§
    ·¯¸jQ•  ã                   ó’  — d Z ddlmZ ddlmZmZmZmZmZm	Z	m
Z
mZ ddl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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gZ% G d„ dej&        ¦  «        Z' G d„ dej&        ¦  «        Z( G d„ dej&        ¦  «        Z) G d„ dej&        ¦  «        Z* ej+        ¦   «         d?dej,        deej-                 ddfd„¦   «         Z. G d„ dej&        ¦  «        Z/ G d„ dej&        ¦  «        Z0 G d„ d ej&        ¦  «        Z1 G d!„ d"ej&        ¦  «        Z2 G d#„ dej&        ¦  «        Z3 G d$„ de3¦  «        Z4d%ee5ej-        f         d&ej&        dee5ej-        f         fd'„Z6d(edee5ef         fd)„Z7 e$ e7d*¬+¦  «         e7d*¬+¦  «         e7d*d,¬-¦  «         e7d*¬+¦  «         e7d*¬+¦  «         e7d*d,¬-¦  «         e7d*¬+¦  «         e7d*¬+¦  «         e7d*d,¬-¦  «         e7d*d.¬/¦  «         e7d*d.¬/¦  «         e7d*d.d,¬0¦  «        d1œ¦  «        Z8	 	 d@d3e5d4e9d5e9d(ede3f
d6„Z:e#dAd4e9d(ede3fd7„¦   «         Z;e#dAd4e9d(ede3fd8„¦   «         Z<e#dAd4e9d(ede3fd9„¦   «         Z=e#dAd4e9d(ede3fd:„¦   «         Z>e#dAd4e9d(ede3fd;„¦   «         Z?e#dAd4e9d(ede3fd<„¦   «         Z@e#dAd4e9d(ede3fd=„¦   «         ZAe#dAd4e9d(ede3fd>„¦   «         ZBdS )Baí  EfficientViM

Efficient ViM: Efficient Vision Mamba with Hidden State Mixer based State Space Duality
- paper: https://arxiv.org/abs/2411.15241 (CVPR 2025)
- code: https://github.com/mlvlab/EfficientViM

Hidden State Mixer based State Space Duality (HSM-SSD) treats the recurrent Mamba-style
state update as a chunked dual (attention-like) form. The mixer here is intentionally
plain PyTorch -- conv, softmax, matmul -- with no mamba_ssm / causal_conv1d / Triton
dependency, so the module remains CPU- and TorchScript-friendly.

The classification head is a 4-way softmax-weighted fusion rather than a single Linear:
three heads consume the pooled per-stage hidden states and the fourth the pooled final
feature map. The distillation variant (`_dist` models) adds a parallel `heads_dist` /
`weights_dist` path that returns a `(logits, logits_dist)` tuple when training and their
mean in eval, mirroring `LevitDistilled`.

For deployment, timm.utils.reparameterize_model(model.eval()) folds Conv-BatchNorm
pairs and depthwise residual branches without changing the feature or classifier APIs.

Ported from the official implementation (MIT license, Copyright (c) 2024 MLVlab), fvcore
FLOP-counting helper dropped.

Modifications by / Copyright 2025 Ross Wightman
é    )Úpartial)ÚAnyÚDictÚListÚOptionalÚSetÚTupleÚTypeÚUnionN)Úfuse_conv_bn_eval©ÚIMAGENET_DEFAULT_MEANÚIMAGENET_DEFAULT_STD)ÚDropPathÚ	LayerNormÚLayerNorm2dÚSqueezeExciteÚcalculate_drop_path_ratesÚtrunc_normal_Úget_device_dtypeé   )Úbuild_model_with_cfg)Úfeature_take_indices)Ú
checkpoint)Úregister_modelÚgenerate_default_cfgsÚEfficientViMÚEfficientViMDistilledc                   ó  ‡ — e Zd ZdZdddddej        ej        dddf
deded	ed
ededededee	ej
                          dee	ej
                          defˆ fd„Z ej        ¦   «         dej
        fd„¦   «         Zdej        dej        fd„Zˆ xZS )ÚConvLayer1dzJConv1d + optional norm + optional act, as in the reference implementation.é   r   r   ç      ð?NÚin_dimÚout_dimÚkernel_sizeÚstrideÚpaddingÚdilationÚgroupsÚnormÚ	act_layerÚbn_weight_initc           	      óê  •— t          ¦   «                              ¦   «          ||dœ}t          j        |||f||||ddœ|¤Ž| _        |r	 ||fi |¤Žnd | _        |	r
 |	¦   «         nd | _        | j        �t          | j        t          j        j	        j
        ¦  «        rXt          j                             | j        j        |
¦  «         t          j                             | j        j        d¦  «         d S d S d S ©N©ÚdeviceÚdtypeF)r&   r'   r(   r)   Úbiasr   )ÚsuperÚ__init__ÚnnÚConv1dÚconvr*   ÚactÚ
isinstanceÚmodulesÚ	batchnormÚ
_BatchNormÚinitÚ	constant_Úweightr2   ©Úselfr#   r$   r%   r&   r'   r(   r)   r*   r+   r,   r0   r1   ÚddÚ	__class__s                 €úh/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/timm/models/efficientvim.pyr4   zConvLayer1d.__init__7   ó  ø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆÝ”IØØØð

