3j'#SrSSKJr SSKJrJrJr SSKrSSKJ r SSK J r J r SSK Jr "SS\ R5rg) zBase 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. ) nullcontext)AnyDictOptionalN)get_state_dict unwrap_model) ModelEmaV3c ^\rSrSrSrS*S\\RS\\RS\ 4U4Sjjjr U4Sjr S+S \\ S S4S jjr S,S \S \\S \\R 4SjjrS-S\ S \\R 4SjjrS.S\S\ S\\RS \R 4SjjrS \ 4SjrS,S \S \\S \\R 4SjjrS+S\\S S4SjjrS-S\ 4SjjrS/S\\R S\ S \\R 4SjjrS/S\\R S\ S \\\44SjjrS0S\\\\4S\ S\\R S\ S S4 S jjrS\4S\ S \\\44S!jjr S1S"\\\4S#\\\\4S\ S\ S S4 S$jjr!S%r"S&\RFS'\RFS \\\RF44S(jr$S)r%U=r&$)2 TrainingTaskaBase class for training tasks. A training task encapsulates a complete forward pass including loss computation. Tasks return a dictionary containing the training loss and other components for logging. The returned dictionary must contain: - 'loss': The training loss for backward pass (required) - 'output': Model output/logits for metric computation (recommended) - Other task-specific loss components for logging (optional) Args: device: Device for task tensors/buffers (defaults to cpu) dtype: Dtype for task tensors/buffers (defaults to torch default) verbose: Enable info logging Example: >>> task = SomeTask(model, criterion, device=torch.device('cuda')) >>> >>> # Prepare for distributed training (if needed) >>> if distributed: >>> task.prepare_distributed(device_ids=[local_rank]) >>> >>> # Training loop >>> result = task(input, target) >>> result['loss'].backward() Ndevicedtypeverbosec>[TU]5 UbUO[R"S5UlUbUO[R"5UlX0lg)Ncpu)super__init__torchr get_default_dtyperr)selfr rr __class__s H/home/wildlama/miniconda3/lib/python3.13/site-packages/timm/task/task.pyrTrainingTask.__init__-sC  & 2f U8K #/UU5L5L5N  c>[R"S5R"U0UD6nURUlURUl[ TU]"U0UD6$)zFMove task to device/dtype, keeping self.device and self.dtype in sync.r)remptytor rr)rargskwargsdummyrs rrTrainingTask.to8sJ A!!4262ll [[ wz4*6**r device_idsreturnc U$)a`Prepare task for distributed training. This method wraps trainable components in DistributedDataParallel (DDP) while leaving non-trainable components (like frozen teacher models) unwrapped. Should be called after task initialization but before training loop. Args: device_ids: List of device IDs for DDP (e.g., [local_rank]) **ddp_kwargs: Additional arguments passed to DistributedDataParallel Returns: self (for method chaining) Example: >>> task = LogitDistillationTask(student, teacher, criterion) >>> task.compile() >>> task.prepare_distributed(device_ids=[args.local_rank]) )rr" ddp_kwargss rprepare_distributed TrainingTask.prepare_distributed?s 2 rbackendmodec g)aCompile hot task components before distributed wrapping. Subclasses should compile the train/eval modules that do the tensor work, not the outer task wrapper. The return value is the eval-facing compiled module/callable used by validation and checkpoint export. Nr%rr)r*compile_kwargss rcompileTrainingTask.compileZsrFemacDU(a [USS5$[USU5$)z1Return the module that owns trainable parameters.trainable_module_emaNtrainable_module)getattr)rr0s rget_trainable_module!TrainingTask.get_trainable_modulehs% 4!7> >t/66rdecay use_warmupc `[UR54UUUS.UD6UlUR$)z+Create an EMA copy of the trainable module.)r7r8r )r r5r2)rr7r8r rs r setup_emaTrainingTask.setup_emansC%/  % % '% ! %  % !