9iSSKrSSKrSSKJr SSKJr SSKJr SSKJ r SSK J r J r J r JrJrJr SSKrSSKrSSKJr SSKJr SS KJrJrJr SS KJr SS KJr SS KJr SS K J!r! SSK"J#r#J$r$ SSK%J&r& SSK'J(r(J)r) SSK*J+r+ SSK,J-r- SSK.J/r/ SSK0J1r1 SSK2J3r3 SSK4J5r5 SSK6J7r7J8r8 SSK9J:r: SSK;Jr>J?r?J@r@ SSKAJBrBJCrCJDrDJErEJFrFJGrGJHrHJIrIJJrJ SSKKJLrL SSKMJNrN SSKOJPrP SSKQJRrR SS KSJTrT SS!KUJVrV SS"KWJXrXJYrYJZrZJ[r[J\r\J]r]J^r^J_r_ S#S$K`Jara S#S%KbJcrc S#S&KdJereJfrfJgrg S#S'KhJiri S#S(KjJkrk S#S)KlJmrm \S*rn\cR"\!RS+9"S,S-\a55rqS.rrg)/N)Mapping)deepcopy) LooseVersion)partial)CallableDictListOptionalTupleUnion) distributed)nn) DataLoaderDatasetSampler)default_collate)DistributedSampler)check_local_model_is_latest)Trainers) build_metrictask_default_metrics)PredictionSavingWrapper)Model TorchModel)TorchCustomDataset)build_custom_dataset) MsDataset)ModelOutputBase) Preprocessor)HOOKS)Priority get_priority)build_lr_scheduler)build_optimizer)Config ConfigDictJSONIteratorEncoder) DEFAULT_MODEL_REVISION ConfigFields ConfigKeysDistributedParallelTypeInvokeModeKeys ModelFile ThirdParty TrainerStages) to_device) create_device)func_receive_dict_inputs)is_swift_available) get_logger)build_from_cfg) compile_model get_dist_infoget_local_rank init_distis_dist is_masteris_on_same_deviceset_random_seed) BaseTrainer)TRAINERS) merge_cfg merge_hooks update_cfg)Hook)build_parallel) is_parallel)zswift.SwiftConfigzswift.PeftConfig) module_namec ^\rSrSrSrSSSSSSSSS\SSSS4S\\\\ R\ 4S\\ S\\ S \\ S \\\ \ \ \ 44S \\\\4S \\\\4S \\\\ \ \44S\\R&R(\R&R*R,4S\\ S\S\\\S\\\\ \ \44S\\ \ \4\44U4SjjjrSrSrS]SjrS^Sjr S^Sjr!Sr"S\#4Sjr$\%S5r&\%S5r'\%S5r(S r)S!r*S"r+\%S#5r,\%S$\\4S%j5r-\%S$\4S&j5r.\%S$\4S'j5r/\%S$\4S(j5r0\%S)5r1\%S*5r2\%S+5r3S^S,\\\\\4S-\#S.\ S \\4S/jjr4S0\S.\ S \S$\54S1jr6\7S^S-\#S.\ S \4S2jj5r8S$\\\44S3jr9S$\\\ \ 44S4jr:S5r;S_S6jr<S`S7\\\\44S8jjr=SS9jr>\%S:5r?S$\\ R\44S;jr@S$\\ R\44S<jrAS$\\ R\44S=jrBS>rCS?rDS@rESArFS7\\\\44SBjrGS^S\HSC\I4SDjjrJS^S\HSC\I4SEjjrKSFrLSGrMSaS0\SH\SI\SJ\NSK\NS\S$\O4SLjjrPSMrQSNrRSOrSSPrTSQ\S$S4SRjrUSS\S$\4STjrVSUrWSVrXSW\ S$S4SXjrYSYrZSZr[S$\ 4S[jr\S\r]U=r^$)bEpochBasedTrainer<a< Epoch based Trainer, a training helper for PyTorch. Args: cfg_file(str): The local config file. model (:obj:`torch.nn.Module` or :obj:`TorchModel` or `str`): The model to be run, or a valid model dir or a model id. If model is None, build_model method will be called. data_collator (`Callable`, *optional*): The function to use to form a batch from a list of elements of `train_dataset` or `eval_dataset`. train_dataset (`MsDataset` or `torch.utils.data.Dataset`, *optional*): The dataset to use for training. Note that if it's a `torch.utils.data.IterableDataset` with some randomization and you are training in a distributed fashion, your iterable dataset should either use a internal attribute `generator` that is a `torch.Generator` for the randomization that must be identical on all processes (and the Trainer will manually set the seed of this `generator` at each epoch) or have a `set_epoch()` method that internally sets the seed of the RNGs used. eval_dataset (`MsDataset` or `torch.utils.data.Dataset`, *optional*): The dataset to use for evaluation. preprocessor (:obj:`Preprocessor`, *optional*): The optional preprocessor. NOTE: If the preprocessor has been called before the dataset fed into