ð ØØØØð

ð 

ð ð

ð 

ˆŒ	ð ,0Ð9�D�D˜Ð'Ð' BÐ'Ð'Ð'°TˆŒ	Ø"+Ð5�9�9‘;”;�;°ˆŒØŒ9Ð ¥Z°´	½2¼:Ô;OÔ;ZÑ%[Ô%[Ð ÝŒG×Ò˜dœiÔ.°Ñ?Ô?Ð?ÝŒG×Ò˜dœiœn¨aÑ0Ô0Ð0Ð0Ð0ð !Ð Ð Ð ó    Úreturnc                 ó¤   — t          | j        t          j        j        j        ¦  «        r&t          | j        | j        ¦  «        | _        d| _        | S ©z3Fold BatchNorm into the convolution for evaluation.N©r9   r*   r5   r:   r;   r<   r   r7   ©rA   s    rD   ÚfusezConvLayer1d.fuseY   óA   € õ �d”i¥¤Ô!5Ô!@ÑAÔAð 	Ý)¨$¬)°T´YÑ?Ô?ˆDŒIØˆDŒIØˆrF   Úxc                 ó    — |                       |¦  «        }| j        �|                      |¦  «        }| j        �|                      |¦  «        }|S ©N©r7   r*   r8   ©rA   rN   s     rD   ÚforwardzConvLayer1d.forwarda   óD   € Ø�IŠI�a‰LŒLˆØŒ9Ð Ø—	’	˜!‘”ˆAØŒ8ÐØ—’˜‘”ˆAØˆrF   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__r5   ÚBatchNorm1dÚReLUÚintr   r
   ÚModuleÚfloatr4   ÚtorchÚno_gradrL   ÚTensorrS   Ú__classcell__©rC   s   @rD   r    r    4   ó<  ø€ € € € € ØTÐTð  !ØØØØØ.0¬nØ35´7Ø$'ØØð 1ð  1àð 1ð ð 1ð ð	 1ð
 ð 1ð ð 1ð ð 1ð ð 1ð ˜4 ¤	œ?Ô+ð 1ð    R¤Y¤Ô0ð 1ð "ð 1ð  1ð  1ð  1ð  1ð  1ðD €U„]�_„_ð�b”ið ð ð ñ „_ðð˜œð ¨%¬,ð ð ð ð ð ð ð ð rF   r    c                   ó  ‡ — e Zd ZdZdddddej        ej        dddf
deded	ed
ededededee	ej
                          dee	ej
                          defˆ fd„Z ej        ¦   «         dej
        fd„¦   «         Zdej        dej        fd„Zˆ xZS )ÚConvLayer2dzJConv2d + optional norm + optional act, as in the reference implementation.r!   r   r   r"   Nr#   r$   r%   r&   r'   r(   r)   r*   r+   r,   c           	      óê  •— t          ¦   «                              ¦   «          ||dœ}t          j        |||f||||ddœ|¤Ž| _        |r	 ||fi |¤Žnd | _        |	r
 |	¦   «         nd | _        | j        �t          | j        t          j        j	        j
        ¦  «        rXt          j                             | j        j        |
¦  «         t          j                             | j        j        d¦  «         d S d S d S r.   )r3   r4   r5   ÚConv2dr7   r*   r8   r9   r:   r;   r<   r=   r>   r?   r2   r@   s                 €rD   r4   zConvLayer2d.__init__m   rE   rF   rG   c                 ó¤   — t          | j        t          j        j        j        ¦  «        r&t          | j        | j        ¦  «        | _        d| _        | S rI   rJ   rK   s    rD   rL   zConvLayer2d.fuse�   rM   rF   rN   c                 ó    — |                       |¦  «        }| j        �|                      |¦  «        }| j        �|                      |¦  «        }|S rP   rQ   rR   s     rD   rS   zConvLayer2d.forward—   rT   rF   )rU   rV   rW   rX   r5   ÚBatchNorm2drZ   r[   r   r
   r\   r]   r4   r^   r_   rL   r`   rS   ra   rb   s   @rD   re   re   j   rc   rF   re   c                   óV   ‡ — e Zd ZdZ	 	 d	dedefˆ fd„Zdej        dej        fd„Zˆ xZ	S )
ÚStemz)4-layer stride-2 conv stem, reduction 16.Nr#   Údimc                 ó0  •— t          ¦   «                              ¦   «          ||dœ}t          j        t	          ||dz  fddddœ|¤Žt	          |dz  |dz  fddddœ|¤Žt	          |dz  |dz  fddddœ|¤Žt	          |dz  |fdddd dœ|¤Ž¦  «        | _        d S )	Nr/   é   r!   é   r   )r%   r&   r'   é   )r%   r&   r'   r+   )r3   r4   r5   Ú
Sequentialre   r7   ©rA   r#   rm   r0   r1   rB   rC   s         €rD   r4   zStem.__init__£   sÖ   ø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆÝ”MÝ˜  q¡ÐS°aÀÈ1ÐSÐSÐPRÐSÐSÝ˜˜q™ #¨¡(ÐU¸À!ÈQÐUÐUÐRTÐUÐUÝ˜˜q™ #¨¡(ÐU¸À!ÈQÐUÐUÐRTÐUÐUÝ˜˜q™ #Ð`°1¸QÈÐUYÐ`Ð`Ð]_Ð`Ð`ñ	
ô 
ˆŒ	ˆ	ˆ	rF   rN   rG   c                 ó,   — |                       |¦  «        S rP   )r7   rR   s     rD   rS   zStem.forward³   s   € Ø�yŠy˜‰|Œ|ÐrF   ©NN)
rU   rV   rW   rX   r[   r4   r^   r`   rS   ra   rb   s   @rD   rl   rl       sˆ   ø€ € € € € Ø3Ð3ð Øð
ð 
àð
ð ð
ð 
ð 
ð 
ð 
ð 
ð ˜œð ¨%¬,ð ð ð ð ð ð ð ð rF   rl   c                   óR   ‡ — e Zd Z	 	 ddedefˆ fd„Zdej        dej        fd„Zˆ xZS )	ÚFFNNr#   rm   c                 ó¨   •— t          ¦   «                              ¦   «          ||dœ}t          ||dfi |¤Ž| _        t          ||dfd ddœ|¤Ž| _        d S )Nr/   r   r   )r+   r,   )r3   r4   re   Úfc1Úfc2rs   s         €rD   r4   zFFN.__init__¸   sj   ø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆÝ˜v s¨AÐ4Ð4°Ð4Ð4ˆŒÝ˜s F¨AÐV¸ÈaÐVÐVÐSUÐVÐVˆŒˆˆrF   rN   rG   c                 óR   — |                       |                      |¦  «        ¦  «        S rP   )rz   ry   rR   s     rD   rS   zFFN.forwardÄ   s   € Ø�xŠx˜Ÿš ™œÑ$Ô$Ð$rF   ru   )	rU   rV   rW   r[   r4   r^   r`   rS   ra   rb   s   @rD   rw   rw   ·   s‹   ø€ € € € € ð
 Øð
Wð 
Wàð
Wð ð
Wð 
Wð 
Wð 
Wð 
Wð 
Wð%˜œð %¨%¬,ð %ð %ð %ð %ð %ð %ð %ð %rF   rw   r7   ÚscalerG   c                 óÔ  — | j         | j        cxk    r| j        k    sn J ‚| j        dk    r| j        dk    sJ ‚| j        \  }}|€#| j        dd…d|dz  |dz  fxx         dz  cc<   dS |                     d¦  «        }| j                             |dd…dddf         ¦  «         | j	        �| j	                             |¦  «         | j        dd…d|dz  |dz  fxx         d|z
  z  cc<   dS )zOFold a residual identity and optional branch gate into a depthwise convolution.)r   r   Nr   rp   r   éÿÿÿÿ)
r)   Úin_channelsÚout_channelsr&   r(   r%   r?   ÚreshapeÚmul_r2   )r7   r|   ÚkhÚkws       rD   Ú_fuse_depthwise_residualr…   È   s,  € ð Œ;˜$Ô*Ð?Ð?Ò?Ð?¨dÔ.?Ò?Ð?Ð?Ð?Ð?Ð?ØŒ;˜&Ò Ð  T¤]°fÒ%<Ð%<Ð%<Ð<ØÔ�F€BˆØ€}ØŒ�A�A�A�q˜" ™' 2¨¡7Ð*Ð+Ð+Ô+¨qÑ0Ð+Ð+Ñ+Ð+Ð+à—’˜bÑ!Ô!ˆØŒ×Ò˜˜q˜q˜q $¨¨dÐ2Ô3Ñ4Ô4Ð4ØŒ9Ð ØŒI�NŠN˜5Ñ!Ô!Ð!ØŒ�A�A�A�q˜" ™' 2¨¡7Ð*Ð+Ð+Ô+¨q°5©yÑ8Ð+Ð+Ñ+Ð+Ð+rF   c                   óÈ   ‡ — e Zd ZU dZej        j        e         ed<   	 	 	 dde	de	de
fˆ fd„Z ej        ¦   «         d	ej        fd
„¦   «         Zdej        d	ej        fd„Zˆ xZS )ÚPatchMergingzXMBConv-style downsample with SE, plus the reference implementation's depthwise bypasses.Úfusedç      @Nr#   r$   Úratioc                 ó�  •— t          ¦   «                              ¦   «          ||dœ}d| _        t          ||z  ¦  «        }t	          j        t          ||fddi|¤Žt          ||fddd|dœ|¤Žt          |dfi |¤Žt          ||fdd d	œ|¤Ž¦  «        | _        t          ||dfd|d d
œ|¤Ž| _	        t          ||dfd|d d
œ|¤Ž| _
        d S )Nr/   Fr%   r   r!   rp   )r%   r&   r'   r)   g      Ð?)r%   r+   )r'   r)   r+   )r3   r4   rˆ   r[   r5   rr   re   r   r7   Údwconv1Údwconv2)	rA   r#   r$   rŠ   r0   r1   rB   Ú
hidden_dimrC   s	           €rD   r4   zPatchMerging.__init__Ý   s  ø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆØˆŒ
Ý˜ 5™Ñ)Ô)ˆ
Ý”MÝ˜ 
Ð@Ð@¸Ð@¸RÐ@Ð@Ý˜
 JÐl¸AÀaÐQRÐ[eÐlÐlÐikÐlÐlÝ˜* dÐ1Ð1¨bÐ1Ð1Ý˜
 GÐQ¸ÀdÐQÐQÈbÐQÐQñ	
ô 
ˆŒ	õ # 6¨6°1Ðe¸aÈÐZ^ÐeÐeÐbdÐeÐeˆŒÝ" 7¨G°QÐhÀÈ'Ð]aÐhÐhÐegÐhÐhˆŒˆˆrF   rG   c                 ó  — | j         s|| j        r
J d¦   «         ‚| j                             ¦   «          | j                             ¦   «          t          | j        j        ¦  «         t          | j        j        ¦  «         d| _         | S )z4Fold the depthwise residual branches for evaluation.ú Fusion requires evaluation mode.T)rˆ   ÚtrainingrŒ   rL   r�   r…   r7   rK   s    rD   rL   zPatchMerging.fuseò   s   € ð Œzð 	Ø”}ÐHÐHÐ&HÑHÔHÐ$ØŒL×ÒÑÔÐØŒL×ÒÑÔÐÝ$ T¤\Ô%6Ñ7Ô7Ð7Ý$ T¤\Ô%6Ñ7Ô7Ð7ØˆDŒJØˆrF   rN   c                 ó   — | j         r|                      |¦  «        n||                      |¦  «        z   }|                      |¦  «        }| j         r|                      |¦  «        n||                      |¦  «        z   }|S rP   )rˆ   rŒ   r7   r�   rR   s     rD   rS   zPatchMerging.forwardþ   sh   € Ø#œzÐBˆD�LŠL˜‰OŒOˆO¨q°4·<²<À±?´?Ñ/BˆØ�IŠI�a‰LŒLˆØ#œzÐBˆD�LŠL˜‰OŒOˆO¨q°4·<²<À±?´?Ñ/BˆØˆrF   )r‰   NN)rU   rV   rW   rX   r^   ÚjitÚFinalÚboolÚ__annotations__r[   r]   r4   r_   r5   r\   rL   r`   rS   ra   rb   s   @rD   r‡   r‡   Ø   sç   ø€ € € € € € ØbÐbàŒ9Œ?˜4Ô Ð Ð Ñ ð ØØðið iàðið ðið ð	ið ið ið ið ið ið* €U„]�_„_ð	�b”ið 	ð 	ð 	ñ „_ð	ð˜œð ¨%¬,ð ð ð ð ð ð ð ð rF   r‡   c                   ó–   ‡ — e Zd ZdZ	 	 	 	 	 ddededed	eeef         fˆ fd
„Zdd„Zde	j
        dee	j
        e	j
        f         fd„Zˆ xZS )ÚHSMSSDax  Hidden State Mixer based State Space Duality.

