o
    wiR                  	   @   s  d 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mZmZmZ dd	lmZ dd
lmZmZ ddlmZ ddlmZ eeZd0d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$G dd dej"Z%G dd dej"Z&G dd dej"Z'G d d! d!ej"Z(G d"d# d#ej"Z)eG d$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*eZ-g d/Z.dS )1zPyTorch ConvNextV2 model.    )OptionalUnionN)nn)BCEWithLogitsLossCrossEntropyLossMSELoss   )ACT2FN)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)BackboneMixin   )ConvNextV2Config        Finput	drop_probtrainingreturnc                 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.
    r   r   r   )r   )dtypedevice)shapendimtorchrandr   r   floor_div)r   r   r   	keep_probr   random_tensoroutput r$   o/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/convnextv2/modeling_convnextv2.py	drop_path)   s   
r&   c                       sT   e Zd ZdZddee ddf fddZdejdejfdd	Z	de
fd
dZ  ZS )ConvNextV2DropPathzXDrop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).Nr   r   c                    s   t    || _d S N)super__init__r   )selfr   	__class__r$   r%   r*   A   s   

zConvNextV2DropPath.__init__hidden_statesc                 C   s   t || j| jS r(   )r&   r   r   r+   r.   r$   r$   r%   forwardE   s   zConvNextV2DropPath.forwardc                 C   s   d| j  S )Nzp=)r   )r+   r$   r$   r%   
extra_reprH   s   zConvNextV2DropPath.extra_reprr(   )__name__
__module____qualname____doc__r   floatr*   r   Tensorr0   strr1   __classcell__r$   r$   r,   r%   r'   >   s
    r'   c                       s<   e Zd ZdZdef fddZdejdejfddZ  Z	S )	ConvNextV2GRNz)GRN (Global Response Normalization) layerdimc                    s>   t    ttddd|| _ttddd|| _d S )Nr   )r)   r*   r   	Parameterr   zerosweightbias)r+   r;   r,   r$   r%   r*   O   s   
zConvNextV2GRN.__init__r.   r   c                 C   sF   t jj|dddd}||jdddd  }| j||  | j | }|S )N   )r   r@   T)ordr;   keepdim)r;   rB   ư>)r   linalgvector_normmeanr>   r?   )r+   r.   global_featuresnorm_featuresr$   r$   r%   r0   T   s   zConvNextV2GRN.forward)
r2   r3   r4   r5   intr*   r   FloatTensorr0   r9   r$   r$   r,   r%   r:   L   s    r:   c                       s8   e Zd ZdZd
 fdd	Zdejdejfdd	Z  ZS )ConvNextV2LayerNormaA  LayerNorm that supports two data formats: channels_last (default) or channels_first.
    The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shape (batch_size, height,
    width, channels) while channels_first corresponds to inputs with shape (batch_size, channels, height, width).
    rD   channels_lastc                    s`   t    tt|| _tt|| _|| _	|| _
| j
dvr*td| j
 |f| _d S )N)rM   channels_firstzUnsupported data format: )r)   r*   r   r<   r   onesr>   r=   r?   epsdata_formatNotImplementedErrornormalized_shape)r+   rS   rP   rQ   r,   r$   r%   r*   d   s   

zConvNextV2LayerNorm.__init__xr   c                 C   s   | j dkrtjj|| j| j| j| j}|S | j dkr]|j	}|
 }|jddd}|| djddd}|| t|| j  }|j|d}| jd d d d f | | jd d d d f  }|S )NrM   rN   r   T)rB   r@   )r   )rQ   r   r   
functional
layer_normrS   r>   r?   rP   r   r6   rG   powsqrtto)r+   rT   input_dtypeusr$   r$   r%   r0   n   s   
	
,zConvNextV2LayerNorm.forward)rD   rM   )	r2   r3   r4   r5   r*   r   r7   r0   r9   r$   r$   r,   r%   rL   ^   s    
rL   c                       s6   e Zd ZdZ fddZdejdejfddZ  Z	S )ConvNextV2EmbeddingszThis class is comparable to (and inspired by) the SwinEmbeddings class
    found in src/transformers/models/swin/modeling_swin.py.
    c                    sL   t    tj|j|jd |j|jd| _t|jd ddd| _	|j| _d S )Nr   kernel_sizestriderD   rN   rP   rQ   )
r)   r*   r   Conv2dnum_channelshidden_sizes
patch_sizepatch_embeddingsrL   	layernormr+   configr,   r$   r%   r*      s   
zConvNextV2Embeddings.__init__pixel_valuesr   c                 C   s4   |j d }|| jkrtd| |}| |}|S )Nr   zeMake sure that the channel dimension of the pixel values match with the one set in the configuration.)r   rc   
ValueErrorrf   rg   )r+   rj   rc   
embeddingsr$   r$   r%   r0      s   



zConvNextV2Embeddings.forward
r2   r3   r4   r5   r*   r   rK   r7   r0   r9   r$   r$   r,   r%   r]   }   s    r]   c                       s8   e Zd ZdZd	 fdd	ZdejdejfddZ  Z	S )
ConvNextV2Layera5  This corresponds to the `Block` class in the original implementation.