this trainer by user's custom code, this parameter should be None, meanwhile remove the 'preprocessor' key from the cfg_file. Else the preprocessor will be instantiated from the cfg_file or assigned from this parameter and this preprocessing action will be executed every time the dataset's __getitem__ is called. optimizers (`Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler._LRScheduler]`, *optional*): A tuple containing the optimizer and the scheduler to use. seed (int): The optional random seed for torch, cuda, numpy and random. max_epochs: (int, optional): Total training epochs. cfg_modify_fn: An input fn which is used to modify the cfg read out of the file. remove_unused_data: Automatically remove unused data keys in mini-batches. The remove action based on the `inspect` on the model's forward method, the removed columns will be moved to the mini-batch's attributes. compile (bool, optional): Compile the model with torch 2.0, default False compile_options (dict, optional): The compile options if compile=True, default None to use the default params of 'TorchModel.compile'. efficient_tuners (dict, optional): The tuners to use to train the model samplers: (:obj:`Sampler` or `Dict[Sampler]`, *optional*): samplers used in the train/eval DataLoader. Examples of cfg_modify_fn: >>> def cfg_modify_fn(cfg): >>> cfg.preprocessor.first_sequence= 'text1' >>> cfg.preprocessor.second_sequence='text2' >>> return cfg NNN*modelcfg_file cfg_modify_fn arg_parse_fn data_collator train_dataset eval_dataset preprocessor optimizersmodel_revisionseed callbackssamplersefficient_tunersc 8>Xl[UR5 SUlXl[R Ul/UlSUlSUl SUl SUl URSS5Ul SUlSUlSUlXl[%U[&5(awUR)XUR+[,R.S55UlUc8[2R4R7UR0[8R:5nXlOUcS5e[2R4R?U5UlSUl[AUS5(a`[CUR0[DR.[DRF[,R.UR+[,R.S50S9 [HTU]X$5 X0l&[OURP5 URSURP5Ul(SU;aURPRUUS5 [WURP5Ul([%U[X[ZR\45(aXl/OURa5Ul/UR(a1URS5nUc0n[cUR^40UD6Ul/URS S5GbUS Ul2S URP;a[g5URPS 'URdURPS S 'S URPS ;apS URPS S ;a"URdURPS S S S 'SURPS S ;a"URdURPS S SS 'SURPS ;aURdURPS SS'O+URPRhRS S5Ul2URkU5uUl6Ul7[2R4RqURd5(d[2Rr"URdSS9 [2R4R7URdSRuURv55n[yUURPRSS5S9Ul=UR|"S,UURP[R URlS.UD6Ul?UR|"S,UURP[RURnS.UD6UlAURUURSS5S9uUlCUlDURSURPRS55UlFURcS5eURSURPRS55UlGURS URPRS!55UlHURS"S5UlIURS#5UlJURS$5UlKURU5 0UlMURbURPRS%5(dhS&URPRh;a/URPRhS&'URPRhS&RS'URS(.5 [URP5nURU5 [U 5(aU /n U =(d /HnURU5 M UR[R5 UR5=(a# [R"UR5S):UlZUR5Ul\UR(dURb UROS*n[U5UlKURRS+:Xa?[UR^5(a%UR^RUR5 UR5 g)-NrFcompilez?Config file should not be None if model is not from pretrained! model_dir) user_agent cfg_optionscompile_optionswork_dirtrain checkpointperiodsave_dirbestloggingout_dirz ./work_dirT)exist_okz{}.log log_levelINFO)log_filerkdatasets model_cfgmoderUremove_unused_data)rr max_epochsztrain.max_epochszzmax_epochs should be provided by the init arguments or configured in the `train.max_epochs` key in the configuration file.train_iters_per_epochztrain.train_iters_per_epochval_iters_per_epochzevaluation.val_iters_per_epochuse_fp16launcherdeviceztrain.hooks.DDPHookhooksDDPHook)typerwr?gpucuda)b_seedr>_metric_valuesrVr-TRAIN_mode_hooks_epoch_iter _inner_iter_stop_trainingget_compiletrain_dataloadereval_dataloader data_loader _samplers isinstancestrget_or_download_model_dirpopr/KEYr^ospathjoinr. CONFIGURATIONinput_model_iddirnamehasattrrr, LOCAL_TRAINERsuper__init__rPrBcfgrebuild_configmerge_from_dictrDrrModulerN