    The Mamba-style gated recurrence is computed in its dual form: input-dependent
    `A` (via `dt`) and `B` combine into a chunk transition matrix applied to the
    flattened spatial tokens, then the resulting hidden state is read out with `C`.
    Everything is expressed with conv / softmax / matmul, no custom kernels.
    r"   é@   ©r"   g      0@Nrm   Ú
ssd_expandÚ	state_dimÚa_init_rangec                 óº  •— t          ¦   «                              ¦   «          ||dœ}|| _        t          | j        |z  ¦  «        | _        || _        t          |d|z  dfd d dœ|¤Ž| _        | j        dz  }t          ||dddf|d d ddœ|¤Ž| _	        t          |d| j        z  dfd d dœ|¤Ž| _
        t          | j        |dfd d ddœ|¤Ž| _        || _        t          j        t          j        | j        fi |¤Ž¦  «        | _        t          j        ¦   «         | _        t          j        t          j        d||¬¦  «        ¦  «        | _        |                      ¦   «          d S )	Nr/   r!   r   )r*   r+   r   )r)   r*   r+   r,   rp   )r*   r+   r,   )r3   r4   r›   r[   Úd_innerrœ   r    Ú	BCdt_projre   ÚdwÚhz_projÚout_projr�   r5   Ú	Parameterr^   ÚemptyÚAÚSiLUr8   ÚonesÚDÚreset_parameters)
rA   rm   r›   rœ   r�   r0   r1   rB   Úconv_dimrC   s
            €rD   r4   zHSMSSD.__init__  s‹  ø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆØ$ˆŒÝ˜4œ?¨SÑ0Ñ1Ô1ˆŒØ"ˆŒå$ S¨!¨i©-¸Ð\ÀÐQUÐ\Ð\ÐY[Ð\Ð\ˆŒØ”> AÑ%ˆÝØØØØØð
ð ØØØð
ð 
ð ð
ð 
ˆŒõ # 3¨¨D¬LÑ(8¸!Ð]À$ÐRVÐ]Ð]ÐZ\Ð]Ð]ˆŒÝ# D¤L°#°qÐl¸tÈtÐdeÐlÐlÐikÐlÐlˆŒà(ˆÔÝ”�eœk¨$¬.Ð?Ð?¸BÐ?Ð?Ñ@Ô@ˆŒÝ”7‘9”9ˆŒÝ”�eœj¨°6ÀÐGÑGÔGÑHÔHˆŒØ×ÒÑÔÐÐÐrF   rG   c                 ó�   — t          j        j        | j        g| j        ¢R Ž  t           j                             | j        ¦  «         d S rP   )r5   r=   Úuniform_r¦   r�   Úones_r©   rK   s    rD   rª   zHSMSSD.reset_parameters4  s>   € Ý
ŒÔ˜œÐ4 $Ô"3Ð4Ð4Ð4Ð4Ý
Œ�Š�d”fÑÔÐÐÐrF   rN   c                 ó$  — |j         \  }}}}|                     d¦  «        }|                      |                      |¦  «                             |d||¦  «        ¦  «                             d¦  «        }t          j        || j        | j        | j        gd¬¦  «        \  }}}	|	| j         	                    ddd¦  «        z    
                    d¦  «        }
|
|z  }||                     dd¦  «        z  }t          j        |                      |¦  «        | j        | j        gd¬¦  «        \  }}|                      ||                      |¦  «        z  || j        z  z   ¦  «        }||z  }|                     |d||¦  «                             ¦   «         }||fS )Nrp   r~   r   ©rm   éþÿÿÿ)ÚshapeÚflattenr¡   r    r�   r^   Úsplitrœ   r¦   ÚviewÚsoftmaxÚ	transposer¢   rŸ   r£   r8   r©   Ú
contiguous)rA   rN   ÚbatchÚ_ÚheightÚwidthÚBCdtÚbÚcÚdtÚaÚabÚhÚzÚys                  rD   rS   zHSMSSD.forward8  sb  € Ø"#¤'Ñˆˆq�&˜%Ø�IŠI�a‰LŒLˆà�wŠw�t—~’~ aÑ(Ô(×0Ò0°¸¸FÀEÑJÔJÑKÔK×SÒSÐTUÑVÔVˆÝ”;˜t d¤n°d´nÀdÄnÐ%UÐ[\Ð]Ñ]Ô]‰ˆˆ1ˆbØ�$”&—+’+˜a  QÑ'Ô'Ñ'×0Ò0°Ñ4Ô4ˆà�‰UˆØ�—’˜R Ñ$Ô$Ñ$ˆåŒ{˜4Ÿ<š<¨™?œ?¨T¬\¸4¼<Ð,HÈaÐPÑPÔP‰ˆˆ1Ø�MŠM˜!˜dŸhšh q™kœk™/¨A°´©JÑ6Ñ7Ô7ˆØ�‰Eˆà�IŠI�e˜R ¨Ñ/Ô/×:Ò:Ñ<Ô<ˆØ�!ˆtˆrF   )r"   r™   rš   NN©rG   N)rU   rV   rW   rX   r[   r]   r	   r4   rª   r^   r`   rS   ra   rb   s   @rD   r˜   r˜     sÓ   ø€ € € € € ðð ð !$ØØ0;ØØð$ ð $ àð$ ð ð$ ð ð	$ ð
    u Ô-ð$ ð $ ð $ ð $ ð $ ð $ ðLð ð ð ð˜œð ¨%°´¸e¼lÐ0JÔ*Kð ð ð ð ð ð ð ð rF   r˜   c                   óø   ‡ — e Zd ZU dZej        j        e         ed<   	 	 	 	 	 	 dde	d	e
d
e
de	de
f
ˆ fd„Zdd„Z ej        ¦   «         dej        fd„¦   «         Zdej        deej        ej        f         fd„Zˆ xZS )ÚEfficientViMBlockzZDepthwise conv -> HSM-SSD mixer -> depthwise conv -> FFN, each with a LayerScale residual.rˆ   r‰   r"   r™   ç        Nrm   Ú	mlp_ratior›   rœ   Ú	drop_pathc           	      ó  •— t          ¦   «                              ¦   «          ||dœ}d| _        t          |f||dœ|¤Ž| _        t          |fddi|¤Ž| _        t          ||dfd|dd d	œ|¤Ž| _        t          ||dfd|dd d	œ|¤Ž| _	        t          |t          ||z  ¦  «        fi |¤Ž| _        t          j        d
t          j        d|||¬¦  «        z  d¬¦  «        | _        |dk    rt%          |¦  «        nt          j        ¦   «         | _        d S )Nr/   F)r›   rœ   Úepsçñhãˆµøä>r!   r   r   )r'   r)   r,   r+   ç-Cëâ6?rq   T©Úrequires_gradrÉ   )r3   r4   rˆ   r˜   Úmixerr   r*   re   rŒ   r�   rw   r[   Úffnr5   r¤   r^   r¨   Úalphar   ÚIdentityrË   )
rA   rm   rÊ   r›   rœ   rË   r0   r1   rB   rC   s
            €rD   r4   zEfficientViMBlock.__init__P  s@  ø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆØˆŒ
Ý˜CÐR¨JÀ)ÐRÐRÈrÐRÐRˆŒ
Ý Ð4Ð4¨Ð4°Ð4Ð4ˆŒ	å" 3¨¨QÐn¸À#ÐVWÐcgÐnÐnÐkmÐnÐnˆŒÝ" 3¨¨QÐn¸À#ÐVWÐcgÐnÐnÐkmÐnÐnˆŒÝ�s�C  i¡Ñ0Ô0Ð7Ð7°BÐ7Ð7ˆŒõ
 ”\Ø•5”:˜a ¨V¸5ÐAÑAÔAÑAØð
ñ 
ô 
ˆŒ
ð 1:¸C²°� )Ñ,Ô,Ð,ÅRÄ[Á]Ä]ˆŒˆˆrF   rG   c                 óP   — t           j                             | j        d¦  «         d S )NrÏ   )r5   r=   r>   rÔ   rK   s    rD   rª   z"EfficientViMBlock.reset_parametersm  s"   € Ý
Œ×Ò˜$œ* dÑ+Ô+Ð+Ð+Ð+rF   c                 ó~  — | j         sµ| j        r
J d¦   «         ‚| j                             ¦   «                              ¦   «         }| j                             ¦   «          | j                             ¦   «          t          | j        j	        |d         ¦  «         t          | j        j	        |d         ¦  «         d| _         | S )z:Fold the gated depthwise residual branches for evaluation.r�   r   rp   T)
rˆ   r‘   rÔ   ÚdetachÚsigmoidrŒ   rL   r�   r…   r7   )rA   rÔ   s     rD   rL   zEfficientViMBlock.fusep  sª   € ð Œzð 	Ø”}ÐHÐHÐ&HÑHÔHÐ$Ø”J×%Ò%Ñ'Ô'×/Ò/Ñ1Ô1ˆEØŒL×ÒÑÔÐØŒL×ÒÑÔÐÝ$ T¤\Ô%6¸¸a¼ÑAÔAÐAÝ$ T¤\Ô%6¸¸a¼ÑAÔAÐAØˆDŒJØˆrF   rN   c                 óÄ  — t          j        | j        ¦  «                             dddd¦  «        }| j        r|                      |¦  «        n6||                      |d         |                      |¦  «        |z
  z  ¦  «        z   }|}|                      |                      |¦  «        ¦  «        \  }}||                      |d         ||z
  z  ¦  «        z   }| j        r|  	                    |¦  «        n6||                      |d         |  	                    |¦  «        |z
  z  ¦  «        z   }||                      |d         |  
                    |¦  «        |z
  z  ¦  «        z   }||fS )Nrq   r~   r   r   rp   r!   )r^   rÙ   rÔ   rµ   rˆ   rŒ   rË   rÒ   r*   r�   rÓ   )rA   rN   rÔ   Úx_prevrÃ   s        rD   rS   zEfficientViMBlock.forward}  s6  € Ý”˜dœjÑ)Ô)×.Ò.¨q°"°a¸Ñ;Ô;ˆð  $œzÐcˆD�LŠL˜‰OŒOˆO¨q°4·>²>À%ÈÄ(ÈdÏlÊlÐ[\ÉoÌoÐ`aÑNaÑBbÑ3cÔ3cÑ/cˆð ˆØ�zŠz˜$Ÿ)š) A™,œ,Ñ'Ô'‰ˆˆ1Ø�T—^’^ E¨!¤H°°F±
Ñ$;Ñ<Ô<Ñ<ˆð  $œzÐcˆD�LŠL˜‰OŒOˆO¨q°4·>²>À%ÈÄ(ÈdÏlÊlÐ[\ÉoÌoÐ`aÑNaÑBbÑ3cÔ3cÑ/cˆð �—’˜u Qœx¨4¯8ª8°A©;¬;¸©?Ñ;Ñ<Ô<Ñ<ˆØ�!ˆtˆrF   )r‰   r"   r™   rÉ   NNrÆ   )rU   rV   rW   rX   r^   r“   r”   r•   r–   r[   r]   r4   rª   r_   r5   r\   rL   r`   r	   rS   ra   rb   s   @rD   rÈ   rÈ   K  s&  ø€ € € € € € ØdÐdàŒ9Œ?˜4Ô Ð Ð Ñ ð
  #Ø #ØØ"ØØðSð SàðSð ðSð ð	Sð
 ðSð ðSð Sð Sð Sð Sð Sð:,ð ,ð ,ð ,ð €U„]�_„_ð
�b”ið 
ð 
ð 
ñ „_ð
ð˜œð ¨%°´¸e¼lÐ0JÔ*Kð ð ð ð ð ð ð ð rF   rÈ   c                   óÔ   ‡ — e Zd Z	 	 	 	 	 	 	 ddedededed	eeej                          d
ededee	e                  fˆ fd„Z
dej        deej        ej        ej        f         fd„Zˆ xZS )ÚEfficientViMStager‰   Nr"   r™   r#   r$   ÚdepthrÊ   Ú
downsampler›   rœ   rË   c           	      ó2  •‡‡‡‡‡‡	‡
— t          ¦   «                              ¦   «          |dk    sJ ‚|| _        d| _        ‰pdg|z  Št	          j        ˆ	ˆˆ
ˆˆˆˆfd„t          |¦  «        D ¦   «         ¦  «        | _        |� |‰|‰	‰
¬¦  «        nd | _        d S )Nr   FrÉ   c                 óH   •— g | ]}t          ‰‰‰‰‰|         ‰‰¬ ¦  «        ‘ŒS ))rm   rÊ   r›   rœ   rË   r0   r1   )rÈ   )	Ú.0Úir0   rË   r1   r#   rÊ   r›   rœ   s	     €€€€€€€rD   ú
<listcomp>z.EfficientViMStage.__init__.<locals>.<listcomp>£  sS   ø€ ð %
ð %
ð %
ð õ ØØ#Ø%Ø#Ø# Aœ,ØØðñ ô ð%
ð %
ð %
rF   )r#   r$   r0   r1   )	r3   r4   rÞ   Úgrad_checkpointingr5   Ú
ModuleListÚrangeÚblocksrß   )rA   r#   r$   rÞ   rÊ   rß   r›   rœ   rË   r0   r1   rC   s    `  ` `````€rD   r4   zEfficientViMStage.__init__‘  sÝ   øøøøøøøø€ õ 	‰Œ×ÒÑÔÐØ�qŠyˆyˆyˆyØˆŒ
Ø"'ˆÔØÐ. # ¨¡ˆ	Ý”mð %
ð %
ð %
ð %
ð %
ð %
ð %
ð %
ð %
ð %
õ ˜5‘\”\ð%
ñ %
ô %
ñ ô ˆŒð WaÐVlˆJˆJ˜f¨g¸fÈEÐRÑRÔRÐRÐrvð 	ŒˆˆrF   rN   rG   c                 óî   — d}| j         D ]I}| j        r2t          j                             ¦   «         st          ||¦  «        \  }}Œ; ||¦  «        \  }}ŒJ|}| j        �|                      |¦  «        }|||fS )a˜  Returns (feature_map, next_stage_input, hidden_state).

