
    h                        d dl Z d dlZd dlmZ d dlmZmZmZmZ d dl	Z
d dlZd dlmZ d dlmc m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%m&Z&m'Z' ddl(m)Z) ddl*m+Z+m,Z,m-Z- e e%d       G d de#                    Z.e e%d       G d de#                    Z/ee% G d de#                    Z0 G d dejb                        Z2	 dLdejb                  dejf                  dejf                  dejf                  deejf                     d e4d!e4fd"Z5 G d# d$ejb                        Z6 G d% d&ejb                        Z7 G d' d(e      Z8 G d) d*ejb                        Z9 G d+ d,ejb                        Z:d- Z;	 dMd.ejf                  d/e4d0e4d1e4d2e4d3ejf                  fd4Z<dNd5Z=d6 Z>d7 Z?e% G d8 d9e             Z@ G d: d;ejb                        ZA G d< d=ejb                        ZB e%d>       G d? d@e@             ZC G dA dBejb                        ZD e%dC       G dD dEe@             ZEe% G dF dGe@             ZF e%dH       G dI dJe@             ZGg dKZHy)O    N)	dataclass)AnyCallableOptionalUnion)_calculate_fan_in_and_fan_out   )ACT2FN)_prepare_4d_attention_mask)GradientCheckpointingLayer)BaseModelOutputBaseModelOutputWithPoolingImageClassifierOutput)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)ModelOutputTransformersKwargsauto_docstringcan_return_tuplefilter_out_non_signature_kwargs)check_model_inputs   )Siglip2ConfigSiglip2TextConfigSiglip2VisionConfigz}
    Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
    )custom_introc                       e Zd ZU dZdZeej                     ed<   dZ	eej                     ed<   dZ
eeej                  df      ed<   dZeeej                  df      ed<   y)Siglip2VisionOutputz
    image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
        The image embeddings obtained by applying the projection layer to the pooler_output.
    Nimage_embedslast_hidden_state.hidden_states
attentions)__name__
__module____qualname____doc__r    r   torchFloatTensor__annotations__r!   r"   tupler#        k/var/www/html/aiagenthome/venv/lib/python3.12/site-packages/transformers/models/siglip2/modeling_siglip2.pyr   r   +   sr    
 15L(5,,-459x 1 129=AM8E%"3"3S"89:A:>Ju00#567>r-   r   ze
    Base class for text model's outputs that also contains a pooling of the last hidden states.
    c                       e Zd ZU dZdZeej                     ed<   dZ	eej                     ed<   dZ
eeej                  df      ed<   dZeeej                  df      ed<   y)Siglip2TextOutputz
    text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
        The text embeddings obtained by applying the projection layer to the pooler_output.
    Ntext_embedsr!   .r"   r#   )r$   r%   r&   r'   r1   r   r(   r)   r*   r!   r"   r+   r#   r,   r-   r.   r0   r0   =   sr    
 04K%++,359x 1 129=AM8E%"3"3S"89:A:>Ju00#567>r-   r0   c                      e Zd ZU dZdZeej                     ed<   dZ	eej                     ed<   dZ
eej                     ed<   dZeej                     ed<   dZeej                     ed<   dZeed<   dZeed	<   d
ee   fdZy)Siglip2Outputa  
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
        Contrastive loss for image-text similarity.
    logits_per_image (`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
        The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
        similarity scores.
    logits_per_text (`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
        The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
        similarity scores.
    text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
        The text embeddings obtained by applying the projection layer to the pooled output of [`Siglip2TextModel`].
    image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim`):
        The image embeddings obtained by applying the projection layer to the pooled output of [`Siglip2VisionModel`].
    text_model_output (`BaseModelOutputWithPooling`):
        The output of the [`Siglip2TextModel`].
    vision_model_output (`BaseModelOutputWithPooling`):
        The output of the [`Siglip2VisionModel`].
    Nlosslogits_per_imagelogits_per_textr1   r    text_model_outputvision_model_outputreturnc                 H     t         fd j                         D              S )Nc              3   d   K   | ]'  }|d vr|   nt        |      j                          ) yw))r7   r8   N)getattrto_tuple).0kselfs     r.   	<genexpr>z)Siglip2Output.to_tuple.<locals>.<genexpr>n   s=      
   LLDGRYZ^`aRbRkRkRmm s   -0)r+   keysr@   s   `r.   r=   zSiglip2Output.to_tuplem   s#     
YY[
 
 	
r-   )r$   r%   r&   r'   r4   r   r(   r)   r*   r5   r6   r1   r    r7   r   r8   r+   r   r=   r,   r-   r.   r3   r3   O   s    & )-D(5$$
%,48hu001837OXe//07/3K%++,304L(5,,-448186:3:
%* 
r-   r3   c            	            e Zd Zdef fdZedej                  dej                  de	dej                  fd       Z
dej                  dej                  dej                  fd	Z xZS )
Siglip2VisionEmbeddingsconfigc                    t         |           || _        |j                  | _        |j
                  | _        t        j                  |j                  | j
                  z  | j
                  z  | j                        | _	        |j                  | _
        t        | j                  dz        | _        t        j                  | j                  | j                        | _        y )N)in_featuresout_featuresg      ?)super__init__rF   hidden_size	embed_dim
patch_sizennLinearnum_channelspatch_embeddingnum_patchesintposition_embedding_size	Embeddingposition_embeddingr@   rF   	__class__s     r.   rK   z Siglip2VisionEmbeddings.__init__u   s    ++ ++!yy++doo=O 

