o –hØã@s¸ddlmZmZddlZddlZddlmZmZmZm Z ddl Z ddl m Z ddl mZddlmZddlmZddlmZddlmZddlmZeƒZe d ¡ZGd d „d eƒZdS) é)ÚABCÚabstractmethodN)ÚListÚTupleÚAnyÚMapping)Údefault_device)Úutils)ÚDEFAULT_BATCH_SIZE)Úevaluate_model)Úget_tqdmzstanza.lemmaclassifierc @sLeZdZdedefdd„Zedd„ƒZdeded ed ed ed d f dd„Z d S)ÚBaseLemmaClassifierTrainerÚ label_decoderÚcountscCsndd„| ¡Dƒ}t| ¡ƒ}|D]}|||t|ƒ||<qt |¡}t d|›d¡tj |d|_ dS)an If applicable, this function will update the loss function of the LemmaClassifierLSTM model to become BCEWithLogitsLoss. The weights are determined by the counts of the classes in the dataset. The weights are inversely proportional to the frequency of the class in the set. E.g. classes with lower frequency will have higher weight. cSsg|]}d‘qS)r©)Ú.0Ú_rrúf/var/www/html/env_mimamsha/lib/python3.10/site-packages/stanza/models/lemma_classifier/base_trainer.pyÚ szFBaseLemmaClassifierTrainer.configure_weighted_loss..zUsing weights z for weighted loss.)ÚweightN) ÚkeysÚsumÚvaluesÚlenÚtorchÚtensorÚloggerÚinfoÚnnÚBCEWithLogitsLossÚ criterion)ÚselfrrÚweightsÚ total_samplesÚ class_idxrrrÚconfigure_weighted_losss  z2BaseLemmaClassifierTrainer.configure_weighted_losscCsdS)z` Build a model using pieces of the dataset to determine some of the model shape Nr)r!rÚ upos_to_idÚ known_wordsÚ target_wordsÚ target_uposrrrÚ build_model#sz&BaseLemmaClassifierTrainer.build_modelÚ num_epochsÚ save_nameÚargsÚ eval_fileÚ train_fileÚreturnNc CsÆtƒ}|s tdƒ‚tj||j| dt¡d}|j}|j} t |ƒ|_ t   d|›¡t   d|›d|j ›¡t   d|j ›¡| || |j|j t|jƒ¡|_tj|j ¡|jd|_|j |¡t   d |›d t|j ¡ƒj›¡tj |¡r„| d d ¡s„td |›dƒ‚|jrŽ| ||j ¡t  !dt|j ¡ƒj›¡|j" t|j ¡ƒj¡|_"dt#dƒ} } t$|ƒD]¬} t%|ƒD]_\} }}}t | ƒt |ƒkrÐt |ƒksânJdt | ƒt |ƒt |ƒf›dƒ‚|j &¡| || |¡}|jrt' (dd„|Dƒ¡jt'j)d |¡}n| |¡}| "||¡}| *¡|j +¡qºt   d| d›d|›d| ,¡›¡|rZt-|j|dd\}}}}t   d|›¡|| krY|} |j .|¡t   d|›d¡q´|j .|¡q´dS) a Trains a model on batches of texts, position indices of the target token, and labels (lemma annotation) for the target token. Args: num_epochs (int): Number of training epochs save_name (str): Path to file where trained model should be saved. eval_file (str): Path to the dev set file for evaluating model checkpoints each epoch. train_file (str): Path to data file, containing tokenized text sentences, token index and true label for token lemma on each line. z,Cannot train model - no train_file supplied!Ú batch_size)Ú get_countsr1z!Loaded dataset successfully from zUsing label decoder: z Output dimension: zTarget words: )ÚlrzTraining model on device: z. ÚforceFz Save name zM already exists; training would overwrite previous file contents. Aborting...z Criterion on Nz-infz>Input sentences, positions, and labels are of unequal length (ú)cSs0g|]}|dkrt ddg¡nt ddg¡‘qS)ré)rr)rÚlabelrrrr]s0z4BaseLemmaClassifierTrainer.train..)ÚdtypezEpoch [r6ú/z ], Loss: T)Ú is_trainingzWeighted f1 for model: z%New best model: weighted f1 score of Ú.)/rÚ ValueErrorr ÚDatasetÚ weighted_lossÚgetr rr&rÚ output_dimrrr(r*r'Úsetr)ÚmodelÚoptimÚAdamÚ parametersr3Ú optimizerÚtoÚnextÚdeviceÚosÚpathÚexistsÚFileExistsErrorr%rÚdebugr ÚfloatÚrangeÚtqdmÚ zero_gradrÚstackÚfloat32ÚbackwardÚstepÚitemr Úsave)r!r+r,r-r.r/rIÚdatasetrr&Ú best_modelÚbest_f1ÚepochÚ sentencesÚ positionsÚ upos_tagsÚlabelsÚoutputsÚtargetsÚlossrÚf1rrrÚtrain)sV   " D &   $  €áz BaseLemmaClassifierTrainer.train) Ú__name__Ú __module__Ú __qualname__rr%rr*ÚintÚstrrerrrrr s  &r )ÚabcrrÚloggingrJÚtypingrrrrrÚtorch.nnrÚ torch.optimrCÚstanza.models.common.utilsrÚstanza.models.lemma_classifierr Ú(stanza.models.lemma_classifier.constantsr Ú.stanza.models.lemma_classifier.evaluate_modelsr Ústanza.utils.get_tqdmr rQÚ getLoggerrr rrrrÚs