o
    wiŇ                  	   @   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 dd	lmZ dd
lmZ ddlmZmZ ddlmZmZmZmZmZmZ ddlmZ ddl m!Z! e rnddl"m#Z#m$Z$ ndd Z$dd Z#e%e&Z'eeddG dd deZ(eeddG dd deZ)eeddG dd deZ*G dd  d e	j+Z,G d!d" d"e	j+Z-G d#d$ d$e	j+Z.dJd'ej/d(e0d)e1d*ej/fd+d,Z2G d-d. d.e	j+Z3G d/d0 d0e	j+Z4G d1d2 d2e	j+Z5G d3d4 d4e	j+Z6G d5d6 d6e	j+Z7G d7d8 d8e	j+Z8G d9d: d:e	j+Z9G d;d< d<e	j+Z:G d=d> d>e	j+Z;eG d?d@ d@eZ<eG dAdB dBe<Z=edCdG dDdE dEe<Z>edFdG dGdH dHe<eZ?g dIZ@dS )Kz9PyTorch Dilated Neighborhood Attention Transformer model.    N)	dataclass)OptionalUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss   )ACT2FN)BackboneOutput)PreTrainedModel) find_pruneable_heads_and_indicesprune_linear_layer)ModelOutputOptionalDependencyNotAvailableauto_docstringis_natten_availableloggingrequires_backends)BackboneMixin   )DinatConfig)
natten2davnatten2dqkrpbc                  O      t  Nr   argskwargs r    e/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/dinat/modeling_dinat.pyr   .      r   c                  O   r   r   r   r   r    r    r!   r   1   r"   r   zO
    Dinat encoder's outputs, with potential hidden states and attentions.
    )custom_introc                   @   sr   e Zd ZU dZ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< dZee
ejdf  ed< dS )DinatEncoderOutputa  
    reshaped_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
        Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
        shape `(batch_size, hidden_size, height, width)`.

        Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
        include the spatial dimensions.
    Nlast_hidden_state.hidden_states
attentionsreshaped_hidden_states)__name__
__module____qualname____doc__r%   r   torchFloatTensor__annotations__r&   tupler'   r(   r    r    r    r!   r$   ;   s   
 	r$   zW
    Dinat 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< dZeeejdf  ed< dS )	DinatModelOutputa  
    pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*, returned when `add_pooling_layer=True` is passed):
        Average pooling of the last layer hidden-state.
    reshaped_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
        Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
        shape `(batch_size, hidden_size, height, width)`.

        Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
        include the spatial dimensions.
    Nr%   pooler_output.r&   r'   r(   )r)   r*   r+   r,   r%   r   r-   r.   r/   r3   r&   r0   r'   r(   r    r    r    r!   r2   Q   s   
 r2   z1
    Dinat outputs for image classification.
    c                   @   r1   )	DinatImageClassifierOutputa7  
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Classification (or regression if config.num_labels==1) loss.
    logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
        Classification (or regression if config.num_labels==1) scores (before SoftMax).
    reshaped_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
        Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
        shape `(batch_size, hidden_size, height, width)`.

        Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
        include the spatial dimensions.
    Nlosslogits.r&   r'   r(   )r)   r*   r+   r,   r5   r   r-   r.   r/   r6   r&   r0   r'   r(   r    r    r    r!   r4   j   s   
 r4   c                       s>   e Zd ZdZ fddZdeej deej	 fddZ
  ZS )DinatEmbeddingsz6
    Construct the patch and position embeddings.
    c                    s4   t    t|| _t|j| _t|j	| _
d S r   )super__init__DinatPatchEmbeddingspatch_embeddingsr   	LayerNorm	embed_dimnormDropouthidden_dropout_probdropoutselfconfig	__class__r    r!   r9      s   

zDinatEmbeddings.__init__pixel_valuesreturnc                 C   s"   |  |}| |}| |}|S r   )r;   r>   rA   )rC   rG   
embeddingsr    r    r!   forward   s   