 "--'*4+;+;S+@'A$"$,,t/?/?"Pr-   positional_embeddingsspatial_shapes
max_lengthr9   c                 b   |j                   d   }| j                   d   }| j                  }t        j                  |||f| j                  |      }| j                  ddd      j                  d      } | j                  j                  dk(  r| j                  t        j                        } t        |      D ]w  }||   \  }}	t        j                  | ||	fddd	
      }
|
j                  |||	z        j                  dd      }
|
j                  |      }
|
||d||	z  f<   |
d   ||||	z  df<   y |S )ac  
        Resize positional embeddings to image-specific size and pad to a fixed size.

        Args:
            positional_embeddings (`torch.Tensor`):
                Position embeddings of shape (height, width, embed_dim)
            spatial_shapes (`torch.LongTensor`):
                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
            max_length (`int`):
                Maximum length of the positional embeddings to pad resized positional embeddings to

        Returns:
            `torch.Tensor`: Embeddings of shape (batch_size, max_length, embed_dim)
        r   )devicedtype   r   cpubilinearFT)sizemodealign_corners	antialiasN)shaper`   r(   emptyr_   permute	unsqueezetypetofloat32rangeFinterpolatereshape	transpose)rZ   r[   r\   
batch_sizerM   source_dtyperesulted_positional_embeddingsiheightwidthresized_embeddingss              r.   resize_positional_embeddingsz4Siglip2VisionEmbeddings.resize_positional_embeddings   s\   ( $))!,
)//3	,22).Y/(//*
& !6 = =aA F P PQR S !'',,5$9$<$<U]]$K!z"A*1-MFE!"%e_#" "4!;!;IvPU~!V!`!`abde!f "4!6!6|!DBT*1.>.>+>?BTUVBW*1fun.>+>?% #( .-r-   pixel_valuesc                 J   | j                   j                  j                  }| j                  |j                  |            }| j                  j                  j                  | j                  | j                  d      }| j                  |||j                  d         }||z   }|S )aH  
        Args:
            pixel_values (`torch.FloatTensor`):
                Pixel values of shape (batch_size, max_num_patches, num_channels * patch_size * patch_size)
            spatial_shapes (`list[tuple[int, int]]`):
                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
        )r`   r^   r   )r\   )	rR   weightr`   rm   rW   rr   rU   r{   rh   )r@   r|   r[   target_dtypepatch_embedsrZ   resized_positional_embeddings
embeddingss           r.   forwardzSiglip2VisionEmbeddings.forward   s     ++2288++LOO,O,OP !% 7 7 > > F F(($*F*F!
 )-(I(I!>l>P>PQR>S )J )
%
 "$AA
r-   )r$   r%   r&   r   rK   staticmethodr(   Tensor
LongTensorrT   r{   r)   r   __classcell__rY   s   @r.   rE   rE   t   s    Q2 Q 8.$||8.((8. 8. 
	8. 8.tE$5$5 uGWGW \a\h\h r-   rE   modulequerykeyvalueattention_maskscalingdropoutc                    t        j                  ||j                  dd            |z  }|||z   }t        j                  j                  |dt         j                        j                  |j                        }t        j                  j                  ||| j                        }t        j                  ||      }	|	j                  dd      j                         }	|	|fS )Nr^   )dimr`   )ptrainingr   ra   )r(   matmulrs   rO   
functionalsoftmaxrn   rm   r`   r   r   
contiguous)
r   r   r   r   r   r   r   kwargsattn_weightsattn_outputs
             r.   eager_attention_forwardr      s     <<s}}R'<=GL!#n4==((2U]](SVVW\WbWbcL==((6??([L,,|U3K''1-88:K$$r-   c            
            e Zd ZdZ fdZ	 ddej                  deej                     deej                  eej                     f   fdZ	 xZ
S )Siglip2Attentionz=Multi-headed attention from 'Attention Is All You Need' paperc                    t         |           || _        |j                  | _        |j
                  | _        | j                  | j                  z  | _        | j                  | j                  z  | j                  k7  r&t        d| j                   d| j                   d      | j                  dz  | _	        |j                  | _        d| _        t        j                  | j                  | j                        | _        t        j                  | j                  | j                        | _        t        j                  | j                  | j                        | _        t        j                  | j                  | j                        | _        y )Nz;embed_dim must be divisible by num_heads (got `embed_dim`: z and `num_heads`: z).      F)rJ   rK   rF   rL   rM   num_attention_heads	num_headshead_dim
ValueErrorscaleattention_dropoutr   	is_causalrO   rP   k_projv_projq_projout_projrX   s     r.   rK   zSiglip2Attention.__init__   s   ++33$..8==4>>)T^^;MdnnM] ^NN#2'  ]]D(
//ii?ii?ii?		$..$..Ar-   r"   r   r9   c           
      :   |j                   \  }}}| j                  |      }| j                  |      }| j                  |      }	|j	                  ||| j
                  | j                        j                  dd      }|j	                  ||| j
                  | j                        j                  dd      }|	j	                  ||| j
                  | j                        j                  dd      }	t        }
| j                  j                  dk7  rt        | j                  j                     }
 |
| |||	|| j                  | j                  | j                  sdn| j                        \  }}|j!                  |||      j#                         }| j%                  |      }||fS )z#Input shape: Batch x Time x Channelr   ra   eager        )r   r   r   )rh   r   r   r   viewr   r   rs   r   rF   _attn_implementationr   r   r   r   r   rr   r   r   )r@   r"   r   r   rt   
seq_lengthrM   queriesrB   valuesattention_interfacer   r   s                r.   r   zSiglip2Attention.forward  sa    -:,?,?)
J	++m,{{=)]+,,z:t~~t}}U__`acdeyyZOYYZ[]^_ZT^^T]]S]]^_abc(?;;++w6"9$++:Z:Z"[$7nnJJ#}}C$,,	%
!\ "))*j)LWWYmmK0L((r-   N)r$   r%   r&   r'   rK   r(   r   r   r+   r   r   r   s   @r.   r   r      sV    GB. 26$)||$) !.$)
 
