o ¢Äiϳã @s ddlZddlmZmZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl Z ddlZddlZddlmZddlZddlmZddlZddlmZmZmZddlmZddlmZGdd „d ejƒZd d „Zed kr e  ¡Z!e!j"d de#dde!j"dddddde!j"dddddde!j"dddddde!j"dddddde! $¡Z%e &e'e%j(dƒ )¡ej*¡Z+ej, -dej, .e+d¡¡ddl/m/Z/ej, .e+d ¡e+d <ej, 0e+d ¡sêe 1e+d ¡ej, 0ej, 2e+d e+d!¡¡se 1ej, 2e+d e+d!¡¡ej, 2e+d e+d!d"¡Z3ej, 2e+d e+d!d#¡Z4ej, 2e+d e+d!d$¡Z5ej, 2e+d e+d!d%¡Z6ej, 2e+d e+d!¡Z7d&d'„e+d(DƒZ8gZ9e:e+d)ƒe:e+d*ƒkrfd+ge:e+d*ƒe+d)<e;e+d*e+d)ƒD]\Zd,d'„e ?e<¡Dƒe=¡qoe%j@dure Ad-¡ej, 0e3¡s›e 1e3¡e:e Be3¡ƒZCeCd.e+d/krÓe/e+d0e+d/eCe+d1d2gd2ggd3¢e3dd4d'„eDe+d/ƒDƒd5 ejE F¡n e Gd6e+d/›d7¡e Ad8¡ej, 0e4¡sïe 1e4¡e:e Be4¡ƒZCeCd.e+d9kre/e+d0e+d9eCe+d1d:gd:ggd3¢e4dd;ejE F¡n e GdZHe+d0D]ZIeH >eeIe+d/e:e+d0ƒd:d?d@¡qQe/eHe+d/eCe+d1dAd2gd2ggd3¢e5ddBd'„eDe+d/ƒDƒd5 ejE F¡n e GdCe+d/›d7¡e AdD¡ej, 0e6¡s¬e 1e6¡e:e Be6¡ƒZCeCd.e+d9krùe+d>ZHe+d0D]ZIeH >eeIe+d9e:e+d0ƒd:d?d@¡qÄe/eHe+d9eCe+d1dAd:gd:ggd3¢e6dd;ejE F¡n e GdEe+d9›d7¡dFZJdGd'„ee4ƒ KdH¡DƒZLgZMeDeJƒD]ZNe jOjP )eLe jQ Rde:eLƒ¡¡\ZSZTeM Ue:eTƒ¡qeVeWe  XeM¡dIƒdIƒdJe+dK<e+dKdLkrRdLe+dK<neYe+dKdLƒdMkradLe+dK<e%jdur ej, 0ej, 2e7dN¡¡rye%jZdur›dOd'„ee3ƒ KdH¡Dƒe+dPZ[ee[e+dKe+dQe9e8dRZ\dSd'„ee4ƒ KdH¡Dƒe+dPZ]ee]e+dKe+dQe9e8dRZ^dTd'„ee5ƒ KdH¡Dƒe+dPZ_ee_e+dKe+dQe9e8dRZ`dUd'„ee6ƒ KdH¡Dƒe+dPZaeeae+dKe+dQe9e8dRZbe AdV¡e c¡Zdeddurþd+ZdneddWZdee\e:e Be3¡ƒe+dKej, 2e7dN¡ejE e¡rdXndYejE e¡s$ednd+dZee`e:e Be5¡ƒe+dKej, 2e7d[¡ejE e¡rAdXndYejE e¡sJednd+dZee^e:e Be4¡ƒe+dKej, 2e7d\¡ejE e¡rgdXndYejE e¡spednd+dZeebe:e Be6¡ƒe+dKej, 2e7d]¡ejE e¡rdXndYejE e¡s–ednd+dZne Gd^¡e%jfdur ejgjdYd_Zheh ie+dKd`¡Zjed+eje+dae+dbdcejdd`ddZkddfdg„Zldhdi„e+dj m¡DƒZniZodkgepe+dj m¡ƒdlgD]Zqeqdkkrødmdn„eoeq<qédodn„eoeq<qéej, 2e7dN¡e+djdk<ej, 2e7d[¡e+djdl<ee+dje+dpeneodqZrGdrds„dsejgjsjtƒZue c¡Zdeddur9d+ZdneddWZdejgjsjveuerƒdeddedtZwe  &e+du¡Zxe  ydvd'„eDdexjzdejdd+ƒDƒ¡Zxe  {exjzd¡ |e j}¡Z~ejgjsjvejgjs e €ex¡e €e~¡¡e:e~ƒdwZxe  &ej, 2e7d\¡¡Ze  &ej, 2e7d]¡¡Z‚e  ƒe  „ejzd¡e  {e‚jzd¡f¡ |e j}¡Z…ejgjsjvejgjs e €e  †ee‚f¡¡e €e…¡¡e:e…ƒdwZ‡ekjˆewe‡exe+dxe+dye+dzd{Z‰ekjŠe‰e+d!e+d d|eej, 2e+d e+d!d}¡ej, 2e+d e+d!d~¡ƒdSdSdS)€éN)ÚoptimÚnn)Útqdm)ÚPath)Úgenerate_adversarial_textsÚ augment_clipsÚmmap_batch_generator)Úcompute_features_from_generator)Ú AudioFeaturescs²eZdZ  d+‡fdd„ Zdd „Zd,d d „Z  d-dd„Zdd„Zdd„Zd.dd„Z d/dd„Z  d0dd „Z d.d!d"„Z d.d#d$„Z d%d&„Zddddgd'gd(dfd)d*„Z‡ZS)1ÚModelé©éé`Údnné€Nc srtƒ ¡||_||_||_t tj ¡rdnd¡|_g|_ g|_ d|_ d|_ d|_ d|_|dkrOGdd„dtjƒ‰G‡fdd „d tjƒ}|||||d |_n|d krbGd d „d tjƒ}|||ƒ|_|d krzdd„|_tjdd|_tjdd|_nddd„}ddd„} ddd„} ||_| |_| |_d|_d|_t t¡|_|d kr§tjjj ntjj!