        `next_stage_input` is the input to the next stage (post-downsample when present), `feature_map`
        is this stage's output at its own resolution -- the tensor `feature_info` points at.
        The feature map is first so that feature-extraction wrappers, which take `x[0]` of a
        tuple-returning module, tap the right tensor.
        N)rè   rå   r^   r“   Úis_scriptingr   rß   )rA   rN   rÃ   ÚblkÚx_outs        rD   rS   zEfficientViMStage.forward³  sŒ   € ð ˆØ”;ð 	ð 	ˆCØÔ&ð ­u¬y×/EÒ/EÑ/GÔ/Gð Ý! # qÑ)Ô)‘��1�1à�s˜1‘v”v‘��1�1àˆØŒ?Ð&Ø—’ Ñ"Ô"ˆAØ�a˜ˆ{ÐrF   )r‰   Nr"   r™   NNN)rU   rV   rW   r[   r]   r   r
   r5   r\   r   r4   r^   r`   r	   rS   ra   rb   s   @rD   rÝ   rÝ   �  sò   ø€ € € € € ð  #Ø48Ø #ØØ/3ØØð 
ð  
àð 
ð ð 
ð ð	 
ð
 ð 
ð !  b¤i¤Ô1ð 
ð ð 
ð ð 
ð    U¤Ô,ð 
ð  
ð  
ð  
ð  
ð  
ðD˜œð ¨%°´¸e¼lÈEÌLÐ0XÔ*Yð ð ð ð ð ð ð ð rF   rÝ   c                   ó  ‡ — e Zd ZdZ	 	 	 	 	 	 	 	 	 	 	 	 d8dededeedf         deedf         dededeedf         dededefˆ fd„Zd9dede	e
eef                  dej        fd„Zej        j        d:deddfd„¦   «         Zd:dej        deddfd„Zej        j        dee         fd „¦   «         Zej        j        d;d"ede
eef         fd#„¦   «         Zej        j        d:d$efd%„¦   «         Zej        j        dej        fd&„¦   «         Zd9dede	e         fd'„Z	 	 	 	 	 d<d)ej        d*e	eeee         f                  d+ed,ed-ed.edeeej                 eeej                 eej                 f         f         fd/„Z	 	 	 d=d*eeee         f         d1ed2edee         fd3„Z d)ej        deej                 fd4„Z!	 d;d)eej                 d5edeej        eej                 f         fd6„Z"d)ej        deej        eej                 f         fd7„Z#ˆ xZ$S )>r   a  EfficientViM backbone with multi-stage hidden-state fused classification head.

    Three HSM-SSD stages operate at reductions 16 / 32 / 64. Each stage yields both a 2-D
    feature map (`x_out`, exposed via `feature_info` / `forward_intermediates`) and the
    final hidden state `h` of its last block, which the classification head fuses together
    with the pooled final feature map under a learned softmax weighting.

