o
    i                   	   @   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
 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 r`d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)dId&ej*d'e+d(e,d)ej*fd*d+Z-G d,d- d-ej&Z.G d.d/ d/ej&Z/G d0d1 d1ej&Z0G d2d3 d3ej&Z1G d4d5 d5ej&Z2G d6d7 d7ej&Z3G d8d9 d9ej&Z4G d:d; d;ej&Z5G d<d= d=ej&Z6eG d>d? d?eZ7eG d@dA dAe7Z8edBdG dCdD dDe7Z9edEdG dFdG dGe7eZ:g dHZ;dS )Jz9PyTorch Dilated Neighborhood Attention Transformer model.    N)	dataclass)OptionalUnion)nn   )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   \/home/ubuntu/.local/lib/python3.10/site-packages/transformers/models/dinat/modeling_dinat.pyr   ,      r   c                  O   r   r   r   r   r   r   r   r   /   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!   9   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$   r%   )r&   r'   r(   r)   r"   r   r*   r+   r,   r0   r#   r-   r$   r%   r   r   r   r   r/   O   s   
 r/   z1
    Dinat outputs for image classification.
    c                   @   r.   )	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$   r%   )r&   r'   r(   r)   r2   r   r*   r+   r,   r3   r#   r-   r$   r%   r   r   r   r   r1   h   s   
 r1   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   r6      s   

zDinatEmbeddings.__init__pixel_valuesreturnc                 C   s"   |  |}| |}| |}|S r   )r8   r;   r>   )r@   rD   
embeddingsr   r   r   forward   s   


zDinatEmbeddings.forward)r&   r'   r(   r)   r6   r   r*   r+   r-   TensorrG   __classcell__r   r   rB   r   r4      s    &r4   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 )r7   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   rK   rK   r   r   )kernel_sizestridepadding)
r5   r6   
patch_sizenum_channelsr:   
ValueErrorr   
SequentialConv2d
projection)r@   rA   rR   rS   hidden_sizerB   r   r   r6      s   

zDinatPatchEmbeddings.__init__rD   rE   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   rK   r   r   )shaperS   rT   rW   permute)r@   rD   _rS   heightwidthrF   r   r   r   rG      s   

zDinatPatchEmbeddings.forward)r&   r'   r(   r)   r6   r   r*   r+   rH   rG   rI   r   r   rB   r   r7      s    "r7   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_layerrE   Nc                    s>   t    || _tj|d| ddddd| _|d| | _d S )NrK   rL   rM   rN   F)rO   rP   rQ   bias)r5   r6   r_   r   rV   	reductionr;   )r@   r_   r`   rB   r   r   r6      s   
zDinatDownsampler.__init__input_featurec                 C   s0   |  |dddddddd}| |}|S )Nr   r   r   rK   )rb   rZ   r;   )r@   rc   r   r   r   rG      s   "
zDinatDownsampler.forward)r&   r'   r(   r)   r   r9   intModuler6   r*   rH   rG   rI   r   r   rB   r   r^      s    "
r^           Finput	drop_probtrainingrE   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.
    rf   r   r   )r   )dtypedevice)rY   ndimr*   randrj   rk   floor_div)rg   rh   ri   	keep_probrY   random_tensoroutputr   r   r   	drop_path   s   
rs   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).Nrh   rE   c                    s   t    || _d S r   )r5   r6   rh   )r@   rh   rB   r   r   r6      s   

zDinatDropPath.__init__r#   c                 C   s   t || j| jS r   )rs   rh   ri   r@   r#   r   r   r   rG      s   zDinatDropPath.forwardc                 C   s   d| j  S )Nzp=)rh   r@   r   r   r   
extra_repr   s   zDinatDropPath.extra_reprr   )r&   r'   r(   r)   r   floatr6   r*   rH   rG   strrw   rI   r   r   rB   r   rt      s
    rt   c                       B   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 )
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 ()rK   r   )ra   )r5   r6   rT   num_attention_headsrd   attention_head_sizeall_head_sizerO   dilationr   	Parameterr*   zerosrpbLinearqkv_biasquerykeyvaluer<   attention_probs_dropout_probr>   r@   rA   r_   	num_headsrO   r   rB   r   r   r6      s   
*zNeighborhoodAttention.__init__Fr#   output_attentionsrE   c                 C   s  |j \}}}| ||d| j| jdd}| ||d| j| jdd}| ||d| j| jdd}|t	| j }t
||| j| j| j}	tjj|	dd}
| |
}
t|
|| j| j}|ddddd }| d d | jf }||}|r||
f}|S |f}|S )	Nr   rK   r_   r   r   rJ   )rY   r   viewr}   r~   	transposer   r   mathsqrtr   r   rO   r   r   
functionalsoftmaxr>   r   rZ   
contiguoussizer   )r@   r#   r   
batch_size
seq_lengthr[   query_layer	key_layervalue_layerattention_scoresattention_probscontext_layernew_context_layer_shapeoutputsr   r   r   rG     s2   	