zDinatEmbeddings.forward)r)   r*   r+   r,   r9   r   r-   r.   r0   TensorrJ   __classcell__r    r    rE   r!   r7      s    &r7   c                       s:   e Zd ZdZ fddZdeej dejfddZ	  Z
S )r:   z
    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
    `hidden_states` (patch embeddings) of shape `(batch_size, height, width, hidden_size)` to be consumed by a
    Transformer.
    c              
      sr   t    |j}|j|j}}|| _|dkrntdttj| j|d ddddtj|d |dddd| _	d S )N   z2Dinat only supports patch size of 4 at the moment.   r	   r	   rN   rN   r   r   )kernel_sizestridepadding)
r8   r9   
patch_sizenum_channelsr=   
ValueErrorr   
SequentialConv2d
projection)rC   rD   rU   rV   hidden_sizerE   r    r!   r9      s   

zDinatPatchEmbeddings.__init__rG   rH   c                 C   s>   |j \}}}}|| jkrtd| |}|dddd}|S )NzeMake sure that the channel dimension of the pixel values match with the one set in the configuration.r   rN   r	   r   )shaperV   rW   rZ   permute)rC   rG   _rV   heightwidthrI   r    r    r!   rJ      s   

zDinatPatchEmbeddings.forward)r)   r*   r+   r,   r9   r   r-   r.   rK   rJ   rL   r    r    rE   r!   r:      s    "r:   c                       sL   e Zd ZdZejfdedejddf fddZde	j
de	j
fd	d
Z  ZS )DinatDownsamplerz
    Convolutional Downsampling Layer.

    Args:
        dim (`int`):
            Number of input channels.
        norm_layer (`nn.Module`, *optional*, defaults to `nn.LayerNorm`):
            Normalization layer class.
    dim
norm_layerrH   Nc                    s>   t    || _tj|d| ddddd| _|d| | _d S )NrN   rO   rP   rQ   F)rR   rS   rT   bias)r8   r9   rb   r   rY   	reductionr>   )rC   rb   rc   rE   r    r!   r9      s   
zDinatDownsampler.__init__input_featurec                 C   s0   |  |dddddddd}| |}|S )Nr   r	   r   rN   )re   r]   r>   )rC   rf   r    r    r!   rJ      s   "
zDinatDownsampler.forward)r)   r*   r+   r,   r   r<   intModuler9   r-   rK   rJ   rL   r    r    rE   r!   ra      s    "
ra           Finput	drop_probtrainingrH   c                 C   sd   |dks|s| S d| }| j d fd| jd   }|tj|| j| jd }|  | || }|S )aF  
    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
    argument.
    ri   r   r   )r   )dtypedevice)r\   ndimr-   randrm   rn   floor_div)rj   rk   rl   	keep_probr\   random_tensoroutputr    r    r!   	drop_path   s   
rv   c                       sT   e Zd ZdZddee ddf fddZdejdejfdd	Z	de
fd
dZ  ZS )DinatDropPathzXDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).Nrk   rH   c                    s   t    || _d S r   )r8   r9   rk   )rC   rk   rE   r    r!   r9      s   

