o
    	۷ib                     @   s  d Z ddlZddlmZ ddlmZmZ ddlZddlmZ ddl	m
Z
mZmZmZ ddlmZ dd	lmZmZ d
dlmZ eeZeeddG dd deZG dd dejZG dd dejZG dd dejZG dd dejZG dd dejZG dd dejZG dd dejZ G dd dejZ!G d d! d!ejZ"G d"d# d#ejZ#G d$d% d%ejZ$eG d&d' d'eZ%eG d(d) d)e%Z&ed*dG d+d, d,e%Z'ed-dG d.d/ d/e%Z(g d0Z)dS )1zPyTorch LeViT model.    N)	dataclass)OptionalUnion)nn   )BaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttentionModelOutput)PreTrainedModel)auto_docstringlogging   )LevitConfigzD
    Output type of [`LevitForImageClassificationWithTeacher`].
    )custom_introc                   @   s^   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ej  ed< dS ),LevitForImageClassificationWithTeacherOutputan  
    logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
        Prediction scores as the average of the `cls_logits` and `distillation_logits`.
    cls_logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
        Prediction scores of the classification head (i.e. the linear layer on top of the final hidden state of the
        class token).
    distillation_logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
        Prediction scores of the distillation head (i.e. the linear layer on top of the final hidden state of the
        distillation token).
    Nlogits
cls_logitsdistillation_logitshidden_states)__name__
__module____qualname____doc__r   r   torchFloatTensor__annotations__r   r   r   tuple r   r   ^/home/ubuntu/vllm_env/lib/python3.10/site-packages/transformers/models/levit/modeling_levit.pyr   &   s   
 r   c                       s,   e Zd ZdZ	d fdd	Zdd Z  ZS )LevitConvEmbeddingsz[
    LeViT Conv Embeddings with Batch Norm, used in the initial patch embedding layer.
    r   c	           	   
      s6   t    tj|||||||dd| _t|| _d S )NF)dilationgroupsbias)super__init__r   Conv2dconvolutionBatchNorm2d
batch_norm)	selfin_channelsout_channelskernel_sizestridepaddingr!   r"   bn_weight_init	__class__r   r   r%   C   s
   
zLevitConvEmbeddings.__init__c                 C   s   |  |}| |}|S N)r'   r)   )r*   
embeddingsr   r   r   forwardL      

zLevitConvEmbeddings.forward)r   r   r   r   r   r   r   r%   r5   __classcell__r   r   r1   r   r    >   s
    	r    c                       (   e Zd ZdZ fddZdd Z  ZS )LevitPatchEmbeddingsz
    LeViT patch embeddings, for final embeddings to be passed to transformer blocks. It consists of multiple
    `LevitConvEmbeddings`.
    c                    s   t    t|j|jd d |j|j|j| _t	
 | _t|jd d |jd d |j|j|j| _t	
 | _t|jd d |jd d |j|j|j| _t	
 | _t|jd d |jd |j|j|j| _|j| _d S )Nr            )r$   r%   r    num_channelshidden_sizesr-   r.   r/   embedding_layer_1r   	Hardswishactivation_layer_1embedding_layer_2activation_layer_2embedding_layer_3activation_layer_3embedding_layer_4r*   configr1   r   r   r%   X   s"   

$
$
 zLevitPatchEmbeddings.__init__c                 C   st   |j d }|| jkrtd| |}| |}| |}| |}| |}| |}| 	|}|
dddS )Nr   zeMake sure that the channel dimension of the pixel values match with the one set in the configuration.r=   )shaper>   
ValueErrorr@   rB   rC   rD   rE   rF   rG   flatten	transpose)r*   pixel_valuesr>   r4   r   r   r   r5   n   s   








zLevitPatchEmbeddings.forwardr7   r   r   r1   r   r:   R   s    r:   c                       s&   e Zd Zd fdd	Zdd Z  ZS )MLPLayerWithBNr   c                    s,   t    tj||dd| _t|| _d S )NF)in_featuresout_featuresr#   )r$   r%   r   LinearlinearBatchNorm1dr)   )r*   	input_dim
output_dimr0   r1   r   r   r%      s   
zMLPLayerWithBN.__init__c                 C   s&   |  |}| |dd|}|S )Nr   r   )rS   r)   rL   
reshape_asr*   hidden_stater   r   r   r5      s   
zMLPLayerWithBN.forward)r   r   r   r   r%   r5   r8   r   r   r1   r   rO   ~   s    rO   c                       s$   e Zd Z fddZdd Z  ZS )LevitSubsamplec                       t    || _|| _d S r3   )r$   r%   r.   
resolution)r*   r.   r]   r1   r   r   r%         

