o *jq,@sddlmZddlZddlmmZddlmZddl m Z ddl m Z ddl mZddlmZddlmZdd lmZmZdd lmZdd lmZdd lmZdd lmZddlmZddl m!Z!ddl"m#Z#ddl$m%Z%gdZ&GdddeZ'ej(ej)dGdddeZ*dS))PathN)DiffusionPipeline)Image)exif_transpose)Dataset) transforms)Trainers)ModelOutputBase OutputKeys)TRAINERS)CheckpointHook)CheckpointProcessor)build_optimizer)EpochBasedTrainer) ConfigDict)ModeKeys)is_dist)za photo of a {}za rendering of a {}za cropped photo of the {}zthe photo of a {}za photo of a clean {}za photo of a dirty {}za dark photo of the {}za photo of my {}za photo of the cool {}za close-up photo of a {}za bright photo of the {}za cropped photo of a {}za photo of the {}za good photo of the {}za photo of one {}za close-up photo of the {}za rendition of the {}za photo of the clean {}za rendition of a {}za photo of a nice {}za good photo of a {}za photo of the nice {}za photo of the small {}za photo of the weird {}za photo of the large {}za photo of a cool {}za photo of a small {}c@s"eZdZddZ  dddZdS)ConesCheckpointProcessorcCs ||_dSN) model_dir)selfrru/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/modelscope/trainers/multi_modal/cones2/cones_trainer.py__init__9s z!ConesCheckpointProcessor.__init__NTcCs`d}d}t|j|j}|j|d|jjddd} | j} | dd} d} d} tD]X}|j| |d|jjddd}|j}t t |d| kdd}|j j|dd d }|dd|||}|j|dd d }|dd|||}| |7} | |7} q)t | | tt|d tj|j|j jd }i}|jj|jjfi||_||d S) z-Save the state dict for Cones model. dog max_lengthTpt)paddingr truncationreturn_tensorsrcudaN)Zattention_maskz/emb.pt) text_encoder)rfrom_pretrainedrtodevice tokenizermodel_max_length input_idsPROMPT_TEMPLETEformatinttorchwheremodelr"savelenZ scheduler from_configconfigZsave_pretrained)rZtrainerZcheckpoint_path_prefix output_dirmetaZsave_optimizersinstance_prompt token_numpipeZtext_inputs_originZtext_inputs_origin_idsindexZprompt_embeds_newZprompt_embeds_origintemplate text_inputsZtext_input_idsZindex_templateZprompt_embeds_nowZ prompt_embedspipelineZscheduler_argsrrrsave_checkpoints<sl        z)ConesCheckpointProcessor.save_checkpoints)NT)__name__ __module__ __qualname__rr<rrrrr7s r) module_namecs8eZdZfddZd dedefddZdd ZZS) ConesDiffusionTrainerc stj|i| |dd|_|dd|_|dd|_|dd|_|d d |_|d d |_|d d|_ t t dd|j d}| t|jtj|jtjddd}||j||j|jdd|jjddj|jd|_dS)Nwith_prior_preservationFr5r class_promptza photo of dogclass_data_dirz/tmp/class_datanum_class_images resolutioniprior_loss_weightg?cSs t|tSr) isinstancer )hookrrrs z0ConesDiffusionTrainer.__init__..r)Z torch_dtypeZsafety_checkerrevisionTrr)rrrr )superrpoprBr5rCrDrErGrHlistfilterhooksZ set_processorrrrr#r,float32r$r%r"r&r'r(detach target_embed)rargskwargsZ ckpt_hookr; __class__rrrwsP   zConesDiffusionTrainer.__init__Ncfg default_argsc CsRz t|jj||dWSty(}z|jd|dtjd|d}~ww)N)rYrZz%Build optimizer error, the optimizer zG is a torch native component, please check if your torch with version: z matches the config.) rr.r" parametersKeyErrorloggererrorr, __version__)rrYrZerrrrs z%ConesDiffusionTrainer.build_optimizercs|d}|jjtj|_|jj|d|j jt j dj }||jjjj}|jj|d|jjjdddd}t |j|j }t |}|jd }t jd |jjj|f|j d } | } |jj||| } |j|d d } |jjjjd kr|} n|jjjjd kr|jj||| } n td |jjjj|j| | | j } dt t |j!ddt | dd d}dt t |j!d|dt | d|dd}||}t"j#| $| $dd}t%j&||i}t'|t(r|)}t'|t*s t+dd|vrbdg}t,g}|D]|-fdd|.Dqi}|D]-}|/|d}|durYt0rP|j12d}t34|5t36|-||7iq-|j8-|n|j8-|d||_9dS)aZ Perform a training step on a batch of inputs. Subclass and override to inject custom behavior. Args: model (`TorchModel`): The model to train. inputs (`Dict[str, Union[torch.Tensor, Any]]`): The inputs and targets of the model. The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument `labels`. Check your model's documentation for all accepted arguments. Return: `torch.Tensor`: The tensor with training loss on this batch. rtarget)ZdtypetextTrr)rrrr r)r%epsilonZ v_predictionzUnknown prediction type g{Gz?NgMbP?mean)Z reductionz$"model.forward()" must return a dictlog_varslosscsg|]}|vr|qSrr).0keyZkey_prr sz4ConesDiffusionTrainer.train_step..r!):trainr.r"rZTRAIN_modeZvaeencoder$r%r,rRZ latent_distsampler2Zscaling_factorr&r'Zsqueezer(Z randn_likeshaperandintZnoise_schedulerZnum_train_timestepslongZ add_noiseZ unsqueezeZprediction_typeZ get_velocity ValueErrorZunetZnormrTFZmse_lossfloatr ZLOSSrIr to_dictdict TypeErrorsetupdatekeysgetrdataclonedistZ all_reduceZdiv_Zget_world_sizeitemZ log_buffer train_outputs)rr.Zinputsr6Zlatentsr:r(noiseZbszZ timestepsZ noisy_latentsZencoder_hidden_statesZ target_priorZmodel_pred_priorZloss_embedding_headZloss_embedding_tailZloss_embeddingrgrZdefault_keys_patternZ match_keysrfrivaluerrjr train_steps           z ConesDiffusionTrainer.train_stepr) r=r>r?rrrwrr __classcell__rrrWrrAts -rA)+pathlibrr,Ztorch.nn.functionalnnZ functionalrtZ diffusersrZPILrZ PIL.ImageOpsrZtorch.utils.datarZ torchvisionrZmodelscope.metainforZmodelscope.outputsr r Zmodelscope.trainers.builderr Z4modelscope.trainers.hooks.checkpoint.checkpoint_hookr Z9modelscope.trainers.hooks.checkpoint.checkpoint_processorr Z%modelscope.trainers.optimizer.builderrZmodelscope.trainers.trainerrZmodelscope.utils.configrZmodelscope.utils.constantrZmodelscope.utils.torch_utilsrr)rZregister_moduleZcones2_inferencerArrrrs,                =