build_modelr7rbr&rcget_preprocessorstrain_preprocessoreval_preprocessorexistsmakedirsformat timestampr5logger build_datasetrSEVALrTget_data_collatortrain_data_collatoreval_data_collatorsafe_get _max_epochs_train_iters_per_epoch_eval_iters_per_epochrvrwrx tune_moduleparallel_groupsappendrCregister_hook_from_cfgcallable register_hook invoke_hookr0 after_initis_dp_group_availabledistget_world_sizedp_group_dist get_metricsmetricsr2r{r=to print_cfg)selfrNrOrPrQrRrSrTrUrVrWrXrYrZr[kwargsrarmrycallback device_name __class__s [/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/trainers/trainer.pyrEpochBasedTrainer.__init__hs.  #"$^^ "$   # 9e4 $#! eS ! !!;;vzz*..$'GIDN77<<(1(?(?A"' ' j)j j'WW__X6DN"&D uk**+OO F$8$8" :>>4(H  0*$((&&txx0 F " HH $ $VM%: ;dhh' ej"))4 5 5J))+DJ ==$jj):;O&"$&tzzE_EDJ ::j$ ' 3":.DMdhh&$.L!,0MMDHHW j )txx00txx0>>&*mmHHW%l3H="$TXXg.|<<&*mmHHW%l3F;"$DHHW--:>--!),Y7 HHNN..z<HDM:>:P:P ;7!7ww~~dmm,, KK  577<< xt~~/NO k6)JL "//"hh00   !..!hh//   =AF=S=H9 $"9"::l&*hh&7&78J&KM+ h.h h+&,jj # HH  ; <'>#&,ZZ ! HH  > ?&A" :u5  :. jj*  )* " == $TXX->->%.'.'dhhnn,*,w' HHNN7 # * *! MM,  DHH% ##E* I  " I!RH   x (( 112//1d6I6I MM77 '') ##)-)@$++eK' 4DK{{6).? .K.K dkk* cUbB[5(d [S5eSSKJn UR UR U5Ulgg)NzGPlease install swift by `pip install ms-swift` to use efficient_tuners.r)Swift)r4 ValueErrorswiftr prepare_modelrN)rr[rs rrEpochBasedTrainer.tune_modulesD  '%'' ] $,,TZZ9IJDJ (rcURRS:XasURRUR5 [ UR5(d3UR (a!UR UR5Ulgggg)z)Place model to device, or to DDP r}N)rxr{rNrrGr to_parallelrs r place_modelEpochBasedTrainer.place_model&sb ;;  v % JJMM$++ &tzz**tzz!--djj9 0:* &rc^Sup4[U[5(a[RU;a;[U[R[5(deU[Rn[R U;a;[U[R [5(deU[R nOUc[ OUnUnUnU(a$SSKJn U4SjnU"5nU"X85nU"UU5nX44$)aGet the data collator for both training and evaluating. Args: data_collator: The input data_collator param. remove_unused_data: Remove the unused data with 'RemoveColumnsCollator'. Returns: The train_data_collator and eval_data_collator, can be None. rLr)RemoveColumnsCollatorc>[R"TRR5n[ UR R 55$N)inspect signaturerNforwardlist parameterskeys)rrs r _set_signature_columns_if_neededMEpochBasedTrainer.get_data_collator.._set_signature_columns_if_neededHs7#--djj.@.@A I0055788r) rrr*rcrvalrmodelscope.utils.data_collatorsr) rrRrrrr collate_fnrr model_inputss ` rr#EpochBasedTrainer.get_data_collator.s3=/ mW - -=0!- 0@0@"A8LLLL&3J4D4D&E#~~.!- "?JJJJ%2:>>%B",9,A}J", !+   M 9<=L"7##3 !67I7C"E "66rcDUb [U5 [5up#US:nU$)zInit dist and returns the dist information. Args: launcher: The launcher info. Returns: _dist: If world_size is greater than 1. r?)r:r8)rrw_ world_sizers rr:EpochBasedTrainer.init_distSs)   h % Q rcbUbUOSn[5(a[5nSU3n[U5$)zxGet the device information. Args: device: The input device info. Returns: device_name: The final device name. r|zcuda:)r;r9r2)rrxr local_ranks r get_deviceEpochBasedTrainer.get_devicecs6!' 2f 99')J!*.K[))rcSnSn[U[5(aUnUnGO[U[5(a[RU;a;[U[R[ 5(deU[Rn[R U;a;[U[R [ 5(deU[R nOR[UR[R5(a)URRbUR5up#Ub[RUlUb[RUlX#4$)zGet the preprocessors information. Args: preprocessor: The input preprocessor info. Returns: The train_preprocessor and eval_preprocessor, can be None. N)rrrr*rcrrrrr)rUbuild_preprocessorr-rrqr)rrUrrs rr#EpochBasedTrainer.get_preprocessorsss"  lL 1 1!-  ,   g . .</!,z/?/?"