    Args:
        in_chans: Number of input image channels.
        num_classes: Number of classifier output classes.
        embed_dim: Channel width of each of the three stages.
        depths: Number of blocks in each of the three stages.
        mlp_ratio: FFN expansion ratio.
        ssd_expand: Expansion of the HSM-SSD hidden state.
        state_dim: Per-stage HSM-SSD state dimension.
        global_pool: Global pooling type, 'avg' or '' to return unpooled branch features.
        drop_rate: Classifier dropout rate.
        drop_path_rate: Stochastic depth rate, linearly ramped across all blocks.
    r!   éè  ©é€   é   é   ©rp   rp   rp   r‰   r"   ©é1   é   é	   ÚavgrÉ   NÚin_chansÚnum_classesÚ	embed_dim.ÚdepthsrÊ   r›   rœ   Úglobal_poolÚ	drop_rateÚdrop_path_ratec                 óÔ  •‡— t          ¦   «                              ¦   «          ||dœŠ|dv sJ ‚|s|dk    sJ ‚t          |¦  «        t          |¦  «        cxk    rt          |¦  «        k    sn J ‚|| _        || _        t          |¦  «        | _        t          |¦  «        | _        d| _        g | _	        t          d||d         dœ‰¤Ž| _        t          |
t          |¦  «        ¦  «        }g }d}d}t          | j        ¦  «        D ]è}|||||         z   …         }|||         z  }|                     t!          dt#          ||         ¦  «        || j        dz
  k     rt#          ||dz            ¦  «        nd||         ||| j        dz
  k     rt$          nd |||         |dœ‰¤Ž¦  «         | j	                             t'          ||         |d	|› �¬
¦  «        ¦  «         || j        dz
  k     r|dz  }Œét)          j        |¦  «        | _        t          |¦  «        |d         z   x| _        | _        t)          j        t5          j        | j        dz   fi ‰¤Ž|dk    ¬¦  «        | _        t)          j        g ˆfd„|D ¦   «         ¢t;          |d         fddi‰¤Ž‘¦  «        | _        t)          j        |	¦  «        | _         |  !                    |‰¦  «        | _"        |  #                    d¬¦  «         d S )Nr/   ©rø   Ú r   F)r#   rm   é   r   )r#   r$   rÞ   rÊ   rß   r›   rœ   rË   zstages.)Únum_chsÚ	reductionÚmodulerp   r~   rÐ   c                 ó.   •— g | ]}t          |fd di‰¤Ž‘ŒS )rÍ   rÎ   )r   )râ   rm   rB   s     €rD   rä   z)EfficientViM.__init__.<locals>.<listcomp>  s.   ø€ ÐBÐBÐB°�i˜Ð,Ð, Ð,¨Ð,Ð,ÐBÐBÐBrF   rÍ   rÎ   ©Úneeds_reset© )$r3   r4   Úlenrý   rú   Ú
num_layersÚtuplerû   rå   Úfeature_inforl   Úpatch_embedr   Úsumrç   ÚappendrÝ   r[   r‡   Údictr5   ræ   ÚstagesÚnum_featuresÚhead_hidden_sizer¤   r^   r¨   Úweightsr   r*   ÚDropoutÚ	head_dropÚ_build_headsÚheadsÚinit_weights)rA   rù   rú   rû   rü   rÊ   r›   rœ   rý   rþ   rÿ   r0   r1   Údprr  Ú	block_idxr  Úi_layerÚ	stage_dprrB   rC   s                      @€rD   r4   zEfficientViM.__init__Ý  sK  øø€ õ 	‰Œ×ÒÑÔÐØ¨Ð/Ð/ˆØ˜kÐ)Ð)Ð)Ð)ØÐ.˜k¨QÒ.Ð.Ð.Ð.Ý�6‰{Œ{�c )™nœnÐ>Ð>Ò>Ð>µ°I±´Ò>Ð>Ð>Ð>Ð>Ð>Ø&ˆÔØ&ˆÔÝ˜f™+œ+ˆŒÝ˜yÑ)Ô)ˆŒØ"'ˆÔØˆÔåÐH x°Y¸q´\ÐHÐHÀRÐHÐHˆÔÝ'¨½¸F¹¼ÑDÔDˆàˆØˆ	ð ˆ	Ý˜Tœ_Ñ-Ô-ð 	ð 	ˆGØ˜I¨	°F¸7´OÑ(CÐCÔDˆIØ˜ œÑ(ˆIØ�MŠMÝ!ð 
Ý˜y¨Ô1Ñ2Ô2Ø<CÀdÄoÐXYÑFYÒ<YÐ<Y�C 	¨'°A©+Ô 6Ñ7Ô7Ð7Ð`aØ  œ/Ø'Ø07¸$¼/ÈAÑ:MÒ0MÐ0M�|˜|ÐTXØ)Ø'¨Ô0Ø'ð
ð 
ð ð
ð 
ñô ð ð Ô×$Ò$Ý˜Y wÔ/¸9ÐM`ÐW^ÐM`ÐM`ÐaÑaÔañô ð ð ˜œ¨1Ñ,Ò,Ð,Ø˜Q‘�	øÝ”m FÑ+Ô+ˆŒå47¸	±N´NÀYÈrÄ]Ñ4RÐRˆÔ˜DÔ1Ý”|¥E¤J¨t¬ÀÑ/BÐ$IÐ$IÀbÐ$IÐ$IÐYdÐghÒYhÐiÑiÔiˆŒÝ”Mð #
ØBÐBÐBÐB¸	ÐBÑBÔBð#
å˜	 "œÐ6Ð6¨4Ð6°2Ð6Ð6ð#
ñ ô ˆŒ	õ œ IÑ.Ô.ˆŒØ×&Ò& {°BÑ7Ô7ˆŒ
Ø×Ò eÐÑ,Ô,Ð,Ð,Ð,rF   rB   rG   c                 óz   ‡‡— ‰pi Št          j        ˆˆfd„g | j        ¢| j        d         ‘R D ¦   «         ¦  «        S )Nc                 óh   •— g | ].}‰d k    rt          j        |‰fi ‰¤Žnt          j        ¦   «         ‘Œ/S )r   )r5   ÚLinearrÕ   )râ   rm   rB   rú   s     €€rD   rä   z-EfficientViM._build_heads.<locals>.<listcomp>#  sP   ø€ ð 
ð 
ð 
àð 2=¸q²°�BŒI�c˜;Ð-Ð-¨"Ð-Ð-Ð-ÅbÄkÁmÄmð
ð 
ð 
rF   r~   )r5   ræ   rû   )rA   rú   rB   s    ``rD   r  zEfficientViM._build_heads!  sb   øø€ ØˆX�2ˆÝŒ}ð 
ð 
ð 
ð 
ð 
à<˜œÐ<¨¬¸Ô);Ð<Ð<ð
ñ 
ô 
ñ ô ð 	rF   Tr	  c                 ó  — t           j                             | j        ¦  «         t	          | d¦  «        r$t           j                             | j        ¦  «         |                      t          | j        |¬¦  «        ¦  «         dS )z?Initialize parameters and restore buffers after ``to_empty()``.Úweights_distr  N)	r5   r=   r®   r  Úhasattrr$  Úapplyr   Ú_init_weights)rA   r	  s     rD   r  zEfficientViM.init_weights(  sk   € õ 	Œ�Š�d”lÑ#Ô#Ð#Ý�4˜Ñ(Ô(ð 	-ÝŒG�MŠM˜$Ô+Ñ,Ô,Ð,Ø�
Š
•7˜4Ô-¸;ÐGÑGÔGÑHÔHÐHÐHÐHrF   Úmc                 ó¦  — t          |t          j        ¦  «        rEt          |j        d¬¦  «         |j        �&t          j                             |j        ¦  «         d S d S t          |t          j        t          j	        t          j
        f¦  «        r|                     ¦   «          d S |r&t          |d¦  «        r|                     ¦   «          d S d S d S )Ng{®Gáz”?)Ústdrª   )r9   r5   r"  r   r?   r2   r=   Úzeros_r   rj   rY   rª   r%  )rA   r(  r	  s      rD   r'  zEfficientViM._init_weights0  sÔ   € Ý�a�œÑ#Ô#ð 	!Ý˜!œ(¨Ð-Ñ-Ô-Ð-ØŒvÐ!Ý”—’˜qœvÑ&Ô&Ð&Ð&Ð&ð "Ð!å˜�BœL­"¬.½"¼.ÐIÑJÔJð 	!Ø×ÒÑ Ô Ð Ð Ð Øð 	!�W QÐ(:Ñ;Ô;ð 	!Ø×ÒÑ Ô Ð Ð Ð ð	!ð 	!ð 	!ð 	!rF   c                 ó>   — d„ |                       ¦   «         D ¦   «         S )Nc                 óB   — h | ]\  }}|                      d ¦  «        ¯|’ŒS )z.D)Úendswith)râ   Únamerº   s      rD   ú	<setcomp>z/EfficientViM.no_weight_decay.<locals>.<setcomp><  s-   € ÐSÐSÐS™˜˜q¸t¿}º}ÈTÑ?RÔ?RÐS�ÐSÐSÐSrF   )Únamed_parametersrK   s    rD   Úno_weight_decayzEfficientViM.no_weight_decay:  s"   € àSÐS D×$9Ò$9Ñ$;Ô$;ÐSÑSÔSÐSrF   FÚcoarsec                 ó0   — t          d|rdng d¢¬¦  «        S )Nz^patch_embedz^stages\.(\d+)))z^stages\.(\d+)\.blocks\.(\d+)N)z^stages\.(\d+)\.downsample©iŸ† )z^normr5  )Ústemrè   )r  )rA   r3  s     rD   Úgroup_matcherzEfficientViM.group_matcher>  s9   € åØ àðÐ$Ð$ðð ð ð		
ñ 	
ô 	
ð 		
rF   Úenablec                 ó6   — || _         | j        D ]	}||_         Œ
d S rP   )rå   r  )rA   r8  Ústages      rD   Úset_grad_checkpointingz#EfficientViM.set_grad_checkpointingK  s0   € à"(ˆÔØ”[ð 	.ð 	.ˆEØ'-ˆEÔ$Ð$ð	.ð 	.rF   c                 ó   — | j         S rP   )r  rK   s    rD   Úget_classifierzEfficientViM.get_classifierQ  s
   € àŒzÐrF   c                 ó"  — t          | ¦  «        }| j        }|| _        |�|dv sJ ‚|| _        | j        s|dk    sJ ‚|                      ||¦  «        | _        | j                             |¦  «         | j                             |dk    ¦  «         d S )Nr  r   )	r   r‘   rú   rý   r  r  Útrainr  Úrequires_grad_)rA   rú   rý   rB   Úwas_trainings        rD   Úreset_classifierzEfficientViM.reset_classifierU  s£   € Ý˜dÑ#Ô#ˆØ”}ˆØ&ˆÔØÐ"Ø +Ð-Ð-Ð-Ð-Ø*ˆDÔØÔÐ3 ;°!Ò#3Ð#3Ð#3Ð3Ø×&Ò& {°BÑ7Ô7ˆŒ
ØŒ
×Ò˜Ñ&Ô&Ð&ØŒ×#Ò# K°!¢OÑ4Ô4Ð4Ð4Ð4rF   ÚNCHWrN   Úindicesr*   Ú
stop_earlyÚ
output_fmtÚintermediates_onlyc                 ó¾  — |dv s
J d¦   «         ‚g }t          t          | j        ¦  «        |¦  «        \  }}	|                      |¦  «        }g }
t          j                             ¦   «         s|s| j        }n| j        d|	dz   …         }t          t          || j	        ¦  «        ¦  «        D ]±\  }\  }} ||¦  «        \  }}}||v rN| 
                    |r7 ||                     dddd¦  «        ¦  «                             dddd¦  «        n|¦  «         |sF|
 
                     ||                     dd¦  «        ¦  «                             dd¦  «        ¦  «         Œ²|r|S t          j                             ¦   «         s|r|	t          | j        ¦  «        dz
  k    r | j	        d         |¦  «        }n|}|
 
                    |                     d¦  «        ¦  «         |
|fS )	a   Forward features that returns intermediates.

        Args:
            x: Input image tensor.
            indices: Take last n blocks if int, all if None, select matching indices if sequence.
            norm: Apply norm layer to compatible intermediates (applies the per-stage LayerNorm
                to the stage's own output map only).
            stop_early: Stop iterating over blocks when last desired intermediate hit.
            output_fmt: Shape of intermediate feature outputs.
            intermediates_only: Only return intermediate features.

