o
    wig                     @   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Zddlm	Z	 ddl
mZmZ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$e	jZ(G d%d& d&e	jZ)eG d'd( d(eZ*eG d)d* d*e*Z+ed+dG d,d- d-e*Z,ed.dG d/d0 d0e*Z-g d1Z.dS )2zPyTorch LeViT model.    N)	dataclass)OptionalUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss   )BaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$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!   e/home/ubuntu/sommelier/.venv/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$   r%   bn_weight_init	__class__r!   r"   r(   E   s
   
zLevitConvEmbeddings.__init__c                 C   s   |  |}| |}|S N)r*   r,   )r-   
embeddingsr!   r!   r"   forwardN      

zLevitConvEmbeddings.forward)r   r   r   r   r   r   r   r(   r8   __classcell__r!   r!   r4   r"   r#   @   s
    	r#   c                       (   e Zd ZdZ fddZdd Z  ZS )LevitPatchEmbeddingsz
    LeViT patch embeddings, for final embeddings to be passed to transformer blocks. It consists of multiple
    `LevitConvEmbeddings`.
    c                    s   t    t|j|jd d |j|j|j| _t	
 | _t|jd d |jd d |j|j|j| _t	
 | _t|jd d |jd d |j|j|j| _t	
 | _t|jd d |jd |j|j|j| _|j| _d S )Nr            )r'   r(   r#   num_channelshidden_sizesr0   r1   r2   embedding_layer_1r   	Hardswishactivation_layer_1embedding_layer_2activation_layer_2embedding_layer_3activation_layer_3embedding_layer_4r-   configr4   r!   r"   r(   Z   s"   

$
$
 zLevitPatchEmbeddings.__init__c                 C   st   |j d }|| jkrtd| |}| |}| |}| |}| |}| |}| 	|}|
dddS )Nr   zeMake sure that the channel dimension of the pixel values match with the one set in the configuration.r@   )shaperA   
ValueErrorrC   rE   rF   rG   rH   rI   rJ   flatten	transpose)r-   pixel_valuesrA   r7   r!   r!   r"   r8   p   s   








zLevitPatchEmbeddings.forwardr:   r!   r!   r4   r"   r=   T   s    r=   c                       s&   e Zd Zd fdd	Zdd Z  ZS )MLPLayerWithBNr   c                    s,   t    tj||dd| _t|| _d S )NF)in_featuresout_featuresr&   )r'   r(   r   LinearlinearBatchNorm1dr,   )r-   	input_dim
output_dimr3   r4   r!   r"   r(      s   
zMLPLayerWithBN.__init__c                 C   s&   |  |}| |dd|}|S )Nr   r   )rV   r,   rO   
reshape_asr-   hidden_stater!   r!   r"   r8      s   
zMLPLayerWithBN.forward)r   r   r   r   r(   r8   r;   r!   r!   r4   r"   rR      s    rR   c                       s$   e Zd Z fddZdd Z  ZS )LevitSubsamplec                       t    || _|| _d S r6   )r'   r(   r1   
resolution)r-   r1   r`   r4   r!   r"   r(         

zLevitSubsample.__init__c                 C   sL   |j \}}}||| j| j|d d d d | jd d | jf |d|}|S )N)rM   viewr`   r1   reshape)r-   r\   
batch_size_channelsr!   r!   r"   r8      s   
zLevitSubsample.forwardr]   r!   r!   r4   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   )r3   r   attention_bias_idxsF
persistent)r'   r(   num_attention_headsscalekey_dimattention_ratioout_dim_keys_valuesout_dim_projectionrR   queries_keys_valuesr   rD   
activation
projectionlist	itertoolsproductrangelenabsappendattention_bias_cacher   	Parameterzerosattention_biasesregister_buffer
LongTensorrc   )r-   rB   rp   rn   rq   r`   points
len_pointsattention_offsetsindicesp1p2offsetr4   r!   r"   r(      s2   



(
zLevitAttention.__init__Tc                    (   t  | |r| jri | _d S d S d S r6   r'   trainr~   r-   moder4   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 r6   trainingr   rk   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 Nrb   r	   dimr   r@   r   )rM   rt   rc   rn   splitrp   rq   permuterP   ro   r   r   softmaxrd   rs   rv   ru   )
r-   r\   re   
seq_lengthrf   rt   querykeyvalue	attentionr!   r!   r"   r8      s   
"zLevitAttention.forwardT
r   r   r   r(   r   no_gradr   r   r8   r;   r!   r!   r4   r"   ri      s    	ri   c                       rh   )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 )Nrj   r   r   r@   rk   Frl   )!r'   r(   rn   ro   rp   rq   rr   rs   resolution_outrR   keys_valuesr^   queries_subsamplequeriesr   rD   ru   rv   r~   rw   rx   ry   rz   r{   r|   r}   r   r   r   r   r   r   rc   )r-   rX   rY   rp   rn   rq   r1   resolution_inr   r   points_r   len_points_r   r   r   r   sizer   r4   r!   r"   r(      s<   