@(KKKK%1*2B2B%C"~~-!,z~~">IIII$0$@! TXX|888800<484K4K4M 1   )&.nn  #  (%-]]  "!44rrcd[US5(aURbURU5nU$)zjA method used to rebuild the config, any subclass can override this method. Returns: The rebuilt config rP)rrPrrs rr EpochBasedTrainer.rebuild_configs1 4 ) )d.@.@.L$$S)C rc<UR[R$)z Get the data parallel group. )rr+DPrs rrEpochBasedTrainer.dp_group ##$;$>$>??rc<UR[R$)z Get the tensor parallel group. )rr+TPrs rtp_groupEpochBasedTrainer.tp_grouprrc<UR[R$)z" Get the pipeline parallel group. )rr+PPrs rpp_groupEpochBasedTrainer.pp_grouprrc<[RUR;$)z5 Get whether the data parallel group is initialized. )r+rrrs rr'EpochBasedTrainer.is_dp_group_available'))T-A-AAArc<[RUR;$)z7 Get whether the tensor parallel group is initialized. )r+rrrs ris_tp_group_available'EpochBasedTrainer.is_tp_group_availablerrc<[RUR;$)z9 Get whether the pipeline parallel group is initialized. )r+rrrs ris_pp_group_available'EpochBasedTrainer.is_pp_group_availablerrcUR$r)rrs rrqEpochBasedTrainer.modes zzrreturncUR$)z.list[:obj:`Hook`]: A list of registered hooks.)rrs rryEpochBasedTrainer.hooks{{rcUR$)zint: Current epoch.)rrs repochEpochBasedTrainer.epochrrcUR$)zint: Current iteration.)rrs riterEpochBasedTrainer.iterszzrcUR$)zint: Iteration in an epoch.)rrs r inner_iterEpochBasedTrainer.inner_iterrcUR$)zint: Maximum training epochs.)rrs rrsEpochBasedTrainer.max_epochsrrc4URUR-$)z!int: Maximum training iterations.)riters_per_epochrs r max_itersEpochBasedTrainer.max_iterss$"6"666rc6^U4SjnTR[R:Xa+TRb TR$U"TR5$TR[R :Xa+TR b TR $U"TR5$g)z"int: Total iterations of one epochc>[U5$![a+nTRRU5 [ S5eSnAff=f)NzPlease implement ``__len__`` method for your dataset, or add `train_iters_per_epoch` and `train_iters_per_epoch` to your configuration file or kwargs)len Exceptionrerrorr)rers r _get_data_len8EpochBasedTrainer.iters_per_epoch.._get_data_lensG <;'' < !!!$ ;<< AN)rqr-rrrrrr)rr's` rr!EpochBasedTrainer.iters_per_epochs < 99 &**6222$T%:%:;; YY(-- '))5111$T%9%9:: (rrorprqc 4U(d[RX#US9$[U[5(aU$[U[5(a3UR (dUR "SUUUS.UD6 UR$[U[5(a|[US[5(ad/nUHEnUR (dUR "SUUUS.UD6 URUR5 MG [SUUSS.UD6nXl U$U[R:XaSOSn URSU 35n U c0n [URR UUUS 9n U R#U 5 [%U UR&5n X l U $![(an [+S U 5 [U[[,45(dUbJ[U4UUS .[/US 5(a[1URR S 9O0D6n X l U sSn A $UsSn A $Sn A ff=f)aBuild input datasets by given model configuration and preprocessor. Args: datasets (Union[Dataset, MsDataset, List[Dataset]]): The input datasets. model_cfg (Config): The model configuration. mode (str): `train`, `eval` or `inference`. See modelscope.utils.constant.ModeKeys preprocessor (Preprocessor, Optional): The preprocessor for input data samples. Returns: Preprocessed datasets. )rprqrU custom_cfgrUrqrN)rorqrUrcrzdataset.)r{rqrorUz** build_dataset error log:rqrUrN)r{r~)rJbuild_dataset_from_cfgrrr is_customto_custom_dataset ds_instancer rtrainerr-rrr&rNr{updatertaskr$printr rdict)rrorprqrUrcustom_datasetsdatasettorch_custom_datasetdataset_mode_key data_configdata_build_configcustom_datasetr&s rrEpochBasedTrainer.build_datasetsE"> (??'@OO($677Hi00)).."#,%1!"! "  +++Hd++ QK1,1,"$'G",,11&'0)5!%&% & $**7+>+>? ((:(,!%( ($ 04,++.2hnn.D7% '008[UT5$r)r)hookrWs r?EpochBasedTrainer.set_checkpoint_file_to_hook..s Jt5G$Hr) modelscope.trainers.hooksrWrfilterryr#rrcheckpoint_fileload_all_statestrict)rcheckpoint_pathr_r`load_ckpt_hooksload_ckpt_hookrWs @rset_checkpoint_file_to_hook-EpochBasedTrainer.set_checkpoint_file_to_hooks  & D"Hzz#$O?