u||Xell33	4$)r-   r   c                   V     e Zd Z fdZdej
                  dej
                  fdZ xZS )
Siglip2MLPc                    t         |           || _        t        |j                     | _        t        j                  |j                  |j                        | _
        t        j                  |j                  |j                        | _        y r   )rJ   rK   rF   r
   
hidden_actactivation_fnrO   rP   rL   intermediate_sizefc1fc2rX   s     r.   rK   zSiglip2MLP.__init__/  sd    #F$5$5699V//1I1IJ99V55v7I7IJr-   r"   r9   c                 l    | j                  |      }| j                  |      }| j                  |      }|S r   )r   r   r   )r@   r"   s     r.   r   zSiglip2MLP.forward6  s4    /**=9/r-   )r$   r%   r&   rK   r(   r   r   r   r   s   @r.   r   r   .  s$    KU\\ ell r-   r   c            	            e Zd Zdeeef   f fdZedej                  dej                  de
e   dej                  fd       Z xZS )Siglip2EncoderLayerrF   c                 D   t         |           |j                  | _        t	        j
                  | j                  |j                        | _        t        |      | _	        t	        j
                  | j                  |j                        | _
        t        |      | _        y Neps)rJ   rK   rL   rM   rO   	LayerNormlayer_norm_epslayer_norm1r   	self_attnlayer_norm2r   mlprX   s     r.   rK   zSiglip2EncoderLayer.__init__>  sm    ++<<F<Q<QR)&1<<F<Q<QRf%r-   r"   r   r   r9   c                     |}| j                  |      } | j                  d||d|\  }}||z   }|}| j                  |      }| j                  |      }||z   }|S )N)r"   r   r,   )r   r   r   r   )r@   r"   r   r   residual_s         r.   r   zSiglip2EncoderLayer.forwardF  s     !((7)4>> 
')
 
q
 !=0 ((7/ =0r-   )r$   r%   r&   r   r   r   rK   r   r(   r   r   r   r)   r   r   r   s   @r.   r   r   =  si    &u%8:K%KL & ||  +,	
 
		 r-   r   c                   j     e Zd ZdZdef fdZe	 ddeej                     de
e   defd       Z xZS )	Siglip2Encoderz
    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
    [`Siglip2EncoderLayer`].

    Args:
        config: Siglip2Config
    rF   c                     t         |           || _        t        j                  t        |j                        D cg c]  }t        |       c}      | _        d| _	        y c c}w )NF)
rJ   rK   rF   rO   
ModuleListro   num_hidden_layersr   layersgradient_checkpointing)r@   rF   r   rY   s      r.   rK   zSiglip2Encoder.__init__h  sV    mm%PVPhPhJi$jJiQ%8%@Ji$jk&+# %ks   A#r   r   r9   c                 T    |}| j                   D ]  } |||fi |} t        |      S )N)r!   )r   r   )r@   inputs_embedsr   r   r"   encoder_layers         r.   r   zSiglip2Encoder.forwardo  s>     &![[M) M ) ??r-   r   )r$   r%   r&   r'   r   rK   r   r   r(   r   r   r   r   r   r   r   s   @r.   r   r   _  s_    ,} ,  26@ !.@ +,	@
 
@ @r-   r   c                        e Zd Zdef fdZee	 	 d
dej                  dej                  dej                  dee   dee   defd	              Z xZS )Siglip2VisionTransformerrF   c                 L   t         |           || _        |j                  }t	        |      | _        t        |      | _        t        j                  ||j                        | _        t        |d      sdn|j                  | _        | j                  rt        |      | _        y y )Nr   vision_use_headT)rJ   rK   rF   rL   rE   r   r   encoderrO   r   r   post_layernormhasattrr   use_head$Siglip2MultiheadAttentionPoolingHeadheadr@   rF   rM   rY   s      r.   rK   z!Siglip2VisionTransformer.__init__  s    &&	1&9%f- ll9&:O:OP$+F4E$FFLbLb==<VDDI r-   r|   r   r[   output_attentionsoutput_hidden_statesr9   c                    ||n| j                   j                  }||n| j                   j                  }| j                  ||      }|0| j                   j                  dk7  rt        ||j                        }n|}| j                  ||||      }|j                  }	| j                  |	      }	| j                  r| j                  |	|      nd}
t        |	|
|j                  |j                        S )z
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.
        Nflash_attention_2)r   r   r   r   )r!   pooler_outputr"   r#   )rF   r   r   r   r   r   r`   r   r!   r   r   r   r   r"   r#   )r@   r|   r   r[   r   r   r"   encoder_attention_maskencoder_outputsr!   r   s              r.   r   z Siglip2VisionTransformer.forward  s     2C1N-TXT_T_TqTq$8$D $++JjJj 	 nE%$++*J*JNa*a%?P]PcPc%d"%3"+/<<'1/!5	 ,8 ,
 ,== //0ABHL		"3^D[_)/')77&11	
 	
