§
    ·¯¸j'#  ã                   ó€   — d Z ddlmZ ddlmZmZmZ ddlZddlm	Z	 ddl
mZmZ ddlmZ  G d„ de	j        ¦  «        ZdS )	zàBase training task abstraction.

This module provides the base TrainingTask class that encapsulates a complete
forward pass including loss computation. Tasks return a dictionary with loss
components and outputs for logging.
é    )Únullcontext)ÚAnyÚDictÚOptionalN)Úget_state_dictÚunwrap_model)Ú
ModelEmaV3c                   ó8  ‡ — e Zd ZdZ	 	 	 d+deej                 deej                 defˆ fd„Z	ˆ fd„Z
	 d,d	ee         d
d fd„Z	 	 d-dedee         d
eej                 fd„Zd.ded
eej                 fd„Z	 	 	 d/dededeej                 d
ej        fd„Zd
efd„Z	 	 d-dedee         d
eej                 fd„Zd,dee         d
dfd„Zd.defd„Zd0deej                 ded
eej                 fd„Zd0deej                 ded
eeef         fd„Z	 	 	 d1d eeeef                  d!edeej                 ded
df
d"„Zdefded
eeef         fd#„Z	 	 	 d2d$eeef         d%eeeef                  ded!ed
df
d&„Z d'„ Z!d(ej"        d)ej"        d
eeej"        f         fd*„Z#ˆ xZ$S )3ÚTrainingTaska  Base class for training tasks.

    A training task encapsulates a complete forward pass including loss computation.
    Tasks return a dictionary containing the training loss and other components for logging.

    The returned dictionary must contain:
        - 'loss': The training loss for backward pass (required)
        - 'output': Model output/logits for metric computation (recommended)
        - Other task-specific loss components for logging (optional)

    Args:
        device: Device for task tensors/buffers (defaults to cpu)
        dtype: Dtype for task tensors/buffers (defaults to torch default)
        verbose: Enable info logging

    Example:
        >>> task = SomeTask(model, criterion, device=torch.device('cuda'))
        >>>
        >>> # Prepare for distributed training (if needed)
        >>> if distributed:
        >>>     task.prepare_distributed(device_ids=[local_rank])
        >>>
        >>> # Training loop
        >>> result = task(input, target)
        >>> result['loss'].backward()
    NTÚdeviceÚdtypeÚverbosec                 óÈ   •— t          ¦   «                              ¦   «          |�|nt          j        d¦  «        | _        |�|nt          j        ¦   «         | _        || _        d S )NÚcpu)ÚsuperÚ__init__Útorchr   Úget_default_dtyper   r   )Úselfr   r   r   Ú	__class__s       €ú^/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/timm/task/task.pyr   zTrainingTask.__init__-   sY   ø€ õ 	‰Œ×ÒÑÔÐØ &Ð 2�f�f½¼ÀUÑ8KÔ8KˆŒØ#Ð/�U�UµUÔ5LÑ5NÔ5NˆŒ
ØˆŒˆˆó    c                 ó¤   •—  t          j        d¦  «        j        |i |¤Ž}|j        | _        |j        | _         t          ¦   «         j        |i |¤ŽS )zFMove task to device/dtype, keeping self.device and self.dtype in sync.r   )r   ÚemptyÚtor   r   r   )r   ÚargsÚkwargsÚdummyr   s       €r   r   zTrainingTask.to8   sO   ø€ à!•”˜A‘”Ô! 4Ð2¨6Ð2Ð2ˆØ”lˆŒØ”[ˆŒ
Ø�u‰wŒwŒz˜4Ð* 6Ð*Ð*Ð*r   Ú
device_idsÚreturnc                 ó   — | S )aÈ  Prepare task for distributed training.

        This method wraps trainable components in DistributedDataParallel (DDP)
        while leaving non-trainable components (like frozen teacher models) unwrapped.