|_"t#j$|j %¡dd|_&dS)Nzcuda:0Úcpuéèrrcs$eZdZ‡fdd„Zdd„Z‡ZS)z Model.__init__..FCNBlockcs2tƒ ¡t ||¡|_t ¡|_t |¡|_dS©N) ÚsuperÚ__init__rÚLinearÚ fcn_layerÚReLUÚreluÚ LayerNormÚ layer_norm)ÚselfÚ layer_dim©Ú __class__©úL/home/kim/smarthome/.venv/lib/python3.10/site-packages/openwakeword/train.pyr9s  z)Model.__init__..FCNBlock.__init__cSs| | | |¡¡¡Sr)rrr©rÚxr!r!r"Úforward?sz(Model.__init__..FCNBlock.forward©Ú__name__Ú __module__Ú __qualname__rr%Ú __classcell__r!r!rr"ÚFCNBlock8s r+cs(eZdZd‡‡fdd„ Zdd„Z‡ZS)úModel.__init__..Netr cs–tƒ ¡t ¡|_t |d|dˆ¡|_t ¡|_t  ˆ¡|_ t  ‡‡fdd„t |ƒDƒ¡|_ t ˆ|¡|_|dkrDt ¡|_dSt ¡|_dS)Nrr csg|]}ˆˆƒ‘qSr!r!©Ú.0Úi)r+rr!r"Ú Ióz8Model.__init__..Net.__init__..)rrrZFlattenÚflattenrÚlayer1rÚrelu1rÚ layernorm1Z ModuleListÚrangeÚblocksÚ last_layerÚSigmoidÚlast_act)rÚ input_shaperÚn_blocksÚ n_classes)r+r )rr"rCs     $ú$Model.__init__..Net.__init__c SsD| | | | |¡¡¡¡}|jD]}||ƒ}q| | |¡¡}|Sr)r4r5r3r2r7r:r8)rr$Úblockr!r!r"r%Ms   ú#Model.__init__..Net.forward)r r r&r!©r+rr"ÚNetBs rB)r<r=Zrnncs&eZdZd‡fdd„ Zdd„Z‡ZS)r,r csXtƒ ¡tj|ddddddd|_t d|¡|_|dkr%t ¡|_ dSt ¡|_ dS) Néÿÿÿÿé@éTç)Z num_layersÚ bidirectionalZ batch_firstZdropoutrr ) rrrZLSTMr3rÚlayer2r9rÚlayer3)rr;r=rr!r"rVs ÿ$r>cSs*| |¡\}}| | |dd…df¡¡S)NrC)r3rIrH)rr$ÚoutÚhr!r!r"r%]sr@©r r&r!r!rr"rBUsr cSs||dk ¡S)Ngà¿)Úsum)ÚpredÚyr!r!r"Údóz Model.__init__..Úbinary)Útaskçà?cSsNtjjj|dd}|dk}||jddd||jddd|k@k ¡}|S©Nr ©Údimr©Zaxis©ÚtorchrÚ functionalÚsoftmaxÚargmaxÚmaxrM)ÚprOÚ thresholdÚprobsZneg_ndcsÚfpr!r!r"Ú multiclass_fphs0z%Model.__init__..multiclass_fpcSsVtjjj|dd}|dk}||jddd||jddd|k@k ¡| ¡}|SrUrY)r_rOÚnegative_class_labelr`raÚpos_ndcsZrcllr!r!r"Úpositive_class_recallnsÿ ÿz-Model.__init__..positive_class_recallcSsJtjjj|dd}|jdd|k}||jdd||k ¡| ¡}|S)Nr rVrX)rZrr[r\r]rM)r_rOrdraZ pos_predsÚaccr!r!r"Úpositive_class_accuracyus$z/Model.__init__..positive_class_accuracyç-Cëâ6?)Úlr)rT)rrT©r)'rrr=r;Úseconds_per_examplerZÚdeviceÚcudaÚ is_availableÚ best_modelsÚbest_model_scoresÚ best_val_fpÚbest_val_accuracyÚbest_val_recallZbest_train_recallrÚModuleÚmodelrbÚ torchmetricsZRecallÚrecallZAccuracyÚaccuracyrgÚn_fpÚval_fpÚ collectionsÚ defaultdictÚlistÚhistoryr[Zbinary_cross_entropyZ cross_entropyÚlossrZAdamÚ parametersÚ optimizer) rr=r;Ú model_typerr<rlrBrcrfrhrrAr"rsB         zModel.