r-   NN)r$   r%   r&   r   rK   r   r   r(   r)   r   r   r   boolr   r   r   r   s   @r.   r   r     s    
E2 
E  -1/3*
''*
 *
 ((	*

 $D>*
 'tn*
 
$*
  *
r-   r   c                    d }||d|z  z
  k  s||d|z  z   kD  rt        j                  dd        |||z
  |z        } |||z
  |z        }| j                  d|z  dz
  d|z  dz
         | j                          | j	                  |t        j                  d      z         | j                  |       | j                  ||       y )Nc                 d    dt        j                  | t        j                  d      z        z   dz  S )N      ?       @)matherfsqrt)xs    r.   norm_cdfz _trunc_normal_.<locals>.norm_cdf  s(    dhhq499S>122c99r-   ra   zjmean is more than 2 std from [a, b] in nn.init.trunc_normal_. The distribution of values may be incorrect.)
stacklevelr   r   )minmax)	warningswarnuniform_erfinv_mul_r   r   add_clamp_)tensormeanstdabr   lus           r.   _trunc_normal_r    s    : 	q1s7{q1s7{ 2;	
 	!d(c!"A!d(c!"A OOAEAIq1uqy) NN KKdiin$%
KK MMaQMr-   r   r   r  r  r  r9   c                     t        j                         5  t        | dd||       | j                  |      j	                  |       ddd       y# 1 sw Y   yxY w)an  Fills the input Tensor with values drawn from a truncated
    normal distribution. The values are effectively drawn from the
    normal distribution :math:`\mathcal{N}(	ext{mean}, 	ext{std}^2)`
    with values outside :math:`[a, b]` redrawn until they are within
    the bounds. The method used for generating the random values works
    best when :math:`a \leq 	ext{mean} \leq b`.

    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
    and the result is subsequently scaled and shifted by the mean and std args.