zNeighborhoodAttention.forwardFr&   r'   r(   r6   r*   rH   r   boolr-   rG   rI   r   r   rB   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   )r5   r6   r   r   denser<   r   r>   r@   rA   r_   rB   r   r   r6   ?  s   
z$NeighborhoodAttentionOutput.__init__r#   input_tensorrE   c                 C      |  |}| |}|S r   r   r>   )r@   r#   r   r   r   r   rG   D  s   

z#NeighborhoodAttentionOutput.forwardr&   r'   r(   r6   r*   rH   rG   rI   r   r   rB   r   r   >  s    $r   c                       sJ   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 )NeighborhoodAttentionModulec                    s4   t    t|||||| _t||| _t | _d S r   )r5   r6   r{   r@   r   rr   setpruned_headsr   rB   r   r   r6   L  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
   r@   r}   r~   r   r   r   r   r   rr   r   r   union)r@   headsindexr   r   r   prune_headsR  s   z'NeighborhoodAttentionModule.prune_headsFr#   r   rE   c                 C   s2   |  ||}| |d |}|f|dd   }|S Nr   r   )r@   rr   )r@   r#   r   self_outputsattention_outputr   r   r   r   rG   d  s   z#NeighborhoodAttentionModule.forwardr   )r&   r'   r(   r6   r   r*   rH   r   r   r-   rG   rI   r   r   rB   r   r   K  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   )r5   r6   r   r   rd   	mlp_ratior   
isinstance
hidden_actry   r   intermediate_act_fnr   rB   r   r   r6   p  s
   
zDinatIntermediate.__init__r#   rE   c                 C   r   r   )r   r   ru   r   r   r   rG   x     

zDinatIntermediate.forwardr   r   r   rB   r   r   o  s    r   c                       r   )DinatOutputc                    s4   t    tt|j| || _t|j| _	d S r   )
r5   r6   r   r   rd   r   r   r<   r=   r>   r   rB   r   r   r6     s   
zDinatOutput.__init__r#   rE   c                 C   r   r   r   ru   r   r   r   rG     r   zDinatOutput.forwardr   r   r   rB   r   r   ~  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 )
DinatLayerrf   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)rO   r   rf   r   rK   T)requires_grad)r5   r6   chunk_size_feed_forwardrO   r   window_sizer   r9   layer_norm_epslayernorm_beforer   	attentionrt   Identityrs   layernorm_afterr   intermediater   rr   layer_scale_init_valuer   r*   oneslayer_scale_parameters)r@   rA   r_   r   r   drop_path_raterB   r   r   r6     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)
r@   r#   r\   r]   r   
pad_valuespad_lpad_tpad_rpad_br   r   r   	maybe_pad  s   zDinatLayer.maybe_padFr#   r   rE   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   rY   r   r   r   rs   r   rr   r   )r@   r#   r   r   r\   r]   channelsshortcutr   r[   
height_pad	width_padattention_outputsr   
was_paddedlayer_outputlayer_outputsr   r   r   rG     s,   
$


zDinatLayer.forward)rf   r   )r&   r'   r(   r6   r   r*   rH   r   r   r-   rG   rI   r   r   rB   r   r     s    r   c                       rz   )

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 ))rA   r_   r   r   r   )r   .0irA   	dilationsr_   r   r   r   r   
<listcomp>  s    z'DinatStage.__init__.<locals>.<listcomp>)r_   r`   F)r5   r6   rA   r_   r   
ModuleListrangelayersr9   
downsamplepointing)r@   rA   r_   depthr   r   r   r   rB   r   r   r6     s   

zDinatStage.__init__Fr#   r   rE   c                 C   sn   |  \}}}}t| jD ]\}}|||}|d }q|}	| jd ur'| |	}||	f}
|r5|
|dd  7 }
|
S r   )r   	enumerater   r   )r@   r#   r   r[   r\   r]   r   layer_moduler   !hidden_states_before_downsamplingstage_outputsr   r   r   rG     s   



zDinatStage.forwardr   r   r   r   rB   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   xr   r   r   r     s    z)DinatEncoder.__init__.<locals>.<listcomp>r   cpu)rk   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 )rK   Nr   )rA   r_   r   r   r   r   r   )	r   rd   r:   depthsr   r   sum
num_levelsr^   )r   i_layerrA   dprr@   r   r   r     s    
*)r5   r6   r   r   r   rA   r*   linspacer   r   r   r   r   levelsr?   rB   r   r   r6     s   
$

zDinatEncoder.__init__FTr#   r   output_hidden_states(output_hidden_states_before_downsamplingreturn_dictrE   c                 C   s  |rdnd }|r
dnd }|rdnd }|r&| dddd}	||f7 }||	f7 }t| jD ]H\}
}|||}|d }|d }|rS|rS| dddd}	||f7 }||	f7 }n|ri|si| dddd}	||f7 }||	f7 }|rs||dd  7 }q+|stdd |||fD S t||||dS )	Nr   r   r   r   rK   c                 s   s    | ]	}|d ur|V  qd S r   r   )r   vr   r   r   	<genexpr><  s    z'DinatEncoder.forward.<locals>.<genexpr>)r"   r#   r$   r%   )rZ   r   r   r-   r!   )r@   r#   r   r   r   r   all_hidden_statesall_reshaped_hidden_statesall_self_attentionsreshaped_hidden_stater   r   r   r   r   r   r   rG     s<   