        Should be called after task initialization but before training loop.

        Args:
            device_ids: List of device IDs for DDP (e.g., [local_rank])
            **ddp_kwargs: Additional arguments passed to DistributedDataParallel

        Returns:
            self (for method chaining)

        Example:
            >>> task = LogitDistillationTask(student, teacher, criterion)
            >>> task.compile()
            >>> task.prepare_distributed(device_ids=[args.local_rank])
        © )r   r   Ú
ddp_kwargss      r   Úprepare_distributedz TrainingTask.prepare_distributed?   s	   € ð2 ˆr   ÚinductorÚbackendÚmodec                 ó   — dS )a'  Compile hot task components before distributed wrapping.

        Subclasses should compile the train/eval modules that do the tensor
        work, not the outer task wrapper. The return value is the eval-facing
        compiled module/callable used by validation and checkpoint export.
        Nr"   ©r   r&   r'   Úcompile_kwargss       r   ÚcompilezTrainingTask.compileZ   s	   € ð ˆtr   FÚemac                 óJ   — |rt          | dd¦  «        S t          | d| ¦  «        S )z1Return the module that owns trainable parameters.Útrainable_module_emaNÚtrainable_module)Úgetattr)r   r,   s     r   Úget_trainable_modulez!TrainingTask.get_trainable_moduleh   s0   € àð 	?Ý˜4Ð!7¸Ñ>Ô>Ð>Ý�tÐ/°Ñ6Ô6Ð6r   ç§èH.ÿï?ÚdecayÚ
use_warmupc                 ó`   — t          |                      ¦   «         f|||dœ|¤Ž| _        | j        S )z+Create an EMA copy of the trainable module.)r3   r4   r   )r	   r1   r.   )r   r3   r4   r   r   s        r   Ú	setup_emazTrainingTask.setup_eman   sM   € õ %/Ø×%Ò%Ñ'Ô'ð%
àØ!Øð	%
ð %
ð
 ð%
ð %
ˆÔ!ð Ô(Ð(r   c                 ó2   — |                       d¬¦  «        duS )z5Return whether this task has an EMA trainable module.T©r,   N)r1   )r   s    r   Úhas_emazTrainingTask.has_ema   s   € à×(Ò(¨TÐ(Ñ2Ô2¸$Ð>Ð>r   c                 ó˜   — |                       ¦   «         sdS t          j        |                      d¬¦  «        f||dœ|¤Ž| _        | j        S )z3Compile the EMA eval model if one has been created.NTr8   )r&   r'   )r9   r   r+   Úget_eval_modelÚeval_model_emar)   s       r   Úcompile_emazTrainingTask.compile_emaƒ   sf   € ð �|Š|‰~Œ~ð 	Ø�4Ý#œmØ×Ò DÐÑ)Ô)ð
àØð
ð 
ð ð	
ð 
ˆÔð Ô"Ð"r   Ústepc                 óÆ   — |                       ¦   «         rL|                      d¬¦  «                             t          |                      ¦   «         ¦  «        |¬¦  «         dS dS )z3Update EMA state from the current trainable module.Tr8   )r>   N)r9   r1   Úupdater   )r   r>   s     r   Ú
update_emazTrainingTask.update_ema”   sd   € à�<Š<‰>Œ>ð 	mØ×%Ò%¨$Ð%Ñ/Ô/×6Ò6µ|ÀD×D]ÒD]ÑD_ÔD_Ñ7`Ô7`ÐgkÐ6ÑlÔlÐlÐlÐlð	mð 	mr   Úexclude_headc                 ó¢   — |                       ¦   «         }|r&d„ |                     ¦   «         D ¦   «         dd…         S |                     ¦   «         S )z/Return parameters to use for gradient clipping.c                 ó   — g | ]}|‘ŒS r"   r"   )Ú.0Úps     r   ú
<listcomp>z4TrainingTask.get_clip_parameters.<locals>.<listcomp>�   s   € Ð=Ð=Ð=˜!�AÐ=Ð=Ð=r   Néþÿÿÿ)r1   Ú
parameters)r   rB   r/   s      r   Úget_clip_parametersz TrainingTask.get_clip_parameters™   sY   € à×4Ò4Ñ6Ô6ÐØð 	CØ=Ð=Ð/×:Ò:Ñ<Ô<Ð=Ñ=Ô=¸c¸r¸cÔBÐBØ×*Ò*Ñ,Ô,Ð,r   Úmodulec                 ó‚   — |€<|rdnd}t          | |¦  «        rt          | |¦  «        S |                      |¬¦  «        }|S )zÃReturn the eval model/callable used for validation.