    Args:
        tensor: an n-dimensional `torch.Tensor`
        mean: the mean of the normal distribution
        std: the standard deviation of the normal distribution
        a: the minimum cutoff value
        b: the maximum cutoff value
    r   r   N)r(   no_gradr  r   r   )r   r   r  r  r  s        r.   trunc_normal_tf_r	    s>    * 
vq#q!,Cd# 
s   0AAc                 ,   t        |       \  }}|dk(  r|}n|dk(  r|}n|dk(  r||z   dz  }|z  }|dk(  r$t        | t        j                  |      dz         y |dk(  rCt	        j
                         5  | j                  t        j                  |             d d d        y |d	k(  rIt        j                  d
|z        }t	        j
                         5  | j                  | |       d d d        y t        d|       # 1 sw Y   y xY w# 1 sw Y   y xY w)Nfan_infan_outfan_avgra   truncated_normalg۶%?r  normaluniformr	   zinvalid distribution )	r   r	  r   r   r(   r  normal_r   r   )	r   r   re   distributionr  r  denomvariancebounds	            r.   variance_scaling_r    s    3F;OFGx				'!Q&u}H))TYYx%8;N%NO		!]]_NNtyy2N3 _		"		!h,']]_OOUFE* _ 0?@@ _ _s   3&C>D
>D
Dc                      t        | dd       y )Nr  r  re   r  r  r   s    r.   lecun_normal_r    s    f8:LMr-   c                      t        | dd       y )Nr  r  r  r  r  s    r.   default_flax_embed_initr    s    f8(Cr-   c                   H    e Zd ZU eed<   dZdZg dZdZdZ	dZ
dZeedZd Zy)Siglip2PreTrainedModelrF   siglip2T)Siglip2TextEmbeddingsrE   r   r   )r"   r#   c                    t        |t              rt        | j                  t              r | j                  j                  j
                  n| j                  j
                  }t        j                  j                  |j                  j                  dt        j                  |      z         yt        |t        j                        rt        |j                         yt        |t              rt        j                  j!                  |j"                  j                         t        j                  j!                  |j$                  j                         t        j                  j!                  |j&                  j                         t        j                  j!                  |j(                  j                         t        j                  j+                  |j"                  j,                         t        j                  j+                  |j$                  j,                         t        j                  j+                  |j&                  j,                         t        j                  j+                  |j(                  j,                         yt        |t.              rt        j                  j!                  |j0                  j                         t        j                  j!                  |j2                  j                         t        j                  j                  |j0                  j,                  d       t        j                  j                  |j2                  j,                  d       yt        |t4              rt        j                  j!                  |j6                  j8                         t        j                  j!                  |j:                  j<                  j8                         t        j                  j+                  |j:                  j>                  j8                         yt        |t@              rrtC        jD                  tC        jF                  d            }|jH                  j8                  jK                  |       |jL                  j8                  jO                          yt        |tP              rnt        j                  j                  |jR                  j                  | j                  j                  j
                  dz  | j                  jT                  z         yt        |t        jV                  t        jX                  f      rLt[        |j                         |j,                  *t        j                  j+                  |j,                         yyt        |t        j\                        rJ|j,                  j8                  jO                          |j                  j8                  jK                  d       yy)zInitialize the weightsr   r  gư>r   r   N)/
isinstancerE   rF   r   vision_configrL   rO   initr  rW   r~   npr   rV   r  r   xavier_uniform_r   r   r   r   zeros_biasr   r   r   r   probedata	attentionin_proj_weightin_proj_biasSiglip2Modelr(   logr   logit_scalefill_
logit_biaszero_Siglip2ForImageClassification
classifierinitializer_factorrP   Conv2dr  r   )r@   r   ry   logit_scale_inits       r.   _init_weightsz$Siglip2PreTrainedModel._init_weights2  s   f56 dkk=9 ))55[[,, 
 GGOOF55<<!bggenBTOU-#FMM2 01GG##FMM$8$89GG##FMM$8$89GG##FMM$8$89GG##FOO$:$:;GGNN6==--.GGNN6==--.GGNN6==--.GGNN6??//0
+GG##FJJ$5$56GG##FJJ$5$56GGOOFJJOOO6GGOOFJJOOO6 DEGG##FLL$5$56GG##F$4$4$C$C$H$HIGGNN6++88==>-$yyc):;##))*:;""((* =>GGOO!!((KK--994?$++B`B``   BII 67&--({{&v{{+ '-KK""$MM$$S) .r-   N)r$   r%   r&   r   r*   base_model_prefixsupports_gradient_checkpointing_no_split_modules_supports_flash_attn_supports_sdpa_supports_flex_attn_supports_attention_backendr   r   _can_record_outputsr;  r,   r-   r.   r   r     sJ    !&*#  N"& -&
,*r-   r   c            	            e Zd Zdef fdZ	 	 	 ddeej                     deej                     deej                     dej                  fdZ
 xZS )	r"  rF   c                 N   t         |           |j                  }t        j                  |j
                  |      | _        t        j                  |j                  |      | _        | j                  dt        j                  |j                        j                  d      d       y )Nposition_ids)r   r^   F)
persistent)rJ   rK   rL   rO   rV   
vocab_sizetoken_embeddingmax_position_embeddingsrW   register_bufferr(   arangeexpandr   s      r.   rK   zSiglip2TextEmbeddings.__init__b  s    &&	!||F,=,=yI"$,,v/M/My"Y 	ELL)G)GHOOPWXej 	 	
r-   	input_idsrF  r   r9   c                 8   ||j                   d   n|j                   d   }| j                  j                  j                   d   }||kD  rt        d| d|       || j                  d d d |f   }|| j                  |      }| j                  |      }||z   }|S )Nr^   r   r   zRSequence length must be less than max_position_embeddings (got `sequence length`: z and max_position_embeddings: )rh   rW   r~   r   rF  rI  )r@   rN  rF  r   r   max_position_embeddingposition_embeddingsr   s           r.   r   zSiglip2TextEmbeddings.forwardn  s     -6,AY__R(}GZGZ[]G^
!%!8!8!?!?!E!Ea!H..d,<=S<TV 
 ,,Q^<L  00;M"55lC"%88
r-   NNN)r$   r%   r&   r   rK   r   r(   r   r)   r   r   r   r   s   @r.   r"  r"  a  sk    

0 

 153759	E,,- u//0   1 12	
 
r-   r"  c                        e Zd Zdef fdZee	 	 	 d	deej                     deej                     deej                     de
e   def
d              Z xZS )
Siglip2TextTransformerrF   c                    t         |           || _        |j                  }t	        |      | _        t        |      | _        t        j                  ||j                        | _        t        j                  ||j                        | _        y r   )rJ   rK   rF   rL   r"  r   r   r   rO   r   r   final_layer_normrP   projection_sizer   r   s      r.   rK   zSiglip2TextTransformer.__init__  si    &&	/7%f- "YF<Q<Q RIIi)?)?@	r-   rN  r   rF  r   r9   c                    |t        d      |j                         }|j                  d|d         }| j                  ||      }d| j                  j
                  v }|rd }n||st        ||j                        } | j                  d||d|}|j                  }	| j                  |	      }	|	d d dd d f   }
| j                  |
      }
t        |	|
      S )NzYou have to specify input_idsr^   )rN  rF  flash)r   r   )r!   r   r,   )r   rd   r   r   rF   r   r   r`   r   r!   rV  r   r   )r@   rN  r   rF  r   input_shaper"   uses_flash_attentionr   r!   pooled_outputs              r.   r   zSiglip2TextTransformer.forward  s     <==nn&NN2{27	),W  '$++*J*JJ!N'0D7H[H[\N+74<< ,
'),
 ,
 ,== 112CD *!R(3		-0)/'
 	
r-   rR  )r$   r%   r&   r   rK   r   r   r   r(   r   r   r   r   r   r   r   s   @r.   rT  rT    s    A0 A  -115/3	(
ELL)(
 !.(
 u||,	(