zDinatDropPath.__init__r&   c                 C   s   t || j| jS r   )rv   rk   rl   rC   r&   r    r    r!   rJ      s   zDinatDropPath.forwardc                 C   s   d| j  S )Nzp=)rk   rC   r    r    r!   
extra_repr   s   zDinatDropPath.extra_reprr   )r)   r*   r+   r,   r   floatr9   r-   rK   rJ   strrz   rL   r    r    rE   r!   rw      s
    rw   c                       J   e Zd Z fddZdd Z	ddejdee de	ej fd	d
Z
  ZS )NeighborhoodAttentionc                    s   t    || dkrtd| d| d|| _t|| | _| j| j | _|| _|| _t	
t|d| j d d| j d | _t	j| j| j|jd| _t	j| j| j|jd| _t	j| j| j|jd| _t	|j| _d S )Nr   zThe hidden size (z6) is not a multiple of the number of attention heads ()rN   r   )rd   )r8   r9   rW   num_attention_headsrg   attention_head_sizeall_head_sizerR   dilationr   	Parameterr-   zerosrpbLinearqkv_biasquerykeyvaluer?   attention_probs_dropout_probrA   rC   rD   rb   	num_headsrR   r   rE   r    r!   r9      s   
*zNeighborhoodAttention.__init__c                 C   s8   |  d d | j| jf }||}|dddddS )Nr   r	   r   rN   rM   )sizer   r   viewr]   )rC   xnew_x_shaper    r    r!   transpose_for_scores  s   
z*NeighborhoodAttention.transpose_for_scoresFr&   output_attentionsrH   c                 C   s   |  | |}|  | |}|  | |}|t| j }t||| j| j	| j
}tjj|dd}| |}t||| j	| j
}|ddddd }| d d | jf }	||	}|re||f}
|
S |f}
|
S )	Nr   rb   r   rN   r	   r   rM   )r   r   r   r   mathsqrtr   r   r   rR   r   r   
functionalsoftmaxrA   r   r]   
contiguousr   r   r   )rC   r&   r   query_layer	key_layervalue_layerattention_scoresattention_probscontext_layernew_context_layer_shapeoutputsr    r    r!   rJ     s   

zNeighborhoodAttention.forwardF)r)   r*   r+   r9   r   r-   rK   r   boolr0   rJ   rL   r    r    rE   r!   r~      s    r~   c                       s8   e Zd Z fddZdejdejdejfddZ  ZS )NeighborhoodAttentionOutputc                    s*   t    t||| _t|j| _d S r   )r8   r9   r   r   denser?   r   rA   rC   rD   rb   rE   r    r!   r9   9  s   
z$NeighborhoodAttentionOutput.__init__r&   input_tensorrH   c                 C      |  |}| |}|S r   r   rA   )rC   r&   r   r    r    r!   rJ   >  s   

z#NeighborhoodAttentionOutput.forwardr)   r*   r+   r9   r-   rK   rJ   rL   r    r    rE   r!   r   8  s    $r   c                       r}   )NeighborhoodAttentionModulec                    s4   t    t|||||| _t||| _t | _d S r   )r8   r9   r~   rC   r   ru   setpruned_headsr   rE   r    r!   r9   F  s   
z$NeighborhoodAttentionModule.__init__c                 C   s   t |dkrd S t|| jj| jj| j\}}t| jj|| j_t| jj|| j_t| jj	|| j_	t| j
j|dd| j
_| jjt | | j_| jj| jj | j_| j|| _d S )Nr   r   r   )lenr   rC   r   r   r   r   r   r   r   ru   r   r   union)rC   headsindexr    r    r!   prune_headsL  s   z'NeighborhoodAttentionModule.prune_headsFr&   r   rH   c                 C   s2   |  ||}| |d |}|f|dd   }|S Nr   r   )rC   ru   )rC   r&   r   self_outputsattention_outputr   r    r    r!   rJ   ^  s   z#NeighborhoodAttentionModule.forwardr   )r)   r*   r+   r9   r   r-   rK   r   r   r0   rJ   rL   r    r    rE   r!   r   E  s    r   c                       2   e Zd Z fddZdejdejfddZ  ZS )DinatIntermediatec                    sJ   t    t|t|j| | _t|jt	rt
|j | _d S |j| _d S r   )r8   r9   r   r   rg   	mlp_ratior   
isinstance
hidden_actr|   r
   intermediate_act_fnr   rE   r    r!   r9   j  s
   
zDinatIntermediate.__init__r&   rH   c                 C   r   r   )r   r   rx   r    r    r!   rJ   r     