        Checkpoint state_dict handling uses unwrap_model separately so DDP and
        compiled wrappers do not leak into saved keys.
        Nr<   Ú
eval_modelr8   )Úhasattrr0   r1   )r   rK   r,   Ú	eval_attrs       r   r;   zTrainingTask.get_eval_model    sV   € ð ˆ>Ø,/ÐAÐ(Ð(°\ˆIÝ�t˜YÑ'Ô'ð 0Ý˜t YÑ/Ô/Ð/Ø×.Ò.°3Ð.Ñ7Ô7ˆFØˆr   c                 ó   — i S )z:Return task-owned state outside the eval model state_dict.r"   )r   rK   r,   s      r   Úget_task_statezTrainingTask.get_task_state­   s   € àˆ	r   ÚstateÚstrictc                 ó   — dS )z8Load task-owned state outside the eval model state_dict.Nr"   )r   rR   rS   rK   r,   s        r   Úload_task_statezTrainingTask.load_task_state±   s	   € ð ˆtr   c                 ó°   — |                       |¬¦  «        }|€i S |rdnd}|rdnd}|                      |¬¦  «        }|t          ||¦  «        i}|r|||<   |S )z3Return checkpoint state entries owned by this task.r8   NÚstate_dict_emaÚ
state_dictÚtask_state_emaÚ
task_state)r;   rQ   r   )r   r,   Ú	unwrap_fnrM   Ú	model_keyÚtask_keyrZ   rR   s           r   Úget_checkpoint_statez!TrainingTask.get_checkpoint_state»   s…   € ð ×(Ò(¨SÐ(Ñ1Ô1ˆ
ØÐØˆIà(+Ð=Ð$Ð$°ˆ	Ø'*Ð<Ð#Ð#°ˆØ×(Ò(¨SÐ(Ñ1Ô1ˆ
à�N¨:°yÑAÔAÐBˆØð 	)Ø(ˆE�(‰OØˆr   rX   rZ   c                 óî   — |                       |¬¦  «        }|€ |rt          d¦  «        ‚t          d¦  «        ‚t          |¦  «                             ||¬¦  «         |                      |||¬¦  «         dS )z-Load model and task-owned checkpoint entries.r8   Nz4Cannot load EMA checkpoint state before setup_ema().z3Cannot load checkpoint state without an eval model.)rS   )rS   r,   )r;   ÚRuntimeErrorr   Úload_state_dictrU   )r   rX   rZ   r,   rS   rM   s         r   Úload_checkpoint_statez"TrainingTask.load_checkpoint_stateÎ   s‰   € ð ×(Ò(¨SÐ(Ñ1Ô1ˆ
ØÐØð [Ý"Ð#YÑZÔZÐZÝÐTÑUÔUÐUÝ�ZÑ Ô ×0Ò0°ÀFÐ0ÑKÔKÐKØ×Ò˜Z°¸CÐÑ@Ô@Ð@Ð@Ð@r   c                 óš   — |                       ¦   «         }|�(|| ur$t          |d¦  «        r|                     ¦   «         S t          ¦   «         S )zÏReturn a no-sync context for gradient accumulation.