#q(!3!5"">2&&~61@A  .0>A  -(.A  % 'rc[RUlUR5UlURUlUR 5 UR5 UR5 URXURSS55 URR5 URUR5 g)a0Start training. Args: checkpoint_path(`str`, `optional`): The previous saving checkpoint to read, usually it's a `some-file-name.pth` file generated by this trainer. load_all_state(`bool`: `optional`): Load all state out of the `checkpoint_path` file, including the state dict of model, optimizer, lr_scheduler, the random state and epoch/iter number. If False, only the model's state dict will be read, and model will be trained again. kwargs: strict(`boolean`): If strict, any unmatched keys will cause an error. r`FN)r-rrget_train_dataloaderrrregister_optimizers_hookregister_processorsprint_hook_infordrrNrc train_loop)rrar_argsrs rrcEpochBasedTrainer.trains"^^ $ 9 9 ;00 %%'   "  (()/He)D F  --.rpredict_datasetscJUR5 UR5 UbSSKJn UR X0US9 UR R 5 [RUl URU5n[US9/nUH nXl M URXg5 g)aStart prediction. Args: predict_datasets(Union[Dataset, List[Dataset]]): The datasets used to predict ground truth. saving_fn(`Callable`): The callable used to save the prediction values to files. Like: >>> class SavingFn: >>> def __init__(self): >>> self.filename = '/tmp/results.txt' >>> >>> def __call__(self, inputs, outputs): >>> import numpy as np >>> ids = inputs.ids >>> predictions = np.argmax(outputs['logits'].cpu().numpy(), axis=1) >>> with open(self.filename, 'a') as f: >>> for id, pred in zip(ids, predictions): >>> f.writelines(f'{id}, {pred}') This saving_fn's result will not be collected to one file, Training with multiprocessing please consider combining these files manually. checkpoint_path(`str`, `optional`): The previous saving checkpoint to read, usually it's a `some-file-name.pth` file or a pure PyTorch `some-file.bin` file generated by this trainer. strict(`boolean`): If strict, any unmatched keys will cause an error. NrrVr` saving_fn)rirjr\rWload_checkpointrNevalr-rrget_predict_dataloaderrr2evaluation_loop) rrnrrrar`rWpredict_dataloadermetric_classesms rpredictEpochBasedTrainer.predicts@   "   & D  . .f / 6 ]] !889IJ1IFGAI  /@rc .UR5 UR5 Ub&SSKJn UR XUR SS5S9 UR R5 [RUl UR5Ul URUl URVs/sHn[U5PM nnUbUR![#US95 UH nXlM UR'URU5nXlU$s snf)aStart evaluation. Args: checkpoint_path(`str`, `optional`): The previous saving checkpoint to read, usually it's a `some-file-name.pth` file or a pure PyTorch `some-file.bin` file generated by this trainer. saving_fn(`Callable`): The callable used to save the prediction values to files. Like: >>> class SavingFn: >>> def __init__(self): >>> self.filename = '/tmp/results.txt' >>> >>> def __call__(self, inputs, outputs): >>> import numpy as np >>> ids = inputs.ids >>> predictions = np.argmax(outputs['logits'].cpu().numpy(), axis=1) >>> with open(self.filename, 'a') as f: >>> for id, pred in zip(ids, predictions): >>> f.writelines(f'{id}, {pred}') kwargs: strict(`boolean`): If strict, any unmatched keys will cause an error. rrVr`Frprq)rirjr\rWrsrrNrtr-rrget_eval_data_loaderrrrrrrr2rvr) rrarrrrWmetricrxry metric_valuess revaluateEpochBasedTrainer.evaluates.   "   & D  . .fjj5.I / K ]] #88://=A\\J\6,v.