        Returns:
            List of intermediate features or tuple of (final features, intermediates).
        )rC  zOutput shape must be NCHW.Nr   r   rp   r!   r~   )r   r  r  r  r^   r“   rê   Ú	enumerateÚzipr*   r  Úpermuter·   r³   )rA   rN   rD  r*   rE  rF  rG  ÚintermediatesÚtake_indicesÚ	max_indexÚfeaturesr  Úfeat_idxr:  Ú
stage_normrì   rÃ   s                    rD   Úforward_intermediatesz"EfficientViM.forward_intermediatesa  s   € ð. ˜YÐ&Ð&Ð&Ð(DÑ&Ô&Ð&ØˆÝ"6µs¸4¼;Ñ7GÔ7GÈÑ"QÔ"QÑˆ�ià×Ò˜QÑÔˆØˆÝŒ9×!Ò!Ñ#Ô#ð 	1¨:ð 	1Ø”[ˆFˆFà”[  )¨a¡- Ô0ˆFÝ-6µs¸6À4Ä9Ñ7MÔ7MÑ-NÔ-Nð 	Oð 	OÑ)ˆHÑ)�u˜jØ˜% ™(œ(‰KˆE�1�aØ˜<Ð'Ð'Ø×$Ò$ÐbfÐ%q Z Z°·²¸aÀÀAÀqÑ0IÔ0IÑ%JÔ%J×%RÒ%RÐSTÐVWÐYZÐ\]Ñ%^Ô%^Ð%^ÐlqÑrÔrÐrØ%ð OØ—’ 
 
¨1¯;ª;°q¸!Ñ+<Ô+<Ñ =Ô =× GÒ GÈÈ1Ñ MÔ MÑNÔNÐNøàð 	!Ø Ð ÝŒ9×!Ò!Ñ#Ô#ð 	¨:ð 	¸ÅcÈ$Ì+ÑFVÔFVÐYZÑFZÒ9ZÐ9ZØ�”	˜"”˜aÑ Ô ˆAˆAàˆAØ�Š˜Ÿ	š	 !™œÑ%Ô%Ð%Ø˜Ð&Ð&rF   r   Ú
prune_normÚ
prune_headc                 ó‚  ‡— t          t          | j        ¦  «        |¦  «        \  Š}|dz   }|| j        k     �rŒ|s
J d¦   «         ‚| j        d|…         | _        | j        d|…         | _        d| j        d         _        t          j        g | j        d|…         ¢t          j	        ¦   «         ‘¦  «        | _        t          j
        t          j        | j        d|…         | j        dd…         g¦  «                             ¦   «         ¦  «        | _        t          | d¦  «        rYt          j
        t          j        | j        d|…         | j        dd…         g¦  «                             ¦   «         ¦  «        | _        | j        d|…         | _        || _        t%          | j        ¦  «        | j        d         z   x| _        | _        |r)t          j        d„ | j        D ¦   «         ¦  «        | _        n9|r7t          j        ˆfd„t+          | j        ¦  «        D ¦   «         ¦  «        | _        |r|                      d¦  «         ‰S )	z¤Prune stages not required for the specified intermediate features.

        With prune_head, also remove output norms unused by the selected intermediates.
        r   z/The multi-stage classifier requires all stages.Nr~   r$  c                 ó4   — g | ]}t          j        ¦   «         ‘ŒS r
  ©r5   rÕ   )râ   rº   s     rD   rä   z:EfficientViM.prune_intermediate_layers.<locals>.<listcomp>­  s   € Ð&HÐ&HÐ&H¸¥r¤{¡}¤}Ð&HÐ&HÐ&HrF   c                 óH   •— g | ]\  }}|‰v r|nt          j        ¦   «         ‘ŒS r
  rW  )râ   rã   r*   rM  s      €rD   rä   z:EfficientViM.prune_intermediate_layers.<locals>.<listcomp>±  s>   ø€ ð 'ð 'ð 'ÙAHÀÀD˜˜\Ð)Ð)��­r¬{©}¬}ð'ð 'ð 'rF   r   )r   r  r  r  r  rß   r5   ræ   r*   rÕ   r¤   r^   Úcatr  rØ   r%  r$  rû   r  r  r  rI  rB  )rA   rD  rS  rT  rN  r  rM  s         @rD   Úprune_intermediate_layersz&EfficientViM.prune_intermediate_layers’  s3  ø€ õ #7µs¸4¼;Ñ7GÔ7GÈÑ"QÔ"QÑˆ�iØ ‘]ˆ
Ø˜œÒ'Ñ'ØÐPÐPÐPÑPÔP�:Øœ+ k z kÔ2ˆDŒKØ $Ô 1°+°:°+Ô >ˆDÔØ)-ˆDŒK˜ŒOÔ&ÝœÐ&N¨¬	°+°:°+Ô(>Ð&NÅÄÁÄÐ&NÑOÔOˆDŒIÝœ<­¬	°4´<ÀÀÀÔ3LÈdÌlÐ[]Ð[^Ð[^ÔN_Ð2`Ñ(aÔ(a×(hÒ(hÑ(jÔ(jÑkÔkˆDŒLÝ�t˜^Ñ,Ô,ð Ý$&¤LÝ”I˜tÔ0°°*°Ô=¸tÔ?PÐQSÐQTÐQTÔ?UÐVÑWÔW×^Ò^Ñ`Ô`ñ%ô %�Ô!ð "œ^¨K¨Z¨KÔ8ˆDŒNØ(ˆDŒOÝ8;¸D¼NÑ8KÔ8KÈdÌnÐ]_ÔN`Ñ8`Ð`ˆDÔ Ô 5Øð 	ÝœÐ&HÐ&H¸d¼iÐ&HÑ&HÔ&HÑIÔIˆDŒIˆIØð 	õ œð 'ð 'ð 'ð 'ÝLUÐVZÔV_ÑL`ÔL`ð'ñ 'ô 'ñ ô ˆDŒIð ð 	%Ø×!Ò! !Ñ$Ô$Ð$ØÐrF   c           	      ó–  — |                       |¦  «        }g }t          | j        | j        ¦  «        D ]Z\  }} ||¦  «        \  }}}|                      ||                     dd¦  «        ¦  «                             dd¦  «        ¦  «         Œ[|                      | j        d         |¦  «                             d¦  «        ¦  «         |S )zKReturn normalized, unpooled hidden states and the final map as NCL tensors.r   rp   r~   )r  rJ  r  r*   r  r·   r³   )rA   rN   rO  r:  r*   rº   rÃ   s          rD   Úforward_featureszEfficientViM.forward_features¸  sÂ   € à×Ò˜QÑÔˆØˆÝ˜tœ{¨D¬IÑ6Ô6ð 	Eð 	E‰KˆE�4Ø�e˜A‘h”h‰GˆAˆq�!Ø�OŠO˜D˜D §¢¨Q°Ñ!2Ô!2Ñ3Ô3×=Ò=¸aÀÑCÔCÑDÔDÐDÐDØ�Š˜˜œ	 "œ aÑ(Ô(×0Ò0°Ñ3Ô3Ñ4Ô4Ð4ØˆrF   Ú
pre_logitsc                 ó@  ‡ ‡‡— ‰ j         s‰S ˆ fd„‰D ¦   «         Š|s‰ j        dk    rt          j        ‰d¬¦  «        S ‰ j                             d¦  «        Šˆˆfd„t          ‰ j        ¦  «        D ¦   «         }t          j        |¦  «         	                    d¦  «        S )Nc                 ó`   •— g | ]*}‰                      |                     d ¦  «        ¦  «        ‘Œ+S ©r~   ©r  Úmean©râ   ÚtrA   s     €rD   rä   z-EfficientViM.forward_head.<locals>.<listcomp>É  ó/   ø€ Ð3Ð3Ð3¨AˆT�^Š^˜AŸFšF 2™JœJÑ'Ô'Ð3Ð3Ð3rF   r   r   r°   r~   c                 óJ   •— g | ]\  }}‰|          |‰|         ¦  «        z  ‘Œ S r
  r
  ©râ   rã   Úheadr  rN   s      €€rD   rä   z-EfficientViM.forward_head.<locals>.<listcomp>Í  ó2   ø€ ÐOÐOÐO©g¨a°�'˜!”*˜t˜t A a¤D™zœzÑ)ÐOÐOÐOrF   )
rý   rú   r^   rY  r  r¶   rI  r  Ústackr  )rA   rN   r]  Úlogitsr  s   ``  @rD   Úforward_headzEfficientViM.forward_headÂ  s³   øøø€ ð
 Ôð 	ØˆHØ3Ð3Ð3Ð3°Ð3Ñ3Ô3ˆØð 	'˜Ô)¨QÒ.Ð.Ý”9˜Q AÐ&Ñ&Ô&Ð&Ø”,×&Ò& rÑ*Ô*ˆØOÐOÐOÐOÐO½À4Ä:Ñ9NÔ9NÐOÑOÔOˆÝŒ{˜6Ñ"Ô"×&Ò& qÑ)Ô)Ð)rF   c                 óV   — |                       |¦  «        }|                      |¦  «        S rP   )r\  rl  rR   s     rD   rS   zEfficientViM.forwardÐ  s)   € Ø×!Ò! !Ñ$Ô$ˆØ× Ò  Ñ#Ô#Ð#rF   )r!   rî   rï   ró   r‰   r"   rô   rø   rÉ   rÉ   NNrP   ©T©F)NFFrC  F)r   FT)%rU   rV   rW   rX   r[   r	   r]   Ústrr4   r   r   r   r5   ræ   r  r^   r“   Úignorer•   r  r\   r'  r   r2  r7  r;  r=  rB  r`   r   r   rR  rZ  r\  rl  rS   ra   rb   s   @rD   r   r   È  sn  ø€ € € € € ðð ð, Ø#Ø)8Ø&/Ø"Ø #Ø)4Ø$Ø"Ø$'ØØðB-ð B-àðB-ð ðB-ð ˜S #˜X”ð	B-ð
 ˜#˜s˜(”OðB-ð ðB-ð ðB-ð ˜S #˜X”ðB-ð ðB-ð ðB-ð "ðB-ð B-ð B-ð B-ð B-ð B-ðHð ¨ð °¸$¸sÀC¸x¼.Ô1Ið ÐUWÔUbð ð ð ð ð „YÔðIð I¨ð I¸ð Ið Ið Iñ ÔðIð!ð !˜rœyð !°tð !Àtð !ð !ð !ð !ð „YÔðT  S¤ð Tð Tð Tñ ÔðTð „YÔð

ð 

 Dð 

°T¸#¸s¸(´^ð 

ð 

ð 

ñ Ôð

ð „YÔð.ð .¨Tð .ð .ð .ñ Ôð.ð
 „YÔð ¤ð ð ð ñ Ôðð
5ð 
5¨Cð 
5¸hÀs¼mð 
5ð 
5ð 
5ð 
5ð 8<ØØ$Ø$Ø',ð/'ð /'àŒ|ð/'ð ˜e C¨¨c¬ NÔ3Ô4ð/'ð ð	/'ð
 ð/'ð ð/'ð !%ð/'ð 
ˆt�E”LÔ! 5¨¨e¬lÔ);¸TÀ%Ä,Ô=OÐ)OÔ#PÐPÔ	Qð/'ð /'ð /'ð /'ðf ./Ø$Ø#ð	$ð $à˜3  S¤	˜>Ô*ð$ð ð$ð ð	$ð
 
ˆcŒð$ð $ð $ð $ðL %¤,ð °4¸¼Ô3Eð ð ð ð ð  %ð*ð *à�E”LÔ!ð*ð ð*ð 
ˆuŒ|˜T %¤,Ô/Ð/Ô	0ð	*ð *ð *ð *ð$˜œð $¨%°´¸dÀ5Ä<Ô>PÐ0PÔ*Qð $ð $ð $ð $ð $ð $ð $ð $rF   c            
       ó2  ‡ — e Zd ZdZdedefˆ fd„Zej        j        de	e
j        e
j        f         fd„¦   «         Zdded	ee         fˆ fd
„Zej        j        ddefd„¦   «         Z	 ddeej                 dedeej        eej                 f         fd„Zˆ xZS )r   a9  EfficientViM with a parallel distilled head.

