o
    iB                  	   @   s  d Z ddlmZ ddlZddl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 ddlmZ eeZd.d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$G d d! d!ejZ%eG d"d# d#eZ&eG d$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 )/zPyTorch ConvNext model.    )OptionalN)nn   )ACT2FN)BackboneOutputBaseModelOutputWithNoAttention(BaseModelOutputWithPoolingAndNoAttention$ImageClassifierOutputWithNoAttention)PreTrainedModel)auto_docstringlogging)BackboneMixin)can_return_tuple   )ConvNextConfig        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!   k/home/ubuntu/veenaModal/venv/lib/python3.10/site-packages/transformers/models/convnext/modeling_convnext.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 )ConvNextDropPathzXDrop 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'   @   s   

zConvNextDropPath.__init__hidden_statesc                 C   s   t || j| jS r%   )r#   r   r   )r(   r+   r!   r!   r"   forwardD   s   zConvNextDropPath.forwardc                 C   s   d| j  S )Nzp=)r   )r(   r!   r!   r"   
extra_reprG   s   zConvNextDropPath.extra_reprr%   )__name__
__module____qualname____doc__r   floatr'   r   Tensorr,   strr-   __classcell__r!   r!   r)   r"   r$   =   s
    r$   c                       sB   e Zd ZdZddd fdd
Zdejdejf fd	d
Z  ZS )ConvNextLayerNormaA  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).
    ư>channels_lastepsdata_formatc                   s8   t  j|fd|i| |dvrtd| || _d S )Nr:   )r8   channels_firstzUnsupported data format: )r&   r'   NotImplementedErrorr;   )r(   normalized_shaper:   r;   kwargsr)   r!   r"   r'   Q   s   
zConvNextLayerNorm.__init__featuresr   c                    sJ   | j dkr|dddd}t |}|dddd}|S t |}|S )z
        Args:
            features: Tensor of shape (batch_size, channels, height, width) OR (batch_size, height, width, channels)
        r<   r      r   r   )r;   permuter&   r,   )r(   r@   r)   r!   r"   r,   W   s   
zConvNextLayerNorm.forward	r.   r/   r0   r1   r'   r   r3   r,   r5   r!   r!   r)   r"   r6   K   s    "r6   c                       s6   e Zd ZdZ fddZdejdejfddZ  Z	S )ConvNextEmbeddingszThis 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strider7   r<   r9   )
r&   r'   r   Conv2dnum_channelshidden_sizes
patch_sizepatch_embeddingsr6   	layernormr(   configr)   r!   r"   r'   j   s   
zConvNextEmbeddings.__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   rI   
ValueErrorrL   rM   )r(   rP   rI   
embeddingsr!   r!   r"   r,   r   s   



zConvNextEmbeddings.forward)
r.   r/   r0   r1   r'   r   FloatTensorr3   r,   r5   r!   r!   r)   r"   rD   e   s    rD   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 )
ConvNextLayera3  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 ([`ConvNextConfig`]): 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| || _|jdkrAtj|jt| dd	nd | _|d
krOt|| _d S t | _d S )N   r   )rF   paddinggroupsr7   r:      r   T)requires_gradr   )r&   r'   r   rH   dwconvr6   rM   Linearpwconv1r   
hidden_actactpwconv2layer_scale_init_value	Parameterr   oneslayer_scale_parameterr$   Identityr#   )r(   rO   dimr#   r)   r!   r"   r'      s   

$zConvNextLayer.__init__r@   r   c                 C   s|   |}|  |}|dddd}| |}| |}| |}| |}| jd ur-| j| }|dddd}|| | }|S )Nr   rA   r   r   )r[   rB   rM   r]   r_   r`   rd   r#   )r(   r@   residualr!   r!   r"   r,      s   






zConvNextLayer.forward)r   rC   r!   r!   r)   r"   rT   }   s    rT   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 )ConvNextStagea  ConvNeXT stage, consisting of an optional downsampling layer + multiple residual blocks.

    Args:
        config ([`ConvNextConfig`]): 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.
    rA   Nc              	      s   t    |ks|dkr"tt|dddtj|||dg| _nt | _p-dg| t fddt|D | _d S )	Nr   r7   r<   r9   rE   r   c                    s   g | ]}t  | d qS ))rf   r#   )rT   ).0jrO   drop_path_ratesout_channelsr!   r"   
<listcomp>   s    z*ConvNextStage.__init__.<locals>.<listcomp>)	r&   r'   r   
ModuleListr6   rH   downsampling_layerrangelayers)r(   rO   in_channelsrm   rF   rG   depthrl   r)   rk   r"   r'      s   


zConvNextStage.__init__r@   r   c                 C   s,   | j D ]}||}q| jD ]}||}q|S r%   )rp   rr   )r(   r@   layerr!   r!   r"   r,      s
   



zConvNextStage.forward)rA   rA   rA   NrC   r!   r!   r)   r"   rh      s    
rh   c                       s<   e Zd Z fddZ	d	dejdee defddZ	  Z
S )
ConvNextEncoderc              	      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)ri   xr!   r!   r"   rn      s    z,ConvNextEncoder.__init__.<locals>.<listcomp>r   cpu)r   rA   r   )rs   rm   rG   rt   rl   )r&   r'   r   ro   stagesr   linspacedrop_path_ratesumdepthssplitrJ   rq   
num_stagesrh   append)r(   rO   rl   prev_chsiout_chsstager)   r!   r"   r'      s&   