\J   ! !"9I"N OAI ,,T-A-A-;= ,Ks-DcUR$r)rrs rrEpochBasedTrainer.metric_values&s"""rc[R"URURS9n[ U[ R 5(d[US5(a UR$[ U[ R 5(aU$g)zInstantiate a pytorch model and return. By default, we will create a model using config from configuration file. You can override this method in a subclass. )rMrNN) rrOr^rrrrrrNrrNs rrEpochBasedTrainer.build_model*s`%%dnntxxH%++w0G0G;;  ryy ) )L*rclURRSS5bY[URS5nUR[ U[ R R5/S95 [U5$[ SUS[ R R5/URS9n[U5$)Nparallel)module device_idsDistributedDataParallelT)r{rfind_unused_parametersr process_group) rrrr3r6torchr}current_devicerFr)rrNdp_cfgs rrEpochBasedTrainer.to_parallel7s 88<< D ) 5dhhz23F MMEuzz/H/H/J.KL N!&) )*#' 1134-- )f%%rc[US5(aURUR5$[U[R R 5(deU$)zLUnwrap the model until it's a naked nn.Module. Args: model: An module. r)r unwrap_modulerrrrrrs rrEpochBasedTrainer.unwrap_moduleHsF 5( # #%%ell3 3eUXX__55 55LrcLUR5 [RUl[ UR UR 5R5n[U[5(aU(dUR"S0UD6nOURU5n[U[5(aUR5n[U[5(d [S5eSU;GaS/n[/5nUH:nURUR!5Vs/sH oU;dM UPM sn5 M< 0n UHnUR#US5n U cM[%5(aaU R&R)5R+S5n [,R."U R1[,R2"555 U RXR5505 M UR6RU 5 OUR6RUS5 X@lgs snf)a Perform a training step on a batch of inputs. Subclass and override to inject custom behavior. Args: model (`TorchModel`): The model to train. inputs (`Dict[str, Union[torch.Tensor, Any]]`): The inputs and targets of the model. The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument `labels`. Check your model's documentation for all accepted arguments. Return: `torch.Tensor`: The tensor with training loss on this batch. z$"model.forward()" must return a dictlog_varslossNr}r~)rcr-rrr3rrNrrrrto_dictr6 TypeErrorsetr3rrr;dataclonerr all_reducediv_ritem log_buffer train_outputs) rrNinputsreceive_dict_inputsrdefault_keys_pattern match_keyskey_pkeyrvalues r train_stepEpochBasedTrainer.train_stepTs$  ^^ 7   tzz * 2 24 fg & &/B!MM3F3M!MM&1M m_ 5 5)113M-..BC C ] *$*8 RJ-!!$1$6$6$8I$8SSLS$8IK.H!%))#t4$yy % 0 0 2 5 5f = 43F3F3H(IJOOS**,$78 " OO " "8 , OO " "=#< =*Js  H! H! cDURRS5 [e)zDeprecated method z*This prediction_step method is deprecated.)rwarnNotImplementedError)rrNrs rprediction_step!EpochBasedTrainer.prediction_steps EF!!rcURcSe0nURbK[UR[5(aUR[R O URUS'UR "UR4URURURS.UDURR RS05D6nU$)a Builder torch dataloader for training. We provide a reasonable default that works well. If you want to use something else, you can change the config for data.train in configuration file, or subclass and override this method (or `get_train_dataloader` in a subclass. z!The train_dataset cannot be None.samplerrrXr dataloader) rSrrr6r*rc_build_dataloader_with_datasetrrrrr)r sampler_cfgrs rrg&EpochBasedTrainer.get_train_dataloaders    %5 5 >> %%/04&6&6&*^^  &";?>>  "99   4// 4  4 hhnn  r2 4 rcURcSe0nURbK[UR[5(aUR[R O URUS'SS0nUR URRRS055 UR"UR4URURURS.UDUD6nU$)zBuilder torch dataloader for evaluation. We provide a reasonable default that works well. If you want to use something else, you can change the config for dataset.eval in configuration file, or subclass and override this method in a subclass. pass z The eval_dataset cannot be None.rshuffleFrr)rTrrr6r*rr3rrRrrrrr)rrdefault_configrs rr}&EpochBasedTrainer.get_eval_data_loaders    $4 4 >> %#-dnn.2$4$4&*^^& 9=  "$U+dhh1155lBGH99   ..      rcURUUR[RURS9n0nUR bK[ UR [5(aUR [RO UR US'SS0nURURRRS055 UR"U4URURUR S.UDUD6nU$)zBuilder torch dataloader for prediction with the config of evaluation. Args: predict_datasets(Union[Dataset, List[Dataset]]): The datasets used to predict ground truth. rnrrFrr)rrr-rrrrr6r*rr3rRrrrrr)rrnr8rrrs rru(EpochBasedTrainer.get_predict_dataloaders$$%hh// %1  >> %#-dnn.2$4$4&*^^& 9=  "$U+dhh1155lBGH99 ..      