zDinatIntermediate.forwardr   r    r    rE   r!   r   i  s    r   c                       r   )DinatOutputc                    s4   t    tt|j| || _t|j| _	d S r   )
r8   r9   r   r   rg   r   r   r?   r@   rA   r   rE   r    r!   r9   y  s   
zDinatOutput.__init__r&   rH   c                 C   r   r   r   rx   r    r    r!   rJ   ~  r   zDinatOutput.forwardr   r    r    rE   r!   r   x  s    r   c                	       sR   e Zd Zd fdd	Zdd Z	ddejdee d	e	ejejf fd
dZ
  ZS )
DinatLayerri   c                    s   t    |j| _|j| _|| _| j| j | _tj||jd| _	t
|||| j| jd| _|dkr4t|nt | _tj||jd| _t||| _t||| _|jdkretj|jtd|f dd| _d S d | _d S )Neps)rR   r   ri   r   rN   T)requires_grad)r8   r9   chunk_size_feed_forwardrR   r   window_sizer   r<   layer_norm_epslayernorm_beforer   	attentionrw   Identityrv   layernorm_afterr   intermediater   ru   layer_scale_init_valuer   r-   oneslayer_scale_parameters)rC   rD   rb   r   r   drop_path_raterE   r    r!   r9     s$   

zDinatLayer.__init__c           
      C   sd   | j }d}||k s||k r.d }}td|| }td|| }	dd||||	f}tj||}||fS )N)r   r   r   r   r   r   r   )r   maxr   r   pad)
rC   r&   r_   r`   r   
pad_valuespad_lpad_tpad_rpad_br    r    r!   	maybe_pad  s   zDinatLayer.maybe_padFr&   r   rH   c                 C   s  |  \}}}}|}| |}| |||\}}|j\}	}
}}	| j||d}|d }|d dkp5|d dk}|rJ|d d d |d |d d f  }| jd urV| jd | }|| | }| |}| 	| 
|}| jd urv| jd | }|| | }|r||d f}|S |f}|S )N)r   r   r	      r   )r   r   r   r\   r   r   r   rv   r   ru   r   )rC   r&   r   
batch_sizer_   r`   channelsshortcutr   r^   
height_pad	width_padattention_outputsr   
was_paddedlayer_outputlayer_outputsr    r    r!   rJ     s,   
$


zDinatLayer.forward)ri   r   )r)   r*   r+   r9   r   r-   rK   r   r   r0   rJ   rL   r    r    rE   r!   r     s    r   c                       sB   e Zd Z fddZ	d	dejdee deej fddZ	  Z
S )

DinatStagec                    sf   t     | _| _t fddt|D | _|d ur+|tjd| _	nd | _	d| _
d S )Nc              	      s&   g | ]}t  | | d qS ))rD   rb   r   r   r   )r   .0irD   	dilationsrb   r   r   r    r!   
<listcomp>  s    z'DinatStage.__init__.<locals>.<listcomp>)rb   rc   F)r8   r9   rD   rb   r   
ModuleListrangelayersr<   
downsamplepointing)rC   rD   rb   depthr   r   r   r   rE   r   r!   r9     s   

zDinatStage.__init__Fr&   r   rH   c                 C   sn   |  \}}}}t| jD ]\}}|||}|d }q|}	| jd ur'| |	}||	f}
|r5|
|dd  7 }
|
S r   )r   	enumerater   r   )rC   r&   r   r^   r_   r`   r   layer_moduler   !hidden_states_before_downsamplingstage_outputsr    r    r!   rJ     s   



zDinatStage.forwardr   )r)   r*   r+   r9   r-   rK   r   r   r0   rJ   rL   r    r    rE   r!   r     s    r   c                       sb   e Zd Z fddZ				ddejdee dee dee d	ee d
ee	e
f fddZ  ZS )DinatEncoderc                    sh   t    t j_ _dd tjd jt	 jddD t
 fddtjD _d S )Nc                 S   s   g | ]}|  qS r    )item)r   r   r    r    r!   r     s    z)DinatEncoder.__init__.<locals>.<listcomp>r   cpu)rn   c                    s|   g | ]:}t  t jd |   j|  j|  j| t jd| t jd|d   |jd k r8tnddqS )rN   Nr   )rD   rb   r   r   r   r   r   )	r   rg   r=   depthsr   r   sum
num_levelsra   )r   i_layerrD   dprrC   r    r!   r     s    
*)r8   r9   r   r   r   rD   r-   linspacer   r   r   r   r   levelsrB   rE   r   r!   r9     s   
$