 +,(
 
$(
  (
r-   rT  zL
    The text model from Siglip2 without any head or projection on top.
    c                        e Zd ZU eed<   def fdZdej                  fdZd Z	e
e	 	 	 ddeej                     deej                     deej                     d	ee   def
d
              Z xZS )Siglip2TextModelrF   c                 d    t         |   |       t        |      | _        | j	                          y r   )rJ   rK   rT  
text_model	post_initrX   s     r.   rK   zSiglip2TextModel.__init__  s&     08r-   r9   c                 B    | j                   j                  j                  S r   r`  r   rI  rC   s    r.   get_input_embeddingsz%Siglip2TextModel.get_input_embeddings  s    ))999r-   c                 :    || j                   j                  _        y r   rc  )r@   r   s     r.   set_input_embeddingsz%Siglip2TextModel.set_input_embeddings  s    5:""2r-   rN  r   rF  r   c                 .     | j                   d|||d|S )a  
        Examples:

        ```python
        >>> from transformers import AutoTokenizer, Siglip2TextModel

        >>> model = Siglip2TextModel.from_pretrained("google/siglip2-base-patch16-224")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/siglip2-base-patch16-224")

        >>> # important: make sure to set padding="max_length" as that's how the model was trained
        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt")

        >>> outputs = model(**inputs)
        >>> last_hidden_state = outputs.last_hidden_state
        >>> pooled_output = outputs.pooler_output  # pooled (EOS token) states
        ```rN  r   rF  r,   )r`  )r@   rN  r   rF  r   s        r.   r   zSiglip2TextModel.forward  s/    4 t 
)%
 	
 	
r-   rR  )r$   r%   r&   r   r*   rK   rO   Modulerd  rf  r   r   r   r(   r   r   r   r   r   r   r   s   @r.   r^  r^    s     0 :bii :;  -115/3	
ELL)
 !.
 u||,	

 +,
 
$
  
r-   r^  c                        e Zd ZdZdef fdZddej                  deej                     dej                  fdZ	 xZ
S )	r   zMultihead Attention Pooling.rF   c                    t         |           t        j                  t	        j
                  dd|j                              | _        t        j                  j                  |j                  |j                  d      | _
        t        j                  |j                  |j                        | _        t        |      | _        |j                  | _        y )Nr   T)batch_firstr   )rJ   rK   rO   	Parameterr(   randnrL   r+  MultiheadAttentionr   r-  r   r   	layernormr   r   r   rX   s     r.   rK   z-Siglip2MultiheadAttentionPoolingHead.__init__  s    \\%++aF4F4F"GH
44V5G5GIcIcqu4vf&8&8f>S>STf%33r-   hidden_stater   r9   c                    |j                   d   }| j                  j                  |dd      }|f|j                   d   |j                   d   }}t        ||j                  |      }|j                  d| j
                  |d      }|j                  d||      }| j                  ||||      d   }|}| j                  |      }|| j                  |      z   }|d d df   S )Nr   r   r^   )	attn_mask)
rh   r+  repeatr   r`   r   rr   r-  rp  r   )r@   rq  r   rt   r+  
target_len
source_lenr   s           r.   r   z,Siglip2MultiheadAttentionPoolingHead.forward  s    !''*


!!*a3%%*[[^\5G5G5J
J7HZHZ\fgN+221dnnjRSTN+33B
JON~~e\<Sa~bcde~~l3$((<"88AqD!!r-   r   )r$   r%   r&   r'   r   rK   r(   r   r   r   r   r   s   @r.   r   r     sA    &42 4"ELL "(5<<BX "didpdp "r-   r   zN
    The vision model from Siglip2 without any head or projection on top.
    c                        e Zd ZU eed<   dZdef fdZdej                  fdZ	e
e	 	 ddej                  dej                  dej                  dee   d	ee   defd
              Z xZS )Siglip2VisionModelrF   r|   c                 d    t         |   |       t        |      | _        | j	                          y r   )rJ   rK   r   vision_modelra  rX   s     r.   rK   zSiglip2VisionModel.__init__  s)     4V< 	r-   r9   c                 B    | j                   j                  j                  S r   )rz  r   rR   rC   s    r.   rd  z'Siglip2VisionModel.get_input_embeddings'  s      ++;;;r-   pixel_attention_maskr[   r   r   c                 .    | j                  |||||      S )a9  
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, Siglip2VisionModel

        >>> model = Siglip2VisionModel.from_pretrained("google/siglip2-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")

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

        >>> inputs = processor(images=image, return_tensors="pt")

        >>> outputs = model(**inputs)
        >>> last_hidden_state = outputs.last_hidden_state
        >>> pooled_output = outputs.pooler_output  # pooled features
        ```r|   r   r[   r   r   )rz  )r@   r|   r|  r[   r   r   s         r.   r   zSiglip2VisionModel.forward*  s,    F   %/)/!5 ! 
 	
r-   r   )r$   r%   r&   r   r*   main_input_namerK   rO   ri  rd  r   r   r(   r)   r   r   r   r   r   r   r   r   s   @r.   rx  rx    s      $O2 <bii <  -1/3'
'''
 $ll'
 ((	'

 $D>'
 'tn'
 
