§
    l¬¸jž  ã                  ó\   — d Z ddlmZ ddlmZ ddlZddlmZ  G d„ dej        ¦  «        ZdS )zVNecks are the interface between a vision backbone and the rest of the detection model.é    )Úannotations)ÚdeepcopyN)Únnc                  óD   ‡ — e Zd ZdZ	 	 ddˆ fd„Zdd„Zdd„Zdd d„Zˆ xZS )!ÚSam3DualViTDetNeckzbA neck that implements a simple FPN as in ViTDet, with support for dual necks (for SAM3 and SAM2).©ç      @ç       @ç      ð?ç      à?FÚtrunkú	nn.ModuleÚposition_encodingÚd_modelÚintÚadd_sam2_neckÚboolc                óˆ  •— t          ¦   «                              ¦   «          || _        || _        t	          j        ¦   «         | _        || _        d}| j        j        d         }t          |¦  «        D �]ª\  }}	t	          j
        ¦   «         }
|	dk    rŽ|
                     dt	          j        ||dz  dd¬¦  «        ¦  «         |
                     dt	          j        ¦   «         ¦  «         |
                     dt	          j        |dz  |d	z  dd¬¦  «        ¦  «         |d	z  }nŠ|	d
k    r5|
                     dt	          j        ||dz  dd¬¦  «        ¦  «         |dz  }nO|	dk    r|}nF|	dk    r-|
                     dt	          j        dd¬¦  «        ¦  «         |}nt          d|	› d�¦  «        ‚|
                     dt	          j        ||d|¬¦  «        ¦  «         |
                     dt	          j        ||dd|¬¦  «        ¦  «         | j                             |
¦  «         �Œ¬d| _        |rt'          | j        ¦  «        | _        dS dS )a  SimpleFPN neck a la ViTDet, very lightly adapted from detectron2.

        Supports a "dual neck" setting with two identical necks (for SAM3 and SAM2) that have different weights.

        Args:
            trunk (nn.Module): The backbone.
            position_encoding (nn.Module): The positional encoding to use.
            d_model (int): The dimension of the model.
            scale_factors (tuple): Scale factors for each FPN level.
            add_sam2_neck (bool): Whether to add a second neck for SAM2.
        Téÿÿÿÿr	   Údconv_2x2_0é   )Úkernel_sizeÚstrideÚgeluÚdconv_2x2_1é   r
   Ú	dconv_2x2r   r   Úmaxpool_2x2zscale_factor=z is not supported yet.Úconv_1x1é   )Úin_channelsÚout_channelsr   ÚbiasÚconv_3x3é   )r!   r"   r   Úpaddingr#   N)ÚsuperÚ__init__r   r   r   Ú
ModuleListÚconvsÚscale_factorsÚchannel_listÚ	enumerateÚ
SequentialÚ
add_moduleÚConvTranspose2dÚGELUÚ	MaxPool2dÚNotImplementedErrorÚConv2dÚappendÚ