zDinatEncoder.__init__FTr&   r   output_hidden_states(output_hidden_states_before_downsamplingreturn_dictrH   c                 C   s  |rdnd }|r
dnd }|rdnd }|r&| dddd}	||f7 }||	f7 }t| jD ]H\}
}|||}|d }|d }|rS|rS| dddd}	||f7 }||	f7 }n|ri|si| dddd}	||f7 }||	f7 }|rs||dd  7 }q+|stdd |||fD S t||||dS )	Nr    r   r	   r   rN   c                 s   s    | ]	}|d ur|V  qd S r   r    )r   vr    r    r!   	<genexpr>6  s    z'DinatEncoder.forward.<locals>.<genexpr>)r%   r&   r'   r(   )r]   r   r   r0   r$   )rC   r&   r   r   r   r  all_hidden_statesall_reshaped_hidden_statesall_self_attentionsreshaped_hidden_stater   r   r   r   r    r    r!   rJ     s<   





zDinatEncoder.forward)FFFT)r)   r*   r+   r9   r-   rK   r   r   r   r0   r$   rJ   rL   r    r    rE   r!   r     s&    
r   c                   @   s    e Zd ZeZdZdZdd ZdS )DinatPreTrainedModeldinatrG   c                 C   st   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r8|j	j
  |jjd dS dS )zInitialize the weightsri   )meanstdNg      ?)r   r   r   rY   weightdatanormal_rD   initializer_rangerd   zero_r<   fill_)rC   moduler    r    r!   _init_weightsF  s   
z"DinatPreTrainedModel._init_weightsN)r)   r*   r+   r   config_classbase_model_prefixmain_input_namer  r    r    r    r!   r  @  s
    r  c                       st   e Zd Zd fdd	Zdd Zdd Ze				dd	eej	 d
ee
 dee
 dee
 deeef f
ddZ  ZS )
DinatModelTc                    s   t  | t| dg || _t|j| _t|jd| jd   | _	t
|| _t|| _tj| j	|jd| _|r=tdnd| _|   dS )zv
        add_pooling_layer (bool, *optional*, defaults to `True`):
            Whether to add a pooling layer
        nattenrN   r   r   N)r8   r9   r   rD   r   r   r   rg   r=   num_featuresr7   rI   r   encoderr   r<   r   	layernormAdaptiveAvgPool1dpooler	post_init)rC   rD   add_pooling_layerrE   r    r!   r9   U  s   

zDinatModel.__init__c                 C      | j jS r   rI   r;   ry   r    r    r!   get_input_embeddingsk     zDinatModel.get_input_embeddingsc                 C   s*   |  D ]\}}| jj| j| qdS )z
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        N)itemsr  layerr   r   )rC   heads_to_pruner%  r   r    r    r!   _prune_headsn  s   zDinatModel._prune_headsNrG   r   r   r  rH   c           
      C   s   |d ur|n| j j}|d ur|n| j j}|d ur|n| j j}|d u r&td| |}| j||||d}|d }| |}d }| jd urW| |	dd
dd}t	|d}|se||f|dd   }	|	S t|||j|j|jdS )Nz You have to specify pixel_valuesr   r   r  r   r   rN   )r%   r3   r&   r'   r(   )rD   r   r   use_return_dictrW   rI   r  r  r  flatten	transposer-   r2   r&   r'   r(   )
rC   rG   r   r   r  embedding_outputencoder_outputssequence_outputpooled_outputru   r    r    r!   rJ   v  s:   