$'
  '
r-   rx  c                   n    e Zd ZU eed<   def fdZ e       e	 	 ddej                  de
ej                     de
ej                     dej                  fd              Z e       e	 	 	 dde
ej                     d	e
ej                     d
e
ej                     dej                  fd              Zee	 	 	 	 	 	 	 	 	 dde
ej                     de
ej                     d	e
ej                     d
e
ej                     de
ej                     de
ej                     de
e   de
e   de
e   defd              Z xZS )r0  rF   c                    t         |   |       t        |j                  t              s"t        dt        |j                         d      t        |j                  t              s"t        dt        |j                         d      |j                  }|j                  }t        j                  |      }t        j                  |      }|j                  | _        |j                  | _        t        j                  t!        j"                  d            | _        t        j                  t!        j"                  d            | _        | j)                          y )NzNconfig.text_config is expected to be of type Siglip2TextConfig but is of type .zRconfig.vision_config is expected to be of type Siglip2VisionConfig but is of type r   )rJ   rK   r$  text_configr   	TypeErrorrl   r%  r   r^  _from_configrx  r`  rz  rO   rm  r(   rn  r2  r4  ra  )r@   rF   r  r%  r`  rz  rY   s         r.   rK   zSiglip2Model.__init__Z  s    &,,.?@++,-Q0 
 &..0CD--./q2 
 ((,, &22;?
)66}E %//(55<<A7,,u{{1~6 	r-   rN  r   rF  r9   c                 F    | j                  |||      }|j                  }|S )aM  
        Returns:
            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
            applying the projection layer to the pooled output of [`Siglip2TextModel`].

        Examples:

        ```python
        >>> from transformers import AutoTokenizer, AutoModel
        >>> import torch

        >>> model = AutoModel.from_pretrained("google/siglip2-base-patch16-224")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/siglip2-base-patch16-224")

        >>> # important: make sure to set padding="max_length" as that's how the model was trained
        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding="max_length", return_tensors="pt")
        >>> with torch.no_grad():
        ...     text_features = model.get_text_features(**inputs)
        ```rh  )r`  r   )r@   rN  r   rF  text_outputsr\  s         r.   get_text_featureszSiglip2Model.get_text_featuresz  s4    6 48??)% 4C 4

 %22r-   r|   r|  r[   c                 F    | j                  |||      }|j                  }|S )a  
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.

        Returns:
            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
            applying the projection layer to the pooled output of [`Siglip2VisionModel`].

        Examples:

        ```python
        >>> import torch
        >>> from transformers import AutoProcessor, AutoModel
        >>> from transformers.image_utils import load_image

        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = load_image(url)

        >>> model = AutoModel.from_pretrained("google/siglip2-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")

        >>> inputs = processor(images=image, return_tensors="pt")

        >>> with torch.no_grad():
        ...     image_features = model.get_image_features(**inputs)
        ```
        )r|   r   r[   )rz  r   )r@   r|   r|  r[   vision_outputsr\  s         r.   get_image_featureszSiglip2Model.get_image_features  s7    J 6:5F5F%/) 6G 6

 '44r-   return_lossr   r   c
           	         ||n| j                   j                  }|	|	n| j                   j                  }	| j                  |||||	      }
| j	                  |||||	      }|
j
                  }|j
                  }||j                  ddd      z  }||j                  ddd      z  }t        j                  ||j                         j                  |j                              }| j                  j                  |j                        | j                  j                  |j                        }}||j                         z  |z   }|j                         }d}|rt        j                  |j!                  d      |j                  	      }t        j"                  |       d|z  z   }t        j$                  j&                  j)                  ||z        }t        j*                  |d
       }|j-                         }t/        |||||||
      S )a  
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.
        return_loss (`bool`, *optional*):
            Whether or not to return the contrastive loss.

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import AutoProcessor, AutoModel
        >>> import torch

        >>> model = AutoModel.from_pretrained("google/siglip2-base-patch16-224")
        >>> processor = AutoProcessor.from_pretrained("google/siglip2-base-patch16-224")

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

        >>> texts = ["a photo of 2 cats", "a photo of 2 dogs"]
        >>> # important: we pass `padding=max_length` since the model was trained with this
        >>> inputs = processor(text=texts, images=image, padding="max_length", return_tensors="pt")

        >>> with torch.no_grad():
        ...     outputs = model(**inputs)

        >>> logits_per_image = outputs.logits_per_image
        >>> probs = torch.sigmoid(logits_per_image) # these are the probabilities
        >>> print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'")
        31.9% that image 0 is 'a photo of 2 cats'
        ```
        Nr~  )rN  r   rF  r   r   ra   r^   T)r   r   keepdimr   )r_   r   )r4   r5   r6   r1   r    r7   r8   )rF   r   r   rz  r`  r   normr(   r   trm   r_   r2  r4  expeyerd   	ones_likerO   r   
logsigmoidsumr   r3   )r@   rN  r|   r|  r[   r   rF  r  r   r   r  r  r    r1   r6   r2  r4  r5   r4   r  m1_diag1logliknlls                          r.   r   zSiglip2Model.forward  s    d 2C1N-TXT_T_TqTq$8$D $++JjJj 	 6:5F5F%/)/!5 6G 6
 48??)%/!5 4C 4
 &33"00 $l&7&7!T&7&RR!K$4$4qb$$4$OO  ,,{LNN4D4G4GHZHZ4[\"&"2"2"5"5k6H6H"I4??K]K]^i^p^pKqZ)KOO,==
J*,,.))O003O<R<RSC881s7BHXX((33H4NOF99V,,C88:D-+#%* .
 	