zLevitSubsample.__init__c                 C   sL   |j \}}}||| j| j|d d d d | jd d | jf |d|}|S )N)rJ   viewr]   r.   reshape)r*   rY   
batch_size_channelsr   r   r   r5      s   
zLevitSubsample.forwardrZ   r   r   r1   r   r[      s    r[   c                       B   e Zd Z fddZe d
 fdd	Zdd Zdd	 Z  Z	S )LevitAttentionc                    sB  t    || _|d | _|| _|| _|| | || d  | _|| | | _t|| j| _	t
 | _t| j|dd| _ttt|t|}t|}i g }}	|D ],}
|D ]'}t|
d |d  t|
d |d  f}||vrwt|||< |	||  qWqSi | _tj
t|t|| _| jdt|	||dd d S )	N      r=   r   )r0   r   attention_bias_idxsF
persistent)r$   r%   num_attention_headsscalekey_dimattention_ratioout_dim_keys_valuesout_dim_projectionrO   queries_keys_valuesr   rA   
activation
projectionlist	itertoolsproductrangelenabsappendattention_bias_cacher   	Parameterzerosattention_biasesregister_buffer
LongTensorr`   )r*   r?   rm   rk   rn   r]   points
len_pointsattention_offsetsindicesp1p2offsetr1   r   r   r%      s2   



(
zLevitAttention.__init__Tc                    (   t  | |r| jri | _d S d S d S r3   r$   trainr{   r*   moder1   r   r   r         

zLevitAttention.trainc                 C   P   | j r| jd d | jf S t|}|| jvr#| jd d | jf | j|< | j| S r3   trainingr~   rh   strr{   r*   device
device_keyr   r   r   get_attention_biases      

z#LevitAttention.get_attention_biasesc           
      C   s   |j \}}}| |}|||| jdj| j| j| j| j gdd\}}}|dddd}|dddd}|dddd}||dd | j	 | 
|j }	|	jdd}	|	| dd||| j}| | |}|S Nr_   r   dimr   r=   r   )rJ   rq   r`   rk   splitrm   rn   permuterM   rl   r   r   softmaxra   rp   rs   rr   )
r*   rY   rb   
seq_lengthrc   rq   querykeyvalue	attentionr   r   r   r5      s   
"zLevitAttention.forwardT
r   r   r   r%   r   no_gradr   r   r5   r8   r   r   r1   r   rf      s    	rf   c                       re   )LevitAttentionSubsamplec	                    s  t    || _|d | _|| _|| _|| | ||  | _|| | | _|| _t	|| j| _
t||| _t	||| | _t | _t	| j|| _i | _ttt|t|}	ttt|t|}
t|	t|
}}i g }}|
D ]>}|	D ]9}d}t|d | |d  |d d  t|d | |d  |d d  f}||vrt|||< |||  qxqttjt|t|| _| jdt| ||dd d S )Nrg   r   r   r=   rh   Fri   )!r$   r%   rk   rl   rm   rn   ro   rp   resolution_outrO   keys_valuesr[   queries_subsamplequeriesr   rA   rr   rs   r{   rt   ru   rv   rw   rx   ry   rz   r   r|   r}   r~   r   r   r`   )r*   rU   rV   rm   rk   rn   r.   resolution_inr   r   points_r   len_points_r   r   r   r   sizer   r1   r   r   r%      s<   