        Tasks that wrap a trainable component with DDP delegate to that
        component's no_sync(). Non-distributed tasks use a no-op context.
        NÚno_sync)r1   rN   rd   r   )r   rK   s     r   rd   zTrainingTask.no_syncÞ   sN   € ð ×*Ò*Ñ,Ô,ˆØÐ &°Ð"4Ð"4½ÀÈÑ9SÔ9SÐ"4Ø—>’>Ñ#Ô#Ð#Ý‰}Œ}Ðr   ÚinputÚtargetc                 ó   — t           ‚)zðPerform forward pass and compute loss.

        Args:
            input: Input tensor [B, C, H, W]
            target: Target labels [B]

        Returns:
            Dictionary with at least 'loss' key containing the training loss
        )ÚNotImplementedError)r   re   rf   s      r   ÚforwardzTrainingTask.forwardé   s
   € õ "Ð!r   )NNT)N)r%   N)F)r2   FN)NF)TNF)NFT)%Ú__name__Ú
__module__Ú__qualname__Ú__doc__r   r   r   r   Úboolr   r   Úlistr$   ÚstrÚnnÚModuler+   r1   Úfloatr6   r9   r=   ÚintrA   rJ   r;   r   r   rQ   rU   r   r^   rb   rd   ÚTensorri   Ú__classcell__)r   s   @r   r   r      s2  ø€ € € € € ðð ð: .2Ø+/Ø ð		ð 	à˜Uœ\Ô*ð	ð ˜EœKÔ(ð	ð ð		ð 	ð 	ð 	ð 	ð 	ð+ð +ð +ð +ð +ð *.ðð à  œðð 
ð	ð ð ð ð: &Ø"&ðð àðð ˜3”-ðð
 
�"”)Ô	ðð ð ð ð7ð 7¨ð 7¸À"Ä)Ô9Lð 7ð 7ð 7ð 7ð "Ø$Ø-1ð	)ð )àð)ð ð)ð ˜Uœ\Ô*ð	)ð 
Œð)ð )ð )ð )ð"?˜ð ?ð ?ð ?ð ?ð &Ø"&ð#ð #àð#ð ˜3”-ð#ð
 
�"”)Ô	ð#ð #ð #ð #ð"mð m˜x¨œ}ð m¸ð mð mð mð mð
-ð -°ð -ð -ð -ð -ðð  X¨b¬iÔ%8ð Àdð ÐW_Ð`bÔ`iÔWjð ð ð ð ðð  X¨b¬iÔ%8ð Àdð ÐW[Ð\_ÐadÐ\dÔWeð ð ð ð ð  Ø*.Øðð à˜D  c œNÔ+ðð ðð ˜RœYÔ'ð	ð
 ðð 
ðð ð ð ð Ø"ðð àðð 
ˆc�3ˆhŒð	ð ð ð ð, 48ØØðAð Aà˜S #˜XœðAð !  c¨3 h¤Ô0ðAð ð	Að
 ðAð 
ðAð Að Að Að 	ð 	ð 	ð"à”<ð"ð ”Lð"ð 
ˆc�5”<ÐÔ	 ð	"ð "ð "ð "ð "ð "ð "ð "r   r   )rm   Ú
contextlibr   Útypingr   r   r   r   Útorch.nnrq   Útimm.utils.modelr   r   Útimm.utils.model_emar	   rr   r   r"   r   r   ú<module>r|      sÆ   ððð ð #Ð "Ð "Ð "Ð "Ð "Ø &Ð &Ð &Ð &Ð &Ð &Ð &Ð &Ð &Ð &à €€€Ø Ð Ð Ð Ð Ð à 9Ð 9Ð 9Ð 9Ð 9Ð 9Ð 9Ð 9Ø +Ð +Ð +Ð +Ð +Ð +ðf"ð f"ð f"ð f"ð f"�2”9ñ f"ô f"ð f"ð f"ð f"r   