    Adds `heads_dist` / `weights_dist`, fused the same way as the primary head. Training with
    `distilled_training=True` returns the `(logits, logits_dist)` tuple, otherwise (and always
    in eval) the two predictions are averaged, matching `LevitDistilled`.
    ÚargsÚkwargsc                 óV  •—  t          ¦   «         j        |i |¤Ž t          | ¦  «        }| j        }t	          j        t          j        | j        dz   fi |¤Ž|dk    ¬¦  «        | _	        |  
                    ||¦  «        | _        | j                             | j        ¦  «         d| _        d S )Nr   r   rÐ   F)r3   r4   r   rú   r5   r¤   r^   r¨   r  r$  r  Ú
heads_distr&  r'  Údistilled_training)rA   rs  rt  rB   rú   rC   s        €rD   r4   zEfficientViMDistilled.__init__Ý  s©   ø€ Ø�‰ŒÔ˜$Ð) &Ð)Ð)Ð)Ý˜dÑ#Ô#ˆØÔ&ˆÝœL­¬°D´OÀaÑ4GÐ)NÐ)NÈ2Ð)NÐ)NÐ^iÐlmÒ^mÐnÑnÔnˆÔØ×+Ò+¨K¸Ñ<Ô<ˆŒØŒ×Ò˜dÔ0Ñ1Ô1Ð1Ø"'ˆÔÐÐrF   rG   c                 ó   — | j         | j        fS rP   )r  rv  rK   s    rD   r=  z$EfficientViMDistilled.get_classifieræ  s   € àŒz˜4œ?Ð*Ð*rF   Nrú   rý   c                 ó  •— t          ¦   «                              ||¬¦  «         t          | ¦  «        }|                      ||¦  «        | _        | j                             | j        ¦  «         | j                             |dk    ¦  «         d S )N)rý   r   )	r3   rB  r   r  rv  r?  r‘   r$  r@  )rA   rú   rý   rB   rC   s       €rD   rB  z&EfficientViMDistilled.reset_classifierê  s}   ø€ Ý‰Œ× Ò  ¸+Ð ÑFÔFÐFÝ˜dÑ#Ô#ˆØ×+Ò+¨K¸Ñ<Ô<ˆŒØŒ×Ò˜dœmÑ,Ô,Ð,ØÔ×(Ò(¨°qªÑ9Ô9Ð9Ð9Ð9rF   Tr8  c                 ó   — || _         d S rP   )rw  )rA   r8  s     rD   Úset_distilled_trainingz,EfficientViMDistilled.set_distilled_trainingñ  s   € à"(ˆÔÐÐrF   FrN   r]  c                 óv  ‡ ‡‡‡— ‰ j         s‰S ˆ fd„‰D ¦   «         Š|s‰ j        dk    rt          j        ‰d¬¦  «        S ‰ j                             d¦  «        Š‰ j                             d¦  «        Šˆˆfd„t          ‰ j        ¦  «        D ¦   «         }ˆˆfd„t          ‰ j	        ¦  «        D ¦   «         }t          j
        |¦  «                             d¦  «        t          j
        |¦  «                             d¦  «        }}‰ j        r)‰ j        r"t          j                             ¦   «         s||fS ||z   dz  S )	Nc                 ó`   •— g | ]*}‰                      |                     d ¦  «        ¦  «        ‘Œ+S r`  ra  rc  s     €rD   rä   z6EfficientViMDistilled.forward_head.<locals>.<listcomp>ü  re  rF   r   r   r°   r~   c                 óJ   •— g | ]\  }}‰|          |‰|         ¦  «        z  ‘Œ S r
  r
  rg  s      €€rD   rä   z6EfficientViMDistilled.forward_head.<locals>.<listcomp>  ri  rF   c                 óJ   •— g | ]\  }}‰|          |‰|         ¦  «        z  ‘Œ S r
  r
  )râ   rã   rh  r$  rN   s      €€rD   rä   z6EfficientViMDistilled.forward_head.<locals>.<listcomp>  s2   ø€ Ð^Ð^Ð^¹¸¸4�| A”¨¨¨a°¬d©¬Ñ3Ð^Ð^Ð^rF   rp   )rý   rú   r^   rY  r  r¶   r$  rI  r  rv  rj  r  rw  r‘   r“   rê   )	rA   rN   r]  rk  Úlogits_distrÄ   Úz_distr  r$  s	   ``     @@rD   rl  z"EfficientViMDistilled.forward_headõ  sM  øøøø€ ð
 Ôð 	ØˆHØ3Ð3Ð3Ð3°Ð3Ñ3Ô3ˆØð 	'˜Ô)¨QÒ.Ð.Ý”9˜Q AÐ&Ñ&Ô&Ð&Ø”,×&Ò& rÑ*Ô*ˆØÔ(×0Ò0°Ñ4Ô4ˆØOÐOÐOÐOÐO½À4Ä:Ñ9NÔ9NÐOÑOÔOˆØ^Ð^Ð^Ð^Ð^Å9ÈTÌ_ÑC]ÔC]Ð^Ñ^Ô^ˆÝ”K Ñ'Ô'×+Ò+¨AÑ.Ô.µ´¸KÑ0HÔ0H×0LÒ0LÈQÑ0OÔ0Oˆ6ˆØÔ"ð 	 t¤}ð 	½U¼Y×=SÒ=SÑ=UÔ=Uð 	Ø�f�9ÐØ�F‘
˜aÑÐrF   rP   rn  ro  )rU   rV   rW   rX   r   r4   r^   r“   rq  r	   r5   ræ   r=  r[   r   rp  rB  r•   r{  r   r`   r   rl  ra   rb   s   @rD   r   r   Õ  sK  ø€ € € € € ðð ð(˜cð (¨Sð (ð (ð (ð (ð (ð (ð „YÔð+  b¤m°R´]Ð&BÔ Cð +ð +ð +ñ Ôð+ð:ð :¨Cð :¸hÀs¼mð :ð :ð :ð :ð :ð :ð „YÔð)ð )¨Tð )ð )ð )ñ Ôð)ð  %ð ð  à�E”LÔ!ð ð ð ð 
ˆuŒ|˜T %¤,Ô/Ð/Ô	0ð	 ð  ð  ð  ð  ð  ð  ð  rF   Ú
state_dictÚmodelc                 óÎ  — d| vr7dD ]4}t          |                      |¦  «        t          ¦  «        r
| |         }  nŒ5t          |t          ¦  «        sd„ |                      ¦   «         D ¦   «         } t          | ¦  «        } |                     ¦   «         D ]R\  }}t          |t          j        ¦  «        r3dD ]0}|› d|› �}|| v r#| |                              |j	        ¦  «        | |<   Œ1ŒS| S )aÔ  Map upstream EfficientViM checkpoints onto this module's names.