    There are two equivalent implementations: [DwConv, LayerNorm (channels_first), Conv, GELU,1x1 Conv]; all in (N, C,
    H, W) (2) [DwConv, Permute to (N, H, W, C), LayerNorm (channels_last), Linear, GELU, Linear]; Permute back

    The authors used (2) as they find it slightly faster in PyTorch.

    Args:
        config ([`ConvNextV2Config`]): Model configuration class.
        dim (`int`): Number of input channels.
        drop_path (`float`): Stochastic depth rate. Default: 0.0.
    r   c                    s   t    tj||dd|d| _t|dd| _t|d| | _t	|j
 | _td| | _td| || _|dkrAt|| _d S t | _d S )N   r   )r_   paddinggroupsrD   rP      r   )r)   r*   r   rb   dwconvrL   rg   Linearpwconv1r	   
hidden_actactr:   grnpwconv2r'   Identityr&   )r+   ri   r;   r&   r,   r$   r%   r*      s   
$zConvNextV2Layer.__init__r.   r   c                 C   sr   |}|  |}|dddd}| |}| |}| |}| |}| |}|dddd}|| | }|S )Nr   r@   r   r   )rt   permuterg   rv   rx   ry   rz   r&   )r+   r.   r   rT   r$   r$   r%   r0      s   





zConvNextV2Layer.forward)r   rm   r$   r$   r,   r%   rn      s    rn   c                       s8   e Zd ZdZd
 fdd	Zdejdejfdd	Z  Z	S )ConvNextV2Stagea  ConvNeXTV2 stage, consisting of an optional downsampling layer + multiple residual blocks.

    Args:
        config ([`ConvNextV2Config`]): Model configuration class.
        in_channels (`int`): Number of input channels.
        out_channels (`int`): Number of output channels.
        depth (`int`): Number of residual blocks.
        drop_path_rates(`list[float]`): Stochastic depth rates for each layer.
    r@   Nc              	      s   t    |ks|dkr!tt|dddtj|||d| _nt | _p,dg| tj fddt|D  | _	d S )	Nr   rD   rN   ra   r^   r   c                    s   g | ]}t  | d qS ))r;   r&   )rn   ).0jri   drop_path_ratesout_channelsr$   r%   
<listcomp>   s    z,ConvNextV2Stage.__init__.<locals>.<listcomp>)
r)   r*   r   
SequentialrL   rb   downsampling_layerr{   rangelayers)r+   ri   in_channelsr   r_   r`   depthr   r,   r   r%   r*      s   


zConvNextV2Stage.__init__r.   r   c                 C   s   |  |}| |}|S r(   )r   r   r/   r$   r$   r%   r0      s   

zConvNextV2Stage.forward)r@   r@   r@   Nrm   r$   r$   r,   r%   r}      s    
r}   c                       sN   e Zd Z fddZ		ddejdee dee dee	e
f fd	d
Z  ZS )ConvNextV2Encoderc              	      s   t    t | _dd tjd|jt|j	dd
|j	D }|jd }t|jD ]$}|j| }t||||dkr;dnd|j	| || d}| j| |}q*d S )	Nc                 S   s   g | ]}|  qS r$   )tolist)r~   rT   r$   r$   r%   r      s    z.ConvNextV2Encoder.__init__.<locals>.<listcomp>r   cpu)r   r@   r   )r   r   r`   r   r   )r)   r*   r   
ModuleListstagesr   linspacedrop_path_ratesumdepthssplitrd   r   
num_stagesr}   append)r+   ri   r   prev_chsiout_chsstager,   r$   r%   r*      s&   

 

zConvNextV2Encoder.__init__FTr.   output_hidden_statesreturn_dictr   c                 C   sj   |rdnd }t | jD ]\}}|r||f }||}q|r"||f }|s/tdd ||fD S t||dS )Nr$   c                 s   s    | ]	}|d ur|V  qd S r(   r$   )r~   vr$   r$   r%   	<genexpr>
  s    z,ConvNextV2Encoder.forward.<locals>.<genexpr>)last_hidden_stater.   )	enumerater   tupler   )r+   r.   r   r   all_hidden_statesr   layer_moduler$   r$   r%   r0      s   


zConvNextV2Encoder.forward)FT)r2   r3   r4   r*   r   rK   r   boolr   r   r   r0   r9   r$   r$   r,   r%   r      s    
r   c                   @   s&   e Zd ZeZdZdZdgZdd ZdS )ConvNextV2PreTrainedModel
convnextv2rj   rn   c                 C   s   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fr:|j	j
  |jjd dS t |trM|jj
  |j	j
  dS dS )zInitialize the weightsr   )rG   stdNg      ?)
isinstancer   ru   rb   r>   datanormal_ri   initializer_ranger?   zero_	LayerNormrL   fill_r:   )r+   moduler$   r$   r%   _init_weights  s   