(((rc$URSS9SL$)z5Return whether this task has an EMA trainable module.Tr0N)r5)rs rhas_emaTrainingTask.has_emas((T(2$>>rc UR5(dg[R"URSS94UUS.UD6UlUR$)z3Compile the EMA eval model if one has been created.NTr=)r)r*)r>rr.get_eval_modeleval_model_emar,s r compile_emaTrainingTask.compile_emasW||~~#mm   D  )     """rstepcUR5(a5URSS9R[UR55US9 gg)z3Update EMA state from the current trainable module.Tr=)rEN)r>r5updater)rrEs r update_emaTrainingTask.update_emasB <<>>  % %$ % / 6 6|DD]D]D_7`gk 6 l r exclude_headcUR5nU(a#UR5Vs/sHo3PM snSS$UR5$s snf)z/Return parameters to use for gradient clipping.N)r5 parameters)rrJr3ps rget_clip_parameters TrainingTask.get_clip_parameterssO446 /::<=s A modulecvUc5U(aSOSn[X5(a [X5$URUS9nU$)zReturn 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. rB eval_modelr=)hasattrr4r5)rrQr0 eval_attrs rrATrainingTask.get_eval_modelsA >,/(\It''t//..3.7F rc0$)z:Return task-owned state outside the eval model state_dict.r%)rrQr0s rget_task_stateTrainingTask.get_task_states rstatestrictcg)z8Load task-owned state outside the eval model state_dict.Nr%)rrZr[rQr0s rload_task_stateTrainingTask.load_task_statesrcURUS9nUc0$U(aSOSnU(aSOSnURUS9nU[X250nU(aXgU'U$)z3Return checkpoint state entries owned by this task.r=state_dict_ema state_dicttask_state_ema task_state)rArXr)rr0 unwrap_fnrS model_keytask_keyrcrZs rget_checkpoint_state!TrainingTask.get_checkpoint_statesg ((S(1  I(+$ '*# ((S(1 N:AB ((O rrarccURUS9nUcU(a [S5e[S5e[U5RXS9 UR X$US9 g)z-Load model and task-owned checkpoint entries.r=Nz4Cannot load EMA checkpoint state before setup_ema().z3Cannot load checkpoint state without an eval model.)r[)r[r0)rA RuntimeErrorrload_state_dictr])rrarcr0r[rSs rload_checkpoint_state"TrainingTask.load_checkpoint_statesa((S(1  "#YZZTU UZ 000K ZC@rcUR5nUb%XLa![US5(aUR5$[5$)zReturn a no-sync context for gradient accumulation. Tasks that wrap a trainable component with DDP delegate to that component's no_sync(). Non-distributed tasks use a no-op context. no_sync)r5rTror)rrQs rroTrainingTask.no_syncs> **,  &"49S9S>># #}rinputtargetc[e)zPerform forward pass and compute loss. Args: input: Input tensor [B, C, H, W] target: Target labels [B] Returns: Dictionary with at least 'loss' key containing the training loss )NotImplementedError)rrqrrs rforwardTrainingTask.forwards "!r)r rrBr2r)NNT)N)inductorN)F)gH.?FN)NF)TNF)NFT)'__name__ __module__ __qualname____firstlineno____doc__rrr rboolrrlistr'strnnModuler.r5floatr:r>rCintrHrOrArrrXr]rrgrlroTensorru__static_attributes__ __classcell__)rs@rr r s:.2+/ U\\* EKK(   +*.   :&"&  3-  "))  77"))9L7"$-1 )))U\\* )  )"?? &"&##3-# "))  #"mx}mm -- Xbii%8 d W_`b`i`iWj Xbii%8dW[\_ad\dWe *. DcN+RYY'     " c3h ,48 AS#XA!c3h0A A  A  A "<<"LL" c5<< ""rr )r| contextlibrtypingrrrrtorch.nnrtimm.utils.modelrrtimm.utils.model_emar rr r%rrrs3 #&& 9+f"299f"r