4%%ffdkkLVLD$!//))	
 	
r-   r  )r?   r@   rA   r   r   r   r   rC   rD   rE   r   r   r
   r=   rF   rG   s   @r,   r  r    s    
{ 
  -1&*04	
llT)
 t#
 #'+	

 +,
 

  
r-   r  )r   r  r  )Nr   )3collections.abcr#   r   rC   torch.nnr'    r   r   activationsr   modeling_layersr   modeling_outputsr   r	   r
   modeling_utilsr   r   processing_utilsr   utilsr   r   r   utils.genericr   r   utils.output_capturingr   configuration_ijepar   Moduler   rI   rD   floatr   r   r   r   r   r   r   r   r   r   r  r  __all__r   r-   r,   <module>r4     s    $   & ! 9 b b F & B B I 5 ,$299 $NNbii Nn !%II%<<% 
% <<	%
 LL4'% T\% % '(%8/. /.dbii "	RYY 			 
")) 
+ < /? / /@@299 @"))  7j% 7j 7jt .
"6 .
.
b Pr-   