 

zConvNextEncoder.__init__Fr+   output_hidden_statesr   c                 C   s@   |r|gnd }| j D ]}||}|d ur|| q
t||dS )N)last_hidden_stater+   )rz   r   r   )r(   r+   r   all_hidden_stateslayer_moduler!   r!   r"   r,      s   

zConvNextEncoder.forward)F)r.   r/   r0   r'   r   r3   r   boolr   r,   r5   r!   r!   r)   r"   rv      s    rv   c                   @   s0   e Zd ZU eed< dZdZdgZi Zdd Z	dS )ConvNextPreTrainedModelrO   convnextrP   rT   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rO|jdurQ|jj| jj dS dS dS )zInitialize the weightsr   )meanstdNg      ?)
isinstancer   r\   rH   weightdatanormal_rO   initializer_rangebiaszero_	LayerNormr6   fill_rT   rd   ra   )r(   moduler!   r!   r"   _init_weights   s   


z%ConvNextPreTrainedModel._init_weightsN)
r.   r/   r0   r   __annotations__base_model_prefixmain_input_name_no_split_modules_can_record_outputsr   r!   r!   r!   r"   r      s   
 r   c                	       sH   e Zd Z fddZee	d	deej dee	 de
fddZ  ZS )
ConvNextModelc                    sJ   t  | || _t|| _t|| _tj|j	d |j
d| _|   d S )NrX   )r&   r'   rO   rD   rR   rv   encoderr   r   rJ   layer_norm_epsrM   	post_initrN   r)   r!   r"   r'   	  s   

zConvNextModel.__init__NrP   r   r   c                 C   sb   |d u r| j j}|d u rtd| |}| j||d}|j}| |ddg}t|||j	dS )Nz You have to specify pixel_valuesr   r   )r   pooler_outputr+   )
rO   r   rQ   rR   r   r   rM   r   r   r+   )r(   rP   r   embedding_outputencoder_outputsr   pooled_outputr!   r!   r"   r,     s   
zConvNextModel.forwardNN)r.   r/   r0   r'   r   r   r   r   rS   r   r   r,   r5   r!   r!   r)   r"   r     s    r   z
    ConvNext Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
    ImageNet.
    )custom_introc                	       sN   e Zd ZdZ fddZee	d
deej	 deej
 defdd	Z  ZS )ConvNextForImageClassificationFc                    sV   t  | |j| _t|| _|jdkr t|jd |j| _nt	 | _| 
  d S )Nr   r   )r&   r'   
num_labelsr   r   r   r\   rJ   
classifierre   r   rN   r)   r!   r"   r'   :  s   


z'ConvNextForImageClassification.__init__NrP   labelsr   c                 K   sP   | j |fi |}|j}| |}d}|dur | j||| jd}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).
        N)r   pooled_logitsrO   )losslogitsr+   )r   r   r   loss_functionrO   r	   r+   )r(   rP   r   r?   outputsr   r   r   r!   r!   r"   r,   I  s   
z&ConvNextForImageClassification.forwardr   )r.   r/   r0   accepts_loss_kwargsr'   r   r   r   r   rS   
LongTensorr	   r,   r5   r!   r!   r)   r"   r   1  s    r   zQ
    ConvNeXt backbone, to be used with frameworks like DETR and MaskFormer.
    c                	       sH   e Zd ZdZ fddZee	d
dejde	e
 defdd	Z  ZS )ConvNextBackboneFc                    s   t  | t  | t|| _t|| _|jd g|j | _i }t	| j
| jD ]\}}t|dd||< q)t|| _|   d S )Nr   r<   )r;   )r&   r'   _init_backbonerD   rR   rv   r   rJ   num_featureszip_out_featureschannelsr6   r   
ModuleDicthidden_states_normsr   )r(   rO   r   r   rI   r)   r!   r"   r'   k  s   

zConvNextBackbone.__init__NrP   r   r   c           	      C   s   |du r| j j}| |}| j|dd}|j}g }t| j|D ]\}}|| jv r4| j| |}|	| qt
t||r?|dS ddS )ah  
        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/convnext-tiny-224")
        >>> model = AutoBackbone.from_pretrained("facebook/convnext-tiny-224")

        >>> inputs = processor(image, return_tensors="pt")
        >>> outputs = model(**inputs)
        ```NTr   )feature_mapsr+   )rO   r   rR   r   r+   r   stage_namesout_featuresr   r   r   tuple)	r(   rP   r   r   r   r+   r   r   hidden_stater!   r!   r"   r,   |  s"   


zConvNextBackbone.forwardr%   )r.   r/   r0   has_attentionsr'   r   r   r   r3   r   r   r   r,   r5   r!   r!   r)   r"   r   c  s    r   )r   r   r   r   )r   F)+r1   typingr   r   r   activationsr   modeling_outputsr   r   r   r	   modeling_utilsr
   utilsr   r   utils.backbone_utilsr   utils.genericr   configuration_convnextr   
get_loggerr.   loggerr3   r2   r   r#   Moduler$   r   r6   rD   rT   rh   rv   r   r   r   r   __all__r!   r!   r!   r"   <module>   s@   
 +$#),@