H
z LevitAttentionSubsample.__init__Tc                    r   r3   r   r   r1   r   r   r     r   zLevitAttentionSubsample.trainc                 C   r   r3   r   r   r   r   r   r   
  r   z,LevitAttentionSubsample.get_attention_biasesc           	      C   s   |j \}}}| |||| jdj| j| j| j gdd\}}|dddd}|dddd}| | 	|}||| j
d | j| jdddd}||dd | j | |j }|jdd}|| dd|d| j}| | |}|S r   )rJ   r   r`   rk   r   rm   rn   r   r   r   r   rM   rl   r   r   r   ra   rp   rs   rr   )	r*   rY   rb   r   rc   r   r   r   r   r   r   r   r5     s"   "zLevitAttentionSubsample.forwardr   r   r   r   r1   r   r      s    -	r   c                       r9   )LevitMLPLayerzE
    MLP Layer with `2X` expansion in contrast to ViT with `4X`.
    c                    s0   t    t||| _t | _t||| _d S r3   )r$   r%   rO   	linear_upr   rA   rr   linear_down)r*   rU   
hidden_dimr1   r   r   r%   .  s   

zLevitMLPLayer.__init__c                 C   s"   |  |}| |}| |}|S r3   )r   rr   r   rX   r   r   r   r5   4  s   


zLevitMLPLayer.forwardr7   r   r   r1   r   r   )  s    r   c                       r9   )LevitResidualLayerz"
    Residual Block for LeViT
    c                    r\   r3   )r$   r%   module	drop_rate)r*   r   r   r1   r   r   r%   @  r^   zLevitResidualLayer.__init__c                 C   sn   | j r.| jdkr.tj|ddd|jd}|| jd| j  }|| 	||  }|S || 	| }|S )Nr   r   )r   )
r   r   r   randr   r   ge_divdetachr   )r*   rY   rndr   r   r   r5   E  s   zLevitResidualLayer.forwardr7   r   r   r1   r   r   ;      r   c                       s0   e Zd ZdZ fddZdd Zdd Z  ZS )
LevitStagezP
    LeViT Stage consisting of `LevitMLPLayer` and `LevitAttention` layers.
    c                    sD  t    g | _|| _|
| _t|D ])}| jtt|||||
| jj	 |dkr;|| }| jtt
||| jj	 q|	d dkr| jd |	d  d | _| jt| jj||d  |	d |	d |	d |	d |
| jd | j| _|	d dkr| jj|d  |	d  }| jtt
| jj|d  || jj	 t| j| _d S )	Nr   	Subsampler      r=   r   )rm   rk   rn   r.   r   r   r<   )r$   r%   layersrI   r   rw   rz   r   rf   drop_path_rater   r   r   r?   r   
ModuleList)r*   rI   idxr?   rm   depthsrk   rn   	mlp_ratiodown_opsr   rc   r   r1   r   r   r%   U  sN   
zLevitStage.__init__c                 C   s   | j S r3   )r   )r*   r   r   r   get_resolution  s   zLevitStage.get_resolutionc                 C   s   | j D ]}||}q|S r3   )r   )r*   rY   layerr   r   r   r5     r6   zLevitStage.forward)r   r   r   r   r%   r   r5   r8   r   r   r1   r   r   P  s
    7r   c                       s*   e Zd ZdZ fddZdddZ  ZS )	LevitEncoderzC
    LeViT Encoder consisting of multiple `LevitStage` stages.
    c                    s   t    || _| jj| jj }g | _| jjdg tt	|j
D ].}t|||j| |j| |j
| |j| |j| |j| |j| |
}| }| j| q"t| j| _d S )N )r$   r%   rI   
image_size
patch_sizestagesr   rz   rw   rx   r   r   r?   rm   rk   rn   r   r   r   r   )r*   rI   r]   	stage_idxstager1   r   r   r%     s*   
zLevitEncoder.__init__FTc                 C   sb   |rdnd }| j D ]}|r||f }||}q	|r||f }|s+tdd ||fD S t||dS )Nr   c                 s   s    | ]	}|d ur|V  qd S r3   r   ).0vr   r   r   	<genexpr>  s    z'LevitEncoder.forward.<locals>.<genexpr>)last_hidden_stater   )r   r   r   )r*   rY   output_hidden_statesreturn_dictall_hidden_statesr   r   r   r   r5     s   