z'ConvNextV2PreTrainedModel._init_weightsN)	r2   r3   r4   r   config_classbase_model_prefixmain_input_name_no_split_modulesr   r$   r$   r$   r%   r     s    r   c                       sX   e Zd Z fddZe			d
deej dee dee de	e
ef fdd	Z  ZS )ConvNextV2Modelc                    sJ   t  | || _t|| _t|| _tj|j	d |j
d| _|   d S )NrC   rr   )r)   r*   ri   r]   rl   r   encoderr   r   rd   layer_norm_epsrg   	post_initrh   r,   r$   r%   r*   ,  s   

zConvNextV2Model.__init__Nrj   r   r   r   c                 C   s   |d ur|n| j j}|d ur|n| j j}|d u rtd| |}| j|||d}|d }| |ddg}|sC||f|dd   S t|||j	dS )Nz You have to specify pixel_valuesr   r   r   rC   r   )r   pooler_outputr.   )
ri   r   use_return_dictrk   rl   r   rg   rG   r   r.   )r+   rj   r   r   embedding_outputencoder_outputsr   pooled_outputr$   r$   r%   r0   9  s(   
zConvNextV2Model.forward)NNN)r2   r3   r4   r*   r   r   r   rK   r   r   r   r   r0   r9   r$   r$   r,   r%   r   )  s    
r   z
    ConvNextV2 Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
    ImageNet.
    )custom_introc                       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 ) ConvNextV2ForImageClassificationc                    sR   t  | |j| _t|| _|jdkrt|jd |jnt | _	| 
  d S )Nr   rC   )r)   r*   
num_labelsr   r   r   ru   rd   r{   
classifierr   rh   r,   r$   r%   r*   g  s   
$z)ConvNextV2ForImageClassification.__init__Nrj   labelsr   r   r   c                 C   sb  |dur|n| j j}| j|||d}|r|jn|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   
regressionsingle_label_classificationmulti_label_classificationrC   r@   )losslogitsr.   )ri   r   r   r   r   problem_typer   r   r   longrJ   r   squeezer   viewr   r   r.   )r+   rj   r   r   r   outputsr   r   r   loss_fctr#   r$   r$   r%   r0   u  s>   


"


z(ConvNextV2ForImageClassification.forward)NNNN)r2   r3   r4   r*   r   r   r   rK   
LongTensorr   r   r   r   r0   r9   r$   r$   r,   r%   r   _  s$    
r   zT
    ConvNeXT V2 backbone, to be used with frameworks like DETR and MaskFormer.
    c                
       sJ   e Zd Z fddZe		d
dejdee dee de	fdd	Z
  ZS )ConvNextV2Backbonec                    s   t  | t  | t|| _t|| _|jd g|j | _i }t	| j
| jD ]\}}t|dd||< q)t|| _|   d S )Nr   rN   )rQ   )r)   r*   _init_backboner]   rl   r   r   rd   num_featureszip_out_featureschannelsrL   r   
ModuleDicthidden_states_normsr   )r+   ri   r   r   rc   r,   r$   r%   r*     s   

zConvNextV2Backbone.__init__Nrj   r   r   r   c                 C   s   |dur|n| j j}|dur|n| j j}| |}| j|d|d}|r&|jn|d }d}t| j|D ]\}}	|| jv rG| j	| |	}	||	f7 }q2|sV|f}
|rT|
|f7 }
|
S t
||r_|ddS dddS )ar  
        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("facebook/convnextv2-tiny-1k-224")
        >>> model = AutoBackbone.from_pretrained("facebook/convnextv2-tiny-1k-224")

        >>> inputs = processor(image, return_tensors="pt")
        >>> outputs = model(**inputs)
        ```NTr   r   r$   )feature_mapsr.   
attentions)ri   r   r   rl   r   r.   r   stage_namesout_featuresr   r
   )r+   rj   r   r   r   r   r.   r   r   hidden_stater#   r$   r$   r%   r0     s:   



zConvNextV2Backbone.forward)NN)r2   r3   r4   r*   r   r   r7   r   r   r
   r0   r9   r$   r$   r,   r%   r     s    r   )r   r   r   r   )r   F)/r5   typingr   r   r   torch.utils.checkpointr   torch.nnr   r   r   activationsr	   modeling_outputsr
   r   r   r   modeling_utilsr   utilsr   r   utils.backbone_utilsr   configuration_convnextv2r   
get_loggerr2   loggerr7   r6   r   r&   Moduler'   r:   rL   r]   rn   r}   r   r   r   r   r   __all__r$   r$   r$   r%   <module>   sD   
 ,!04FM