__init__cCs |jdkrt |j|¡dSdS)z> Saves the weights of a trained Pytorch model r N)r=rZÚsaverv)rÚ output_pathr!r!r"Ú save_model‰s ÿzModel.save_modelÚcsˆ|‰|jdkrtjj|j d¡t |j¡d||gddS|jdkrBG‡fdd„dtj ƒ}tjj|ƒt |j¡d||gddSdS)Nr rr)Z output_namescs&eZdZ‡‡fdd„Zdd„Z‡ZS)zModel.export_to_onnx..Mcstƒ ¡ˆj d¡|_dS)Nr)rrrvÚto©r)r Úobjr!r"ršs z(Model.export_to_onnx..M.__init__cSstjjj| |¡ddS)Nr rV)rZrr[r\rvr#r!r!r"r% óz'Model.export_to_onnx..M.forwardr&r!©rŠrr"ÚM™sr) r=rZÚonnxÚexportrvrˆÚrandr;rru)rr…Z class_mappingrr!rŒr"Úexport_to_onnxs   ÿ   ÿôzModel.export_to_onnxrrFçü©ñÒMbP?c Cstd|dt tj|||t|||ƒ¡}|||}|dkr/t |||k||¡}t ||k||¡}|S)NrTr r)ÚnpÚcosÚpiÚfloatÚwhere) rZ global_stepÚ warmup_stepsÚholdÚ total_stepsZstart_lrÚ target_lrZ learning_rateZ warmup_lrr!r!r"Úlr_warmup_cosine_decay§s  ÿ ÿzModel.lr_warmup_cosine_decaycCs | |¡Sr©rvr#r!r!r"r%Àó z Model.forwardcCstj|jd|jddS)NrLr)Z input_sizerm)Ú torchinfoÚsummaryrvr;r‰r!r!r"r Ãr‹z Model.summarycCsž|dur|j}t |d¡}| ¡}|D] }||d9<q|D]}| ¡}| ¡D] \}}|||7<q+q!|D] }||t|ƒ<q;| |¡|S)zHAverages the weights of the provided models together to make a new modelNr)rpÚcopyÚdeepcopyZ state_dictÚitemsÚlenZload_state_dict)rÚmodelsZaveraged_modelZaveraged_model_dictÚkeyrvZ model_dictÚvaluer!r!r"Úaverage_modelsÆsÿ zModel.average_modelsçš™™™™™&@rTçš™™™™™É?cs~dgtˆjƒ}|D]Q}|d ˆj¡|d ˆj¡}}ttˆjƒtˆjƒddD]0\} } t ¡ | |ƒ} || ˆ | |d¡  ¡  ¡  ¡|| <Wdƒn1sUwYq*q ‡fdd„|Dƒ}‡fd d„t|ƒDƒ} ‡fd d„| Dƒ} t | ƒ|krŠt  d |›d ¡dSˆj| t | ¡}ˆj| t | ¡d }t  d|›dt| ƒ›dt  | ¡›d|| t | ¡›¡|S)a? Select the top model based on the false positive rate on the validation data Args: false_positive_validate_data (torch.DataLoader): A dataloader with validation data n (int): The number of models to select Returns: list: A list of the top n models rr z,Find best checkpoints by false positive rate©ÚtotalÚdesc©.NNcsg|]}|ˆ‘qSr!r!)r.rb)Ú val_set_hrsr!r"r0õr1z,Model._select_best_model..csg|] \}}|ˆkr|‘qSr!r!)r.Úndxrb)Úmax_fp_per_hourr!r"r0÷scsg|] }ˆj|d‘qS)Ú val_recall)rq)r.r°r‰r!r"r0øózNo models with recall >= z found!Útraining_step_ndxzBest model from training step z out of zmodels has recall of z and false positive rate of )r¤rprˆrmrÚ enumeraterZÚno_gradrbÚdetachrÚnumpyr^ÚloggingÚwarningr“r]rqÚinfo)rZfalse_positive_validate_datar¯r±Z min_recallZfalse_positive_ratesÚbatchÚx_valÚy_valZmdl_ndxrvÚval_psZcandidate_model_ndxZcandidate_model_recallÚ best_modelZbest_model_training_stepr!)r±rr¯r"Ú_select_best_modelás4 "ÿ *þ€þ ÿþzModel._select_best_modeléPÃrc CsNd}t d¡d}t d|t|ƒ¡ ¡} t |t|dƒ|d¡ tj¡} |j||||| | |d|d||d t d ¡|d }|d }|j |krS|d }t d ¡t d|t|ƒ¡ ¡} t d|d¡ tj ¡} |j||||| | |d|d||d t d¡|d }|j |kr“|d }t d ¡t d|t|ƒ¡ ¡} t d|d¡ tj ¡} |j||||| | |d|d||d t d¡t  |j dd¡} t  |j dd¡} t  |j dd ¡} g}t |j|jƒD]\}}|d| kr|d| kr|d| kr| |¡qåt|ƒdkr|j|d}n|j}t ¡t|D]}|d |j¡|d |j¡}}||ƒ}q| ||d¡ ¡ ¡ ¡}| ||d tj¡¡ ¡ ¡ ¡}d}|D]"}|d |j¡|d |j¡}}||ƒ}|| ||d¡7}qY|| ¡ ¡ ¡}Wdƒn 1s‘wYt d|›d|›d|›d¡|S)zìA sequence of training steps that produce relatively strong models automatically, based on validation data and performance targets provided. After training merges the best checkpoints and returns a single model. r©z…################################################## Starting training sequence 1... ##################################################rir gÐ?ééé) ÚXÚX_valÚfalse_positive_val_dataÚ max_stepsÚnegative_weight_scheduleÚ val_stepsr˜Ú hold_stepsrjr¯z…################################################## Starting training sequence 2... ##################################################é rEzCIncreasing weight on negative examples to reduce false positives...z…################################################## Starting training sequence 3... ##################################################zBMerging checkpoints above the 90th percentile into single model...Ú val_accuracyéZr²Ú val_fp_per_hrr)r¥r®Nz( ################ Final Model Accuracy: z Final Model Recall: z' Final Model False Positives per Hour: z ################ )r¹r»r“ZlinspaceÚintÚtolistÚastypeÚint64Ú train_modelrrZint16Ú percentilerÚziprprqÚappendr¤r¨rvrZr¶rˆrmrxr·rr¸ryrb)rÚX_trainrÇrÈÚstepsÚmax_negative_weightÚtarget_fp_per_hourr¯rjÚweightsrËZaccuracy_percentileZrecall_percentileZ fp_percentiler¥rvZscoreZcombined_modelr¼r$rOr¿Zcombined_model_recallZcombined_model_accuracyZcombined_model_fpr½r¾Zcombined_model_fp_per_hrr!r!r"Ú auto_trains  " ù    ù    ù  € " $"ò ÿ ÿzModel.auto_trainc CsÌt|ƒdkr |d}| |j¡}g}t|ddD]H}|d}g}td|jdddƒD]}| |dd…||d…dd…f¡q+t |¡}|durO|  |¡}n||ƒ}| |  ¡  ¡  ¡d¡qt  |¡S) aÑ Predict on Tensors of openWakeWord features corresponding to single audio clips Args: features (torch.Tensor): A Tensor of openWakeWord features with shape (batch, features) model (torch.nn.Module): A Pytorch model to use for prediction (default None, which will use self.model) Returns: torch.Tensor: An array of predictions of shape (batch, prediction), where 0 is negative and 1 is positive rÅrzPredicting on clips)r­rr rN)r¤rˆrmrr6ÚshaperØrZÚvstackrvr·rr¸r“)rÚfeaturesrvÚ predictionsr$r¼r/Úpredsr!r!r"Úpredict_on_featuresps  &   zModel.predict_on_featurescCs2tddd}|j|dd}|jt |¡|d}|S)a£ Predict on Tensors of 16-bit 16 khz audio data Args: clips (np.ndarray): A Numpy array of audio