zDinatEncoder.forward)FFFT)r&   r'   r(   r6   r*   rH   r   r   r   r-   r!   rG   rI   r   r   rB   r   r      s&    
r   c                   @   s&   e Zd ZU eed< dZdZdd ZdS )DinatPreTrainedModelrA   dinatrD   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 weightsrf   )meanstdNg      ?)r   r   r   rV   weightdatanormal_rA   initializer_rangera   zero_r9   fill_)r@   moduler   r   r   _init_weightsL  s   
z"DinatPreTrainedModel._init_weightsN)r&   r'   r(   r   r,   base_model_prefixmain_input_namer  r   r   r   r   r  F  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
        nattenrK   r   r   N)r5   r6   r   rA   r   r   r   rd   r:   num_featuresr4   rF   r   encoderr   r9   r   	layernormAdaptiveAvgPool1dpooler	post_init)r@   rA   add_pooling_layerrB   r   r   r6   [  s   

zDinatModel.__init__c                 C      | j jS r   rF   r8   rv   r   r   r   get_input_embeddingsq     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   )r@   heads_to_pruner"  r   r   r   r   _prune_headst  s   zDinatModel._prune_headsNrD   r   r   r   rE   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   rK   )r"   r0   r#   r$   r%   )rA   r   r   use_return_dictrT   rF   r  r  r  flattenr   r*   r/   r#   r$   r%   )
r@   rD   r   r   r   embedding_outputencoder_outputssequence_outputpooled_outputrr   r   r   r   rG   |  s:   


zDinatModel.forward)T)NNNN)r&   r'   r(   r6   r  r$  r   r   r*   r+   r   r   r-   r/   rG   rI   r   r   rB   r   r  Y  s(    
r  z
    Dinat Model transformer with an image classification head on top (a linear layer on top of the final hidden state
    of the [CLS] token) e.g. for ImageNet.
    c                       sn   e Zd Z fddZe					ddeej deej dee	 dee	 dee	 d	e
eef fd
dZ  ZS )DinatForImageClassificationc                    s\   t  | t| dg |j| _t|| _|jdkr#t| jj|jnt	 | _
|   d S )Nr  r   )r5   r6   r   
num_labelsr  r  r   r   r  r   
classifierr  r?   rB   r   r   r6     s   
"z$DinatForImageClassification.__init__NrD   labelsr   r   r   rE   c                 C   s   |dur|n| j j}| j||||d}|d }| |}d}	|dur*| ||| j }	|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   rK   )r2   r3   r#   r$   r%   )	rA   r&  r  r.  loss_functionr1   r#   r$   r%   )r@   rD   r/  r   r   r   r   r+  r3   r2   rr   r   r   r   rG     s,   
z#DinatForImageClassification.forward)NNNNN)r&   r'   r(   r6   r   r   r*   r+   
LongTensorr   r   r-   r1   rG   rI   r   r   rB   r   r,    s*    
r,  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 )rK   )rd   r:   r   rA   r   r   r     s    z*DinatBackbone.__init__.<locals>.<listcomp>)r5   r6   _init_backboner   r4   rF   r   r  r:   r   r   r   r  zip_out_featuresr   r   r9   
ModuleDicthidden_states_normsr  )r@   rA   r8  stagerS   rB   r3  r   r6     s   

&zDinatBackbone.__init__c                 C   r  r   r  rv   r   r   r   r  	  r   z"DinatBackbone.get_input_embeddingsNrD   r   r   r   rE   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   rK   r   r   )feature_mapsr#   r$   )rA   r&  r   r   rF   r  r%   r5  stage_namesout_featuresrY   rZ   r   r   r8  r#   r   r$   )r@   rD   r   r   r   r(  r   r#   r:  r9  hidden_stater   rS   r\   r]   rr   r   r   r   rG     sD   !


zDinatBackbone.forward)NNN)r&   r'   r(   r6   r  r   r*   rH   r   r   r   rG   rI   r   r   rB   r   r2    s$    r2  )r,  r  r  r2  )rf   F)<r)   r   dataclassesr   typingr   r   r*   r   activationsr   modeling_outputsr   modeling_utilsr	   pytorch_utilsr
   r   utilsr   r   r   r   r   r   utils.backbone_utilsr   configuration_dinatr   natten.functionalr   r   
get_loggerr&   loggerr!   r/   r1   re   r4   r7   r^   rH   rx   r   rs   rt   r{   r   r   r   r   r   r   r   r  r  r,  r2  __all__r   r   r   r   <module>   sv    
$ F$G/FR>b