r default_argsc[URUR5UUS9$![a7nURR SUS[ RS35 UeSnAff=f)Nrrz%Build optimizer error, the optimizer G is a torch native component, please check if your torch with version:  matches the config.)r$rrNKeyErrorrr%r __version__rrrr&s rr$!EpochBasedTrainer.build_optimizersx """4::.)+ +  KK  7u=<48 :  $   '( (()--iT(aTRT/5 SnT(aTRS5S:XaTRSS5 0TETEnT(aTRT/5 SnTRS5S;aSTlT(aTRS5S;aTRSS5 0TETEnX4$)aThis function used to fit `optimizer_hook` key and `lr_scheduler_hook` key for easycv configs. The logic is: If the optimizer_hook is provided and it's not TorchAMPOptimizerHook or ApexAMPOptimizerHook, (which means the hook is a complete one for optimization, which does not need the OptimizerHook), The OptimizerHook will not be registered, or else the OptimizerHook will be registered. Same logic to the LrSchedulerHook, the only difference is the condition of lr_scheduler_hook is PlateauLrSchedulerHook. If TorchAMPOptimizerHook or ApexAMPOptimizerHook is provided, self.use_fp16 will be set to False in case of the duplication of registration. Nr{r)TorchAMPOptimizerHookApexAMPOptimizerHookF)rrrrv) _lr_options_optim_optionslr_hookr optim_hookrrs r_fit_to_old_keysDEpochBasedTrainer.register_optimizers_hook.._fit_to_old_keys7s++WI6Kgkk&15MM FD)77w7 ++ZL9!N~~f%*BB % !7 ==  ' 'D2DmDE G rbatch_size_per_gpuworkers_per_gpurrc .Sn Sn UR5(aR[RRUR5n [RR UR5n U(aUn Un OUn Un UR SS5n U cU(a@[U[RRR5(d [UU U US9n O9Sn [U[RRR5(dXXS'SnUb[[XUS9OSn[[R5[S5:aXxS 'O US LaUR R#S 5 [%U4U U U UUR S S 5US.UD6nU$)aBuild dataloader using input dataset and cfg. Used by `EpochBasedTrainer.train()` and `EpochBasedTrainer.evaluate()`. In distributed training, each GPU/process has a dataloader. In non-distributed training, there is only one dataloader for all GPUs. Args: dataset (Dataset): A PyTorch dataset. batch_size_per_gpu (int): Number of training samples on each GPU, i.e., batch size of each GPU. workers_per_gpu (int): How many subprocesses to use for data loading for each GPU. dist (bool): Distributed training/test or not. Default: True. shuffle (bool): Whether to shuffle the data at every epoch. Default: True. seed (int, Optional): Seed to be used. Default: 0. runner_type (str): Type of runner. Default: `EpochBasedRunner` persistent_workers (bool): If True, the data loader will not shutdown the worker processes after a dataset has been consumed once. This allows to maintain the workers `Dataset` instances alive. This argument is only valid when PyTorch>=1.7.0. Default: False. kwargs: any keyword argument to be used to initialize DataLoader Returns: DataLoader: A PyTorch dataloader. rr?rN) num_replicasrankrr) num_workersrrXz1.7.0persistent_workersTzNpersistent_workers is invalid because your pytorch version is lower than 1.7.0 pin_memoryF) batch_sizerr batch_samplerrworker_init_fn)rrr get_rankrrrrutilsrIterableDatasetrrrrrrr@r)rr8rrrrrXrrrrrrrrinit_fnrs rr0EpochBasedTrainer._build_dataloader_with_datasethsF  % % ' '$$--dmm!*;!7 &&z;;? ",#$   !@!@A J7  !?!?