r-   r   rR  )	NNNNNNNNN)r$   r%   r&   r   r*   rK   r   r   r(   r   r   r)   r  r   r  r   r   r3   r   r   r   s   @r.   r0  r0  V  s   } @ %& 26/3	 <<  !.  u||,	 
 
		   ' D %& 597;59	*u001* 'u||4* !!1!12	*
 
		*  '*Z  15487;591537&*,0/3e
E,,-e
 u001e
 'u||4	e

 !!1!12e
 !.e
 u//0e
 d^e
 $D>e
 'tne
 
e
  e
r-   r0  z
    Siglip2 vision encoder with an image classification head on top (a linear layer on top of the pooled final hidden states of
    the patch tokens) e.g. for ImageNet.
    c                        e Zd ZdZdeddf fdZee	 	 	 	 	 	 ddee	j                     dee	j                     dee	j                     dee	j                     d	ee   d
ee   defd              Z xZS )r6  r|   rF   r9   Nc                 ~   t         |   |       |j                  | _        t        j	                  |j
                        }|j                  | _        |j                  dkD  r4t        j                  |j
                  j                  |j                        nt        j                         | _        | j                          y )Nr   )rJ   rK   
num_labelsrx  r  r%  rz  rO   rP   rL   Identityr7  ra  )r@   rF   rz  rY   s      r.   rK   z&Siglip2ForImageClassification.__init__@  s      ++ *66v7K7KL(55 OUN_N_bcNcBIIf**668I8IJikititiv 	
 	r-   r|  r[   labelsr   r   c                 ,   ||n| j                   j                  }||n| j                   j                  }| j                  |||||      }|j                  }|Q|d   j                  |j                        }	t        j                  ||	z  d      t        j                  |	d      z  }nt        j                  |d      }| j                  |      }
d}|| j                  ||
| j                         }t        ||
|j                  |j                        S )a  
        pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
            Mask to avoid performing attention on padding pixel indices.
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width) of the input images.
        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).

        Examples:

        ```python
        >>> from transformers import AutoImageProcessor, Siglip2ForImageClassification
        >>> import torch
        >>> from PIL import Image
        >>> import requests

        >>> torch.manual_seed(3)  # doctest: +IGNORE_RESULT
        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
        >>> image = Image.open(requests.get(url, stream=True).raw)

        >>> # note: we are loading a `Siglip2Model` from the hub here,
        >>> # so the head will be randomly initialized, hence the predictions will be random if seed is not set above.
        >>> image_processor = AutoImageProcessor.from_pretrained("google/siglip2-base-patch16-224")
        >>> model = Siglip2ForImageClassification.from_pretrained("google/siglip2-base-patch16-224")

        >>> inputs = image_processor(images=image, return_tensors="pt")
        >>> outputs = model(**inputs)
        >>> logits = outputs.logits
        >>> # model predicts one of the two classes
        >>> predicted_class_idx = logits.argmax(-1).item()
        >>> print("Predicted class:", model.config.id2label[predicted_class_idx])
        Predicted class: LABEL_1
        ```
        N)r   r[   r   r   ).Nr   r  )r4   logitsr"   r#   )rF   r   r   rz  r!   rm   r_   r(   r  r   r7  loss_functionr   r"   r#   )r@   r|   r|  r[   r  r   r   outputssequence_output	pool_maskr  r4   s               r.   r   z%Siglip2ForImageClassification.forwardR  s!   ^ 2C1N-TXT_T_TqTq$8$D $++JjJj 	 /3.?.?/)/!5 /@ /
 "33  +,Y7::?;Q;QRI#ii)(CKeiiXaghNiiO#jja@O 1%%ffdkkBD$!//))	
 	
r-   )NNNNNN)r$   r%   r&   r  r   rK   r   r   r   r(   r   r   r   r   r   r   r   s   @r.   r6  r6  7  s     %O}  $  047;59)-,0/3O
u||,O
 'u||4O
 !!1!12	O

 &O
 $D>O
 'tnO
 
O
  O
r-   r6  )r0  r   r^  rx  r6  )r   )r   r   g       r   )r   r  r  )Ir   r   dataclassesr   typingr   r   r   r   numpyr'  r(   torch.nnrO   torch.nn.functionalr   rp   torch.nn.initr   activationsr
   modeling_attn_mask_utilsr   modeling_layersr   modeling_outputsr   r   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   r   r   utils.genericr   configuration_siglip2r   r   r   r   r0   r3   ri  rE   r   floatr   r   r   r   r   r   r  r	  r  r  r  r   r"  rT  r^  r   rx  r0  r6  __all__r,   r-   r.   <module>r     s  *   ! 1 1      7 ! B 9 b b F & w w / X X 
	?+ 	? 	? 
	? 	? 	?  
K  
   
Fbbii bX %II%<<% 
% <<	%
 U\\*% % %.;)ryy ;)| 4 D@RYY @D9
ryy 9
x! J \_$LL$ %$27$BG$SX$
\\$4A2ND A*_ A* A*H%BII %P5
RYY 5
p 
.
- .

.
b"299 "> 
8
/ 8

8
v ]
) ]
 ]
@ f
$: f
f
Rr-   