zLevitEncoder.forward)FTr7   r   r   r1   r   r     s    r   c                       r9   )LevitClassificationLayerz$
    LeViT Classification Layer
    c                    s(   t    t|| _t||| _d S r3   )r$   r%   r   rT   r)   rR   rS   )r*   rU   rV   r1   r   r   r%     s   
z!LevitClassificationLayer.__init__c                 C   s   |  |}| |}|S r3   )r)   rS   )r*   rY   r   r   r   r   r5     r6   z LevitClassificationLayer.forwardr7   r   r   r1   r   r     r   r   c                   @   s,   e Zd ZU eed< dZdZdgZdd ZdS )LevitPreTrainedModelrI   levitrN   r   c                 C   sz   t |tjtjfr#|jjjd| jjd |j	dur!|j	j
  dS dS t |tjtjfr;|j	j
  |jjd dS dS )zInitialize the weightsg        )meanstdNg      ?)
isinstancer   rR   r&   weightdatanormal_rI   initializer_ranger#   zero_rT   r(   fill_)r*   r   r   r   r   _init_weights  s   
z"LevitPreTrainedModel._init_weightsN)	r   r   r   r   r   base_model_prefixmain_input_name_no_split_modulesr   r   r   r   r   r     s   
 r   c                       X   e Zd Z fddZe			d
deej dee dee de	e
ef fdd	Z  ZS )
LevitModelc                    s2   t  | || _t|| _t|| _|   d S r3   )r$   r%   rI   r:   patch_embeddingsr   encoder	post_initrH   r1   r   r   r%     s
   

zLevitModel.__init__NrN   r   r   returnc                 C   s   |d ur|n| j j}|d ur|n| j j}|d u rtd| |}| j|||d}|d }|jdd}|s?||f|dd   S t|||jdS )Nz You have to specify pixel_valuesr   r   r   r   r   )r   pooler_outputr   )	rI   r   use_return_dictrK   r   r   r   r   r   )r*   rN   r   r   r4   encoder_outputsr   pooled_outputr   r   r   r5     s(   
zLevitModel.forwardNNN)r   r   r   r%   r   r   r   r   boolr   r   r   r5   r8   r   r   r1   r   r     s    
r   z
    Levit Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
    ImageNet.
    c                       sd   e Zd Z fddZe				ddeej deej dee	 dee	 de
eef f
d	d
Z  ZS )LevitForImageClassificationc                    sX   t  | || _|j| _t|| _|jdkr t|jd |jntj	
 | _|   d S Nr   r_   )r$   r%   rI   
num_labelsr   r   r   r?   r   r   Identity
classifierr   rH   r1   r   r   r%     s   

z$LevitForImageClassification.__init__NrN   labelsr   r   r   c           
      C   s   |dur|n| j j}| j|||d}|d }|d}| |}d}|dur.| ||| j }|sD|f|dd  }	|durB|f|	 S |	S t|||jd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).
        Nr   r   r   r=   )lossr   r   )rI   r   r   r   r   loss_functionr	   r   )
r*   rN   r   r   r   outputssequence_outputr   r   outputr   r   r   r5   -  s    

z#LevitForImageClassification.forward)NNNN)r   r   r   r%   r   r   r   r   r   r   r   r   r	   r5   r8   r   r   r1   r   r     s$    
r   ap  
    LeViT Model transformer with image classification heads on top (a linear layer on top of the final hidden state and
    a linear layer on top of the final hidden state of the distillation token) e.g. for ImageNet. .. warning::
           This model supports inference-only. Fine-tuning with distillation (i.e. with a teacher) is not yet
           supported.
    c                       r   )&LevitForImageClassificationWithTeacherc                    s   t  | || _|j| _t|| _|jdkr t|jd |jntj	
 | _|jdkr4t|jd |jntj	
 | _|   d S r   )r$   r%   rI   r   r   r   r   r?   r   r   r   r   classifier_distillr   rH   r1   r   r   r%   [  s   


z/LevitForImageClassificationWithTeacher.__init__NrN   r   r   r   c           
      C   s   |d ur|n| j j}| j|||d}|d }|d}| || |}}|| d }|s;|||f|dd   }	|	S t||||jdS )Nr   r   r   r=   )r   r   r   r   )rI   r   r   r   r   r   r   r   )
r*   rN   r   r   r   r   r   distill_logitsr   r   r   r   r   r5   p  s   
z.LevitForImageClassificationWithTeacher.forwardr   )r   r   r   r%   r   r   r   r   r   r   r   r   r5   r8   r   r   r1   r   r   R  s    	
r   )r   r   r   r   )*r   ru   dataclassesr   typingr   r   r   r   modeling_outputsr   r   r	   r
   modeling_utilsr   utilsr   r   configuration_levitr   
get_loggerr   loggerr   Moduler    r:   rO   r[   rf   r   r   r   r   r   r   r   r   r   r   __all__r   r   r   r   <module>   sP   
,>SE..62