H
z LevitAttentionSubsample.__init__Tc                    r   r6   r   r   r4   r!   r"   r     r   zLevitAttentionSubsample.trainc                 C   r   r6   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   )rM   r   rc   rn   r   rp   rq   r   r   r   r   rP   ro   r   r   r   rd   rs   rv   ru   )	r-   r\   re   r   rf   r   r   r   r   r!   r!   r"   r8     s"   "zLevitAttentionSubsample.forwardr   r   r!   r!   r4   r"   r      s    -	r   c                       r<   )LevitMLPLayerzE
    MLP Layer with `2X` expansion in contrast to ViT with `4X`.
    c                    s0   t    t||| _t | _t||| _d S r6   )r'   r(   rR   	linear_upr   rD   ru   linear_down)r-   rX   
hidden_dimr4   r!   r"   r(   0  s   

zLevitMLPLayer.__init__c                 C   s"   |  |}| |}| |}|S r6   )r   ru   r   r[   r!   r!   r"   r8   6  s   


zLevitMLPLayer.forwardr:   r!   r!   r4   r"   r   +  s    r   c                       r<   )LevitResidualLayerz"
    Residual Block for LeViT
    c                    r_   r6   )r'   r(   module	drop_rate)r-   r   r   r4   r!   r"   r(   B  ra   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-   r\   rndr!   r!   r"   r8   G  s   zLevitResidualLayer.forwardr:   r!   r!   r4   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	   )rp   rn   rq   r1   r   r   r?   )r'   r(   layersrL   r   rz   r}   r   ri   drop_path_rater   r   r   rB   r   
ModuleList)r-   rL   idxrB   rp   depthsrn   rq   	mlp_ratiodown_opsr   rf   r   r4   r!   r"   r(   W  sN   
zLevitStage.__init__c                 C   s   | j S r6   )r   )r-   r!   r!   r"   get_resolution  s   zLevitStage.get_resolutionc                 C   s   | j D ]}||}q|S r6   )r   )r-   r\   layerr!   r!   r"   r8     r9   zLevitStage.forward)r   r   r   r   r(   r   r8   r;   r!   r!   r4   r"   r   R  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(   rL   
image_size
patch_sizestagesr   r}   rz   r{   r   r   rB   rp   rn   rq   r   r   r   r   )r-   rL   r`   	stage_idxstager4   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 r6   r!   ).0vr!   r!   r"   	<genexpr>  s    z'LevitEncoder.forward.<locals>.<genexpr>)last_hidden_stater   )r   r    r
   )r-   r\   output_hidden_statesreturn_dictall_hidden_statesr   r!   r!   r"   r8     s   



zLevitEncoder.forward)FTr:   r!   r!   r4   r"   r     s    r   c                       r<   )LevitClassificationLayerz$
    LeViT Classification Layer
    c                    s(   t    t|| _t||| _d S r6   )r'   r(   r   rW   r,   rU   rV   )r-   rX   rY   r4   r!   r"   r(     s   
z!LevitClassificationLayer.__init__c                 C   s   |  |}| |}|S r6   )r,   rV   )r-   r\   r   r!   r!   r"   r8     r9   z LevitClassificationLayer.forwardr:   r!   r!   r4   r"   r     r   r   c                   @   s&   e Zd ZeZdZdZdgZdd ZdS )LevitPreTrainedModellevitrQ   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   rU   r)   weightdatanormal_rL   initializer_ranger&   zero_rW   r+   fill_)r-   r   r!   r!   r"   _init_weights  s   
z"LevitPreTrainedModel._init_weightsN)	r   r   r   r   config_class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 r6   )r'   r(   rL   r=   patch_embeddingsr   encoder	post_initrK   r4   r!   r"   r(     s
   

zLevitModel.__init__NrQ   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   )	rL   r   use_return_dictrN   r   r   r   r   r   )r-   rQ   r   r   r7   encoder_outputsr   pooled_outputr!   r!   r"   r8     s(   
zLevitModel.forwardNNN)r   r   r   r(   r   r   r   r   boolr   r    r   r8   r;   r!   r!   r4   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   rb   )r'   r(   rL   
num_labelsr   r   r   rB   r   r   Identity
classifierr   rK   r4   r!   r"   r(     s   

z$LevitForImageClassification.__init__NrQ   labelsr   r   r   c                 C   sb  |dur|n| j j}| j|||d}|d }|d}| |}d}|dur| j jdu rP| jdkr6d| j _n| jdkrL|jtj	ksG|jtj
krLd| j _nd| j _| j jdkrnt }	| jdkrh|	| | }n+|	||}n%| j jdkrt }	|	|d| j|d}n| j jdkrt }	|	||}|s|f|d	d  }
|dur|f|
 S |
S t|||jd
S )a  
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        Nr   r   r   
regressionsingle_label_classificationmulti_label_classificationrb   r@   )lossr   r   )rL   r   r   r   r   problem_typer   dtyper   longintr   squeezer   rc   r   r   r   )r-   rQ   r   r   r   outputssequence_outputr   r   loss_fctoutputr!   r!   r"   r8   /  s@   



"


z#LevitForImageClassification.forward)NNNN)r   r   r   r(   r   r   r   r   r   r   r   r    r   r8   r;   r!   r!   r4   r"   r     s$    
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(   rL   r   r   r   r   rB   r   r   r   r   classifier_distillr   rK   r4   r!   r"   r(   o  s   


z/LevitForImageClassificationWithTeacher.__init__NrQ   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   )rL   r   r   r   r   r  r   r   )
r-   rQ   r   r   r  r  r   distill_logitsr   r	  r!   r!   r"   r8     s   
z.LevitForImageClassificationWithTeacher.forwardr   )r   r   r   r(   r   r   r   r   r   r   r    r   r8   r;   r!   r!   r4   r"   r
  f  s    	
r
  )r   r
  r   r   )/r   rx   dataclassesr   typingr   r   r   torch.utils.checkpointr   torch.nnr   r   r   modeling_outputsr
   r   r   r   modeling_utilsr   utilsr   r   configuration_levitr   
get_loggerr   loggerr   Moduler#   r=   rR   r^   ri   r   r   r   r   r   r   r   r   r   r
  __all__r!   r!   r!   r"   <module>   sT   
,>SE..H2