sam2_convsr   )Úselfr   r   r   r+   r   Úuse_biasÚdimÚ_ÚscaleÚcurrentÚout_dimÚ	__class__s               €úq/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/ultralytics/models/sam/sam3/necks.pyr(   zSam3DualViTDetNeck.__init__   s¡  ø€ õ& 	‰Œ×ÒÑÔÐØˆŒ
Ø!2ˆÔÝ”]‘_”_ˆŒ
à*ˆÔØˆØ”:Ô*¨2Ô.ˆå! -Ñ0Ô0ð 5	'ñ 5	'‰HˆAˆuÝ”m‘o”oˆGà˜Š|ˆ|Ø×"Ò"Ø!ÝÔ& s¨C°1©HÀ!ÈAÐNÑNÔNñô ð ð ×"Ò"ØÝ”G‘I”Iñô ð ð ×"Ò"Ø!ÝÔ& s¨a¡x°¸±ÀqÐQRÐSÑSÔSñô ð ð  ™(��Ø˜#’�Ø×"Ò"ØÝÔ& s¨C°1©HÀ!ÈAÐNÑNÔNñô ð ð  ™(��Ø˜#’�Ø��Ø˜#’�Ø×"Ò"Ø!Ý”L¨Q°qÐ9Ñ9Ô9ñô ð ð ��å)Ð*W¸%Ð*WÐ*WÐ*WÑXÔXÐXà×ÒØÝ”	Ø 'Ø!(Ø !Ø!ð	ñ ô ñô ð ð ×ÒØÝ”	Ø 'Ø!(Ø !ØØ!ðñ ô ñ	ô 	ð 	ð ŒJ×Ò˜gÑ&Ô&Ð&Ñ&àˆŒØð 	3å& t¤zÑ2Ô2ˆDŒOˆOˆOð	3ð 	3ó    Útensor_listúlist[torch.Tensor]Úreturnúctuple[list[torch.Tensor], list[torch.Tensor], list[torch.Tensor] | None, list[torch.Tensor] | None]c                óÚ   — |                       |¦  «        }|d         }|                      || j        ¦  «        \  }}| j        €||ddfS |                      || j        ¦  «        \  }}||||fS )z8Get feature maps and positional encodings from the neck.r   N)r   Úsam_forward_feature_levelsr*   r6   )r7   rA   ÚxsÚxÚsam3_outÚsam3_posÚsam2_outÚsam2_poss           r?   ÚforwardzSam3DualViTDetNeck.forwardj   s|   € ð �ZŠZ˜Ñ$Ô$ˆØˆrŒFˆØ!×<Ò<¸QÀÄ
ÑKÔKÑˆ�(ØŒ?Ð"Ø˜X t¨TÐ1Ð1Ø!×<Ò<¸QÀÄÑPÔPÑˆ�(Ø˜ 8¨XÐ5Ð5r@   rH   útorch.Tensorr*   únn.ModuleListú-tuple[list[torch.Tensor], list[torch.Tensor]]c                óÜ   — g g }}|D ]b} ||¦  «        }|                      |¦  «         |                      |                      |¦  «                             |j        ¦  «        ¦  «         Œc||fS )zNRun neck convolutions and compute positional encodings for each feature level.)r5   r   ÚtoÚdtype)r7   rH   r*   ÚoutsÚpossÚconvÚfeats          r?   rF   z-Sam3DualViTDetNeck.sam_forward_feature_levelsv   sz   € ð ˜ˆdˆØð 	Eð 	EˆDØ�4˜‘7”7ˆDØ�KŠK˜ÑÔÐØ�KŠK˜×.Ò.¨tÑ4Ô4×7Ò7¸¼
ÑCÔCÑDÔDÐDÐDØ�TˆzÐr@   NÚimgszúlist[int] | Nonec                óJ   — |�|nddg}| j                              |¦  «         dS )z*Set the image size for the trunk backbone.Nið  )r   Ú	set_imgsz)r7   rX   s     r?   r[   zSam3DualViTDetNeck.set_imgsz�   s0   € àÐ*��°°t°ˆØŒ
×Ò˜UÑ#Ô#Ð#Ð#Ð#r@   )r   F)r   r   r   r   r   r   r   r   )rA   rB   rC   rD   )rH   rN   r*   rO   rC   rP   )N)rX   rY   )	Ú__name__Ú
__module__Ú__qualname__Ú__doc__r(   rM   rF   r[   Ú__classcell__)r>   s   @r?   r   r      s™   ø€ € € € € ØlÐlð +Ø#ðV3ð V3ð V3ð V3ð V3ð V3ð V3ðp
6ð 
6ð 
6ð 
6ð	ð 	ð 	ð 	ð$ð $ð $ð $ð $ð $ð $ð $ð $r@   r   )	r_   Ú
__future__r   Úcopyr   Útorchr   ÚModuler   © r@   r?   ú<module>rf      s�   ðð
 ]Ð \à "Ð "Ð "Ð "Ð "Ð "à Ð Ð Ð Ð Ð à €€€Ø Ð Ð Ð Ð Ð ðu$ð u$ð u$ð u$ð u$˜œñ u$ô u$ð u$ð u$ð u$r@   