    Module names match the reference implementation. Normalization parameters are flattened,
    training-checkpoint envelopes are unwrapped, and the parallel distillation head is
    dropped for non-distilled models. The released checkpoints
    nest the weights under `model` (final) and `model_ema` (EMA); the EMA weights match the
    paper's reported top-1, so they are preferred when present.
    zpatch_embed.conv.0.conv.weight)Ú	model_emarƒ  r‚  c                 óP   — i | ]#\  }}|                      d ¦  «        °|dk    ° ||“Œ$S )zheads_dist.r$  )Ú
startswith)râ   ÚkÚvs      rD   ú
<dictcomp>z(checkpoint_filter_fn.<locals>.<dictcomp>  s;   € ÐvÐvÐv™t˜q !À!Ç,Â,È}ÑB]ÔB]ÐvÐabÐftÒatÐat�a˜ÐatÐatÐatrF   )r?   r2   ú.)
r9   Úgetr  r   ÚitemsÚnamed_modulesr5   r   r�   Únormalized_shape)r‚  rƒ  Úenvelope_keyr/  r  ÚsuffixÚkeys          rD   Úcheckpoint_filter_fnr“  	  s  € ð (¨zÐ9Ð9Ø@ð 	ð 	ˆLÝ˜*Ÿ.š.¨Ñ6Ô6½Ñ=Ô=ð Ø'¨Ô5�
Ø�ðõ �eÕ2Ñ3Ô3ð wàvÐv z×'7Ò'7Ñ'9Ô'9ÐvÑvÔvˆ
å�jÑ!Ô!€JØ×+Ò+Ñ-Ô-ð Wð W‰ˆˆfÝ�f�bœlÑ+Ô+ð 	WØ,ð Wð W�ØÐ(Ð( Ð(Ð(�Ø˜*Ð$Ð$Ø&0°¤o×&=Ò&=¸fÔ>UÑ&VÔ&V�J˜s‘OøøØÐrF   rt  c                  ó:   — dddd ddt           t          dddd	d
dœ| ¥S )Nr  rî   )r!   éà   r•  g      ì?Úbicubiczpatch_embed.conv.0.conv)úheads.0úheads.1úheads.2úheads.3zarXiv:2411.15241z&https://github.com/mlvlab/EfficientViMÚmit)Úurlrú   Ú
input_sizeÚ	pool_sizeÚcrop_pctÚinterpolationrb  r*  Ú
first_convÚ
classifierÚ	paper_idsÚ
origin_urlÚlicenser   )rt  s    rD   Ú_cfgr¦  (  sE   € àØØ#ØØØ"Ý%Ý#Ø/ð CØ'Ø>Øðð ð  ð!ð rF   ztimm/)Ú	hf_hub_id)r—  r˜  r™  rš  zheads_dist.0zheads_dist.1zheads_dist.2zheads_dist.3)r§  r¢  )r!   rñ   rñ   )r§  r�  )r§  r�  r¢  )zefficientvim_m1.e300_in1kzefficientvim_m1.e450_in1kzefficientvim_m1_dist.in1kzefficientvim_m2.e300_in1kzefficientvim_m2.e450_in1kzefficientvim_m2_dist.in1kzefficientvim_m3.e300_in1kzefficientvim_m3.e450_in1kzefficientvim_m3_dist.in1kzefficientvim_m4.e300_in1kzefficientvim_m4.e450_in1kzefficientvim_m4_dist.in1kFÚvariantÚ
pretrainedÚ	distilledc                 ón   — t          dd¬¦  «        }t          |rt          nt          | |ft          |dœ|¤ŽS )N)r   r   rp   Úgetter)Úout_indicesÚfeature_cls)Úpretrained_filter_fnÚfeature_cfg)r  r   r   r   r“  )r¨  r©  rª  rt  r°  s        rD   Ú_create_efficientvimr±  �  sX   € õ  9¸(ÐCÑCÔC€KÝØ!*Ð<ÕÐµØØðõ 2Øðð ð ðð ð rF   c           	      óX   — t          ddd¬¦  «        }t          dd| it          |fi |¤Ž¤ŽS )z6EfficientViM-M1: 6.7M params, 72.9 top-1 @ 300 epochs.©rð   éÀ   é@  ró   rô   ©rû   rü   rœ   Úefficientvim_m1r©  )r·  ©r  r±  ©r©  rt  Ú
model_argss      rD   r·  r·  ¡  óN   € õ Ø!ØØðñ ô €Jõ
  ÐgÐg¸jÐgÍDÐQ[ÐLfÐLfÐ_eÐLfÐLfÐgÐgÐgrF   c           	      óX   — t          ddd¬¦  «        }t          dd| it          |fi |¤Ž¤ŽS )z7EfficientViM-M2: 13.9M params, 75.4 top-1 @ 300 epochs.rï   ró   rô   r¶  Úefficientvim_m2r©  )r½  r¸  r¹  s      rD   r½  r½  ¬  r»  rF   c           	      óX   — t          ddd¬¦  «        }t          dd| it          |fi |¤Ž¤ŽS )z7EfficientViM-M3: 16.6M params, 77.6 top-1 @ 300 epochs.©r•  rµ  rò   ró   rô   r¶  Úefficientvim_m3r©  )rÀ  r¸  r¹  s      rD   rÀ  rÀ  ·  r»  rF   c           	      óX   — t          ddd¬¦  «        }t          dd| it          |fi |¤Ž¤ŽS )zFEfficientViM-M4: 19.6M params, 79.4 top-1 @ 300 epochs, 256x256 input.r¿  ©r!   rq   rp   ©r™   é    r  r¶  Úefficientvim_m4r©  )rÅ  r¸  r¹  s      rD   rÅ  rÅ  Â  sN   € õ Ø!ØØðñ ô €Jõ
  ÐgÐg¸jÐgÍDÐQ[ÐLfÐLfÐ_eÐLfÐLfÐgÐgÐgrF   c           	      ó\   — t          ddd¬¦  «        }t          	 d| ddœt          |fi |¤Ž¤ŽS )	z4EfficientViM-M1, distilled: 6.7M params, 74.6 top-1.r³  ró   rô   r¶  Úefficientvim_m1_distT©r©  rª  )rÇ  r¸  r¹  s      rD   rÇ  rÇ  Í  ób   € õ Ø!ØØðñ ô €Jõ
  ØðàØðð õ ˆzÐ
$Ð
$˜VÐ
$Ð
$ð	ð ð rF   c           	      ó\   — t          ddd¬¦  «        }t          	 d| ddœt          |fi |¤Ž¤ŽS )	z5EfficientViM-M2, distilled: 13.9M params, 76.7 top-1.rï   ró   rô   r¶  Úefficientvim_m2_distTrÈ  )rË  r¸  r¹  s      rD   rË  rË  Ý  rÉ  rF   c           	      ó\   — t          ddd¬¦  «        }t          	 d| ddœt          |fi |¤Ž¤ŽS )	z5EfficientViM-M3, distilled: 16.6M params, 79.1 top-1.r¿  ró   rô   r¶  Úefficientvim_m3_distTrÈ  )rÍ  r¸  r¹  s      rD   rÍ  rÍ  í  rÉ  rF   c           	      ó\   — t          ddd¬¦  «        }t          	 d| ddœt          |fi |¤Ž¤ŽS )	zDEfficientViM-M4, distilled: 19.6M params, 80.7 top-1, 256x256 input.r¿  rÂ  rÃ  r¶  Úefficientvim_m4_distTrÈ  )rÏ  r¸  r¹  s      rD   rÏ  rÏ  ý  sb   € õ Ø!ØØðñ ô €Jõ
  ØðàØðð õ ˆzÐ
$Ð
$˜VÐ
$Ð
$ð	ð ð rF   rP   )FFro  )CrX   Ú	functoolsr   Útypingr   r   r   r   r   r	   r
   r   r^   Útorch.nnr5   Útorch.nn.utils.fusionr   Ú	timm.datar   r   Útimm.layersr   r   r   r   r   r   r   Ú_builderr   Ú	_featuresr   Ú_manipulater   Ú	_registryr   r   Ú__all__r\   r    re   rl   rw   r_   rg   r`   r…   r‡   r˜   rÈ   rÝ   r   r   rp  r“  r¦  Údefault_cfgsr•   r±  r·  r½  rÀ  rÅ  rÇ  rË  rÍ  rÏ  r
  rF   rD   ú<module>rÜ     s¡  ððð ð4 Ð Ð Ð Ð Ð Ø EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ EÐ Eà €€€Ø Ð Ð Ð Ð Ð Ø 3Ð 3Ð 3Ð 3Ð 3Ð 3à AÐ AÐ AÐ AÐ AÐ AÐ AÐ Aðð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð ð +Ð *Ð *Ð *Ð *Ð *Ø +Ð +Ð +Ð +Ð +Ð +Ø #Ð #Ð #Ð #Ð #Ð #Ø <Ð <Ð <Ð <Ð <Ð <Ð <Ð <àÐ2Ð
3€ð3ð 3ð 3ð 3ð 3�"”)ñ 3ô 3ð 3ðl3ð 3ð 3ð 3ð 3�"”)ñ 3ô 3ð 3ðlð ð ð ð ˆ2Œ9ñ ô ð ð.%ð %ð %ð %ð %ˆ"Œ)ñ %ô %ð %ð" €„�„ð9ð 9 2¤9ð 9°X¸e¼lÔ5Kð 9ÐW[ð 9ð 9ð 9ñ „ð9ð*ð *ð *ð *ð *�2”9ñ *ô *ð *ðZCð Cð Cð Cð CˆRŒYñ Cô Cð CðLBð Bð Bð Bð B˜œ	ñ Bô Bð BðJ5ð 5ð 5ð 5ð 5˜œ	ñ 5ô 5ð 5ðpJ$ð J$ð J$ð J$ð J$�2”9ñ J$ô J$ð J$ðZ1 ð 1 ð 1 ð 1 ð 1 ˜Lñ 1 ô 1 ð 1 ðhØ˜˜eœlÐ*Ô+ðàŒyðð 
ˆ#ˆuŒ|Ð
Ôðð ð ð ð>�3ð ˜4  S œ>ð ð ð ð ð* %Ð$Ø!% Øð"ñ "ô "ð "& Øð"ñ "ô "ð "& Øð	
ð"ñ "ô "ð "& Øð"ñ "ô "ð "& Øð"ñ "ô "ð "& Øð	
ð"ñ "ô "ð "& Øð"ñ "ô "ð "& Øð"ñ "ô "ð "& Øð	
ð"ñ "ô "ð "& ØØ ð"ñ "ô "ð "& ØØ ð"ñ "ô "ð "& ØØ ð	
ð"ñ "ô "ðEP&ð P&ñ Pô P€ðj !Øðð Øðàðð ðð ð	ð
 ðð ð ð ð" ðhð h ð h¸ð hÀð hð hð hñ „ðhð ðhð h ð h¸ð hÀð hð hð hñ „ðhð ðhð h ð h¸ð hÀð hð hð hñ „ðhð ðhð h ð h¸ð hÀð hð hð hñ „ðhð ðð  Tð ¸Sð À\ð ð ð ñ „ðð ðð  Tð ¸Sð À\ð ð ð ñ „ðð ðð  Tð ¸Sð À\ð ð ð ñ „ðð ðð  Tð ¸Sð À\ð ð ð ñ „ðð ð rF   