clips with shape (batch, samples) model (torch.nn.Module): A Pytorch model to use for prediction (default None, which will use self.model) Returns: np.ndarray: An array of predictions of shape (batch, prediction), where 0 is negative and 1 is positive ré)rmÚncpur©Ú batch_sizer)r Z embed_clipsrärZÚ from_numpy)rZclipsrvÚFrárãr!r!r"Úpredict_on_clipsŽs zModel.predict_on_clipscCst|jdkr tdƒ‚t dtj ||d¡›d¡t |¡}t j j |  d¡t   |j¡dtj ||d¡dd d S) zDSaves the trained openwakeword model to both onnx and tflite formatsr zˆExporting models to both onnx and tflite with more than one class is currently not supported! Use the `export_to_onnx` function instead.z#### Saving ONNX mode as 'ú.onnxú'rré )Z opset_versionN)r=Ú ValueErrorr¹r»ÚosÚpathÚjoinr¡r¢rZrŽrrˆrr;)rrvÚ model_nameÚ output_dirZ model_to_saver!r!r"Ú export_model£s   ÿzModel.export_modeléúric * CsÐ| |j¡|j |j¡d} d} t g¡ |j¡}t g¡ |j¡}tt|dƒ|ddD]7\}}|d |j¡|d |j¡}}|d tj¡}|jj D]}|j ||||| d|d<qO|j  ¡| |¡}||dk|  ¡dk@}||dk|  ¡d k@}t  ||dk|  ¡dk@||dk|  ¡d k@f¡}|d tj¡}t  ||f¡}t|ƒdkrÉt |jd¡|d}|dk}d||<|d}n|jdkræt |jd¡||}|dk}d||<|d}|jddkrƒ| ||jdkrø|n|| |j¡¡}|| }| |jd7} |jdd kr|}|}| d kr/| d7} t  ||f¡}t  ||f¡}nT| ¡|j ¡d} d} |jd  | ¡ ¡ ¡¡| ||jdkrU|n|¡}|j|7_|jd  | ||¡ ¡ ¡ ¡¡t g¡ |j¡}t g¡ |j¡}|| vrå|dkrå|duråd}t|ƒD]:\}}t ¡(|d |j¡|d |j¡}} | |¡}!|| |!| d¡7}Wdƒn 1sÌwYq˜||  ¡ ¡ ¡}"|jd  |"¡|| vrh|dkrh|durhd}#d}$t|ƒD]_\}}t ¡M|d |j¡}g}%td|jdddƒD]}&|% |dd…|&|&d…dd…f¡qt |%¡}%| |%¡}'t |'dkƒrF|#d7}#n|$d7}$Wdƒn 1sUwYqü|jd |#|#|$¡|| vrò|dkrò|duròt|ƒD]R\}}t ¡@|d |j¡|d |j¡}} | |¡}!| |!| d¡ ¡ ¡ ¡}(| !|!| d tj"¡¡})| |!| d¡}Wdƒn 1sÇwYq{|jd |) ¡ ¡ ¡¡|jd |(¡|jd | ¡ ¡ ¡¡|| vr[|dkr[|jddt# $|jdd¡kr[|jddt# $|jdd¡kr[|j% t& '|j¡¡|j( ||jdd|jdd|jdd|j )d dg¡ddœ¡|jdd|_*|jdd|_+||dkredSq-dS)Nr rZTrainingr«r®)r˜r™ršr›rjr’g+‡ÙÎ÷ï?rr€rxrÐrrTZpositive_test_clips_recallrÎr²Úval_n_fprCé2rÄ)r´r÷r²rÎrÐ),rˆrmrvrZZTensorrrµÚfloat32r‚Z param_groupsrœZ zero_gradZsqueezeÚcatr¤Úonesrßr=r€ZbackwardÚsteprrØr·rr¸rbrzrxr¶r6ràÚanyryrÔr“rÖrpr¡r¢rqÚgetrtrs)*rrÆrÉr˜rÌrÇrÈZpositive_test_clipsrÊrËrjr¯Zaccumulation_stepsZaccumulated_samplesZaccumulated_predictionsZaccumulated_labelsZstep_ndxÚdatar$rOZy_ÚgrâZ neg_high_lossZ pos_high_lossÚwrer€rbr{Z val_step_ndxr½r¾Zval_predictionsrÐÚtpÚfnr¼r/rãr²Zval_accr!r!r"rÕ²sØ "   ÿ  6   $  $ " ý€ (   €ö€  " û€""  ýÿ†zModel.train_model)r r rrr N)r‡)rrrrFr’r)r©rTrª)rÂrrª)r'r(r)rr†r‘rœr%r r¨rÁrÞrärërõrÕr*r!r!rr"r s6ÿo  ú  $ ÿ k ýr c Csêddl}ddlm}ddl}| |¡}||dd}t ¡N}| tj   |d¡¡|j j   tj   |d¡¡}| ¡} t d|›d¡t|d ƒ } |  | ¡Wdƒn1sVwYWdƒdSWdƒdS1snwYdS) zRConverts an ONNX version of an openwakeword model to the Tensorflow tflite format.rN)ÚprepareZCPU©rmZtf_modelz#### Saving tflite mode to 'ríÚwb)rŽZonnx_tf.backendrZ tensorflowÚloadÚtempfileÚTemporaryDirectoryZ export_graphrðrñròZliteZTFLiteConverterZfrom_saved_modelÚconvertr¹r»ÚopenÚwrite) Zonnx_model_pathr…rŽrÚtfZ onnx_modelZtf_repZtmp_dirÚ converterZ tflite_modelÚfr!r!r"Úconvert_onnx_to_tflite>s(      ÿ ú ý ú ÷ rÚ__main__z--training_configz/The path to the training config file (required)T)ÚhelpÚtypeÚrequiredz--generate_clipsz-Execute the synthetic data generation processÚ store_trueÚFalseF)rÚactionÚdefaultrz--augment_clipsz/Execute the synthetic data augmentation processz --overwritezNOverwrite existing openwakeword features when the --augment_clips flag is usedz --train_modelz"Execute the model training processÚrZpiper_sample_generator_path)Úgenerate_samplesrôróZpositive_trainZ positive_testZnegative_trainZ negative_testcCs"g|] }t |¡D]}|j‘q qSr!)rðÚscandirrñ)r.Újr/r!r!r"r0s"r0Ú rir_pathsZ!background_paths_duplication_rateÚbackground_pathsr cCsg|]}|j‘qSr!)rñr-r!r!r"r0”ózŒ################################################## Generating positive clips for training ##################################################gffffffî?Z n_samplesÚ target_phraseZtts_batch_sizeg\Âõ(\ï?)gè?çð?gô?cCóg|] }t ¡jd‘qS©z.wav©ÚuuidÚuuid4Úhexr-r!r!r"r0¢r³) ÚtextÚ max_samplesrèÚ noise_scalesÚnoise_scale_wsÚ length_scalesrôÚauto_reduce_batch_sizeZ file_namesz8Skipping generation of positive clips for training, as ~z already existz‹################################################## Generating positive clips for testing ##################################################Z n_samples_valr!)r(r)rèr*r+r,rôr-z3Skipping generation of positive clips testing, as ~zŒ################################################## Generating negative clips for training ##################################################Zcustom_negative_phrasesrª)Z input_textÚNZinclude_partial_phraseZinclude_input_wordsécCr"r#r$r-r!r!r"r0Çr³z8Skipping generation of negative clips for training, as ~z‹################################################## Generating negative clips for testing ##################################################z7Skipping generation of negative clips for testing, as ~røcCóg|]}t|ƒ‘qSr!©Ústrr-r!r!r"r0år1z*.wavrià.