@O a %^^ q5D000!"8$   ]<< = D  KK1 K"""366 001rcURR5 [URUR5R5n[ R "5 [U[5(a$U(dURR"S0UD6nOURR U5nSSS5 U$!,(df  W$=f)zaPerform a training step on a batch of inputs. Subclass and override to inject custom behavior. Nr~) rNrtr3rrrno_gradrr)rrrresults revaluation_step!EpochBasedTrainer.evaluation_steps 6   tzz * 2 24]]_$((1D++3d3++D1   _  s AB== C c Sn[URRS5(aBURRRn[ UR4SUR 0UD6nUR [R5 UR(apSSK J n U"UUURUUURRRSS5URRRSS5URS 9nO%SS K Jn U"UUURUUURS 9nUR [R 5 U$) zAEvaluation loop used by `EpochBasedTrainer.evaluate()`. N visualizationr8r)multi_gpu_test cache_dir gpu_collectF)rxrx vis_closuretmpdirrdata_loader_iters_per_gpu)single_gpu_test)rxrxrdata_loader_iters)rrrRrrrTrr0 before_valr#modelscope.trainers.utils.inferencerrxrrr after_val)rrrxrvis_cfgrrrs rrv!EpochBasedTrainer.evaluation_loops$ 488&& 8 8hh))77G!""J,0,=,=JAHJK 112 :: J*{{-'xx**..{DA HH//33M5I*.*D*DFM L+{{-'"&"<"< >M 001rc [S5e)aBvisualization function for evaluation results. Examples: >>> # draw list of images as numpy array >>> images = draw_images(num_of_visualization) >>> # set displayed name for each image >>> filenames = get_image_display_names() >>> vis_results = {'images': images, 'filenames' : filenames} >>> # visualization results will be displayed in group named eva_vis >>> self.visualization_buffer.output['eval_vis'] = vis_results Args: results (list(dict)): a list of result dict. dataset (Dataset): torch dataset object to access original data. zIc?;It344%)%>%>%@M"188C&A[[E'.J:" w/ j1122 ''- ! 99%56++O +#?? +s D"D''EE))rrrrr@rrrrrrrrrrrrPrrrxrrrTrrrwrrrrNr^rrVrrrrSrrrvrb)Fr)NT)NF)FTrF)_rI __module__ __qualname____firstlineno____doc__r(r r rrrrrrrrrr roptim Optimizerr _LRSchedulerintr rEr TunerConfigrrrrr:rrr%rpropertyrrrrrr rqryrrrrsrrrrrB staticmethodr.rrrdrcrzrrrrrrrrgr}rur&r6r$r#rrhboolrrrkr rvrrrrir5rrrjrA__static_attributes__ __classcell__)rs@rrJrJ<s(XBF&*04/3GKAE@DEIHN,B.2EI37)sE*bii"<=>ssms$H- s #8, s $E(D9A:B5C+C%DE s$E)W*<$=>s#5G);#<=s#5)-c<.?)@*A$BCsekk33#kk66CCDEs%SMs !s" T +#s$uWd3<.@%@AB%s&$Dk)9$:$/%01'ssjK:#7J * 5>&@@ @@ @@ B B B tDzsc C    77;;8>B O %gy$w-&G HO !'O  O %-\$: O bwc&2%7"=A#&#%(#-9##(5E, *D$E5 T%T "230 / #!/@!% .A"'g(>"?.A`+Z## U299j#89 &E"))Z*?$@&" eBIIz,A&B 7+r" 04uW=A']>K8L: : T j!LFJG`5:7;34:?X07X;>X9<X.2 X 15 X .1 X5?Xt 2D$"HL,($(4(2t..> -3 -4 - +  s  rrJc.X-U-U-n[U5 gr)r>) worker_idrrrX worker_seeds rrrs$y047KK r)srrcollections.abcrr,rdistutils.versionr functoolsrtypingrrr r r r rrr rrtorch.utils.datarrrtorch.utils.data.dataloaderrtorch.utils.data.distributedrmodelscope.hub.check_modelrmodelscope.metainformodelscope.metricsrr,modelscope.metrics.prediction_saving_wrapperrmodelscope.models.baserr1modelscope.msdatasets.dataset_cls.custom_datasetsr9modelscope.msdatasets.dataset_cls.custom_datasets.builderr modelscope.msdatasets.ms_datasetrmodelscope.outputsrmodelscope.preprocessors.baser!modelscope.trainers.hooks.builderr "modelscope.trainers.hooks.priorityr!r"'modelscope.trainers.lrscheduler.builderr#%modelscope.trainers.optimizer.builderr$modelscope.utils.configr%r&r'modelscope.utils.constantr(r)r*r+r,r-r.r/r0modelscope.utils.data_utilsr1modelscope.utils.devicer2modelscope.utils.file_utilsr3modelscope.utils.import_utilsr4modelscope.utils.loggerr5modelscope.utils.registryr6modelscope.utils.torch_utilsr7r8r9r:r;r<r=r>baser@builderrArrBrCrD hooks.hookrEparallel.builderrFparallel.utilsrGr[register_modulerErJrr~rrrs #*?? %997;B(A46.63EFAKK66621@<.4;;;>>,';<  h&6&67a a 8a H+!r