zDinatModel.forward)T)NNNN)r)   r*   r+   r9   r"  r'  r   r   r-   r.   r   r   r0   r2   rJ   rL   r    r    rE   r!   r  S  s(    
r  z
    Dinat Model transformer with an image classification head on top (a linear layer on top of the final hidden state
    of the [CLS] token) e.g. for ImageNet.
    c                       sn   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	 d	e
eef fd
dZ  ZS )DinatForImageClassificationc                    s\   t  | t| dg |j| _t|| _|jdkr#t| jj|jnt	 | _
|   d S )Nr  r   )r8   r9   r   
num_labelsr  r	  r   r   r  r   
classifierr  rB   rE   r    r!   r9     s   
"z$DinatForImageClassification.__init__NrG   labelsr   r   r  rH   c                 C   sb  |dur|n| j j}| j||||d}|d }| |}d}	|dur| j jdu rL| jdkr2d| j _n| jdkrH|jtjksC|jtj	krHd| j _nd| j _| j jdkrjt
 }
| jdkrd|
| | }	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|j|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   
regressionsingle_label_classificationmulti_label_classificationr   rN   )r5   r6   r&   r'   r(   )rD   r)  r	  r2  problem_typer1  rm   r-   longrg   r   squeezer   r   r   r4   r&   r'   r(   )rC   rG   r3  r   r   r  r   r/  r6   r5   loss_fctru   r    r    r!   rJ     sL   


"


z#DinatForImageClassification.forward)NNNNN)r)   r*   r+   r9   r   r   r-   r.   
LongTensorr   r   r0   r4   rJ   rL   r    r    rE   r!   r0    s*    
r0  zL
    NAT backbone, to be used with frameworks like DETR and MaskFormer.
    c                       s\   e Zd Z fddZdd Ze			ddejdee	 dee	 d	ee	 d
e
f
ddZ  ZS )DinatBackbonec                    s   t    t    t| dg t | _t | _ jg fddt	t
 jD  | _i }t| j| jD ]\}}t|||< q8t|| _|   d S )Nr  c                    s   g | ]}t  jd |  qS )rN   )rg   r=   r   rD   r    r!   r     s    z*DinatBackbone.__init__.<locals>.<listcomp>)r8   r9   _init_backboner   r7   rI   r   r  r=   r   r   r   r  zip_out_featuresr   r   r<   
ModuleDicthidden_states_normsr  )rC   rD   rB  stagerV   rE   r=  r!   r9     s   

&zDinatBackbone.__init__c                 C   r   r   r!  ry   r    r    r!   r"    r#  z"DinatBackbone.get_input_embeddingsNrG   r   r   r  rH   c                 C   s,  |dur|n| j j}|dur|n| j j}|dur|n| j j}| |}| j||dddd}|j}d}t| j|D ]A\}	}
|	| j	v ry|
j
\}}}}|
dddd }
|
||| |}
| j|	 |
}
|
||||}
|
dddd }
||
f7 }q8|s|f}|r||jf7 }|S t||r|jnd|jd	S )
a/  
        Examples:

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

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

        >>> processor = AutoImageProcessor.from_pretrained("shi-labs/nat-mini-in1k-224")
        >>> model = AutoBackbone.from_pretrained(
        ...     "shi-labs/nat-mini-in1k-224", out_features=["stage1", "stage2", "stage3", "stage4"]
        ... )

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

        >>> outputs = model(**inputs)

        >>> feature_maps = outputs.feature_maps
        >>> list(feature_maps[-1].shape)
        [1, 512, 7, 7]
        ```NT)r   r   r   r  r    r   rN   r	   r   )feature_mapsr&   r'   )rD   r)  r   r   rI   r  r(   r?  stage_namesout_featuresr\   r]   r   r   rB  r&   r   r'   )rC   rG   r   r   r  r,  r   r&   rD  rC  hidden_stater   rV   r_   r`   ru   r    r    r!   rJ     sD   !


zDinatBackbone.forward)NNN)r)   r*   r+   r9   r"  r   r-   rK   r   r   r   rJ   rL   r    r    rE   r!   r<    s$    r<  )r0  r  r  r<  )ri   F)Ar,   r   dataclassesr   typingr   r   r-   torch.utils.checkpointr   torch.nnr   r   r   activationsr
   modeling_outputsr   modeling_utilsr   pytorch_utilsr   r   utilsr   r   r   r   r   r   utils.backbone_utilsr   configuration_dinatr   natten.functionalr   r   
get_loggerr)   loggerr$   r2   r4   rh   r7   r:   ra   rK   r{   r   rv   rw   r~   r   r   r   r   r   r   r   r  r  r0  r<  __all__r    r    r    r!   <module>   sz    
$ >$G/FRQb