Ú total_lengthi}i zpositive_features_train.npycCr0r!r1r-r!r!r"r0ôr1Zaugmentation_roundsZaugmentation_batch_size)r3rèZbackground_clip_pathsZ RIR_pathscCr0r!r1r-r!r!r"r0úr1cCr0r!r1r-r!r!r"r0r1cCr0r!r1r-r!r!r"r0r1z›################################################## Computing openwakeword features for generated samples ##################################################rEZgpur)Zn_totalZ clip_durationÚ output_filermræznegative_features_train.npyzpositive_features_test.npyznegative_features_test.npyzVOpenwakeword features already exist, skipping data augmentation and feature generationri€>rƒZ layer_sizei)r=r;rƒrrlrcsZˆˆjdksˆˆjdkr+t ˆ¡‰t ‡‡fdd„tdˆjdˆˆƒDƒ¡}|SˆS)zbSimple transformation function to ensure negative data is the appropriate shape for the model sizer cs$g|]}ˆ||ˆ…dd…f‘qSrr!r-©Únr$r!r"r0:s$zf..r)rßr“ràÚarrayr6)r$r6Z new_batchr!r5r"r6s  ,ÿrcCsi|]}|t“qSr!)r)r.r¦r!r!r"Ú @rr8Zfeature_data_filesZpositiveZadversarial_negativecCódd„|DƒS)NcSóg|]}d‘qSrLr!r-r!r!r"r0DrQú..r!©r$r!r!r"rPDórPcCr9)NcSr:rkr!r-r!r!r"r0FrQr;r!r<r!r!r"rPFr=Zbatch_n_per_class)Z n_per_classZdata_transform_funcsZlabel_transform_funcsc@seZdZdd„Zdd„ZdS)Ú IterDatasetcCs ||_dSr©Ú generator)rr@r!r!r"rUržzIterDataset.__init__cCs|jSrr?r‰r!r!r"Ú__iter__XszIterDataset.__iter__N)r'r(r)rrAr!r!r!r"r>Ts r>)rèZ num_workersZprefetch_factorZ#false_positive_validation_data_pathcCs g|] }t||td…‘qSrk)ÚX_val_fpr;r-r!r!r"r0ds rçrÚrÛZtarget_false_positives_per_hour)rÙrÇrÈrÚrÛrÜ)rvrórôrìz.tflite)r)‹rZrrrŸrwr¡rðÚsysrr%r¸r“Zscipyr|Úargparser¹rÚyamlÚpathlibrZ openwakewordZopenwakeword.datarrrZopenwakeword.utilsr r rur rr'ÚArgumentParserÚparserÚ add_argumentr2Ú parse_argsÚargsrr Ztraining_configÚreadÚLoaderÚconfigrñÚinsertÚabspathrÚexistsÚmkdirròZpositive_train_output_dirZpositive_test_output_dirZnegative_train_output_dirZnegative_test_output_dirZfeature_save_dirrrr¤r×Zbackground_pathZduplication_rateÚextendrZgenerate_clipsr»ÚlistdirZn_current_samplesr6rnZ empty_cacherºZadversarial_textsr r6ÚglobZpositive_clipsZduration_in_samplesr/ÚioZwavfileÚrandomÚrandintÚsrZdatrØrÑÚroundZmedianÚabsÚ overwriteZpositive_clips_trainZpositive_clips_train_generatorZpositive_clips_testZpositive_clips_test_generatorZnegative_clips_trainZnegative_clips_train_generatorZnegative_clips_testZnegative_clips_test_generatorÚ cpu_countZn_cpusrorÕÚutilsrêZget_embedding_shaper;ZowwrÚkeysZdata_transformsZlabel_transformsr~r¦Zbatch_generatorrÿZIterableDatasetr>Z DataLoaderrÙrBr7rßZzerosrÓrùZX_val_fp_labelsZ TensorDatasetréZ X_val_posZ X_val_negZhstackrûÚlabelsràrÇrÞrÀrõr!r!r!r"Ús2    ) üûûûû  "   û   ý     ü ü     ü ý  $"  $  ý  ý  ý  ý   ü ü ü ü   ÿ    ü ÿ*þ.þû ú ÿÍ[