o –hlã@s–dZddlZddlZddlZddlmZddlmm m Z e  d¡Z dd„Zddd„ZGd d „d ejƒZGd d „d ejƒZGd d„dejƒZdS)z Different loss functions. éNÚstanzacCs"t |¡}d|tj<t |¡}|S©Nr)ÚtorchÚonesÚconstantÚPAD_IDÚnnÚNLLLoss)Ú vocab_sizeÚweightÚcrit©r úT/var/www/html/env_mimamsha/lib/python3.10/site-packages/stanza/models/common/loss.pyÚ SequenceLosss   rFcCs„t|tƒr t |¡}tj|dd\}}|tt |¡ƒ}t |¡|}|r,dt |¡}t  d  |¡¡t j t  |¡ d¡d}|S)zË Either return a loss function which reweights all examples so the classes have the same effective weight, or dampened reweighting using log() so that the biggest class has some priority T)Ú return_countsézReweighting cross entropy by {}ztorch.FloatTensor)r )Ú isinstanceÚlistÚnpÚarrayÚuniqueÚfloatÚsumÚlogÚloggerÚdebugÚformatrÚCrossEntropyLossrÚ from_numpyÚtype)ÚlabelsÚ log_dampenedÚ all_labelsÚ_ÚweightsÚlossr r rÚweighted_cross_entropy_losss  ÿr&cs*eZdZdZd‡fdd„ Zdd„Z‡ZS) Ú FocalLosszø Uses the model's assessment of how likely the correct answer is to weight the loss for a each error multi-category focal loss, in other words from "Focal Loss for Dense Object Detection" https://arxiv.org/abs/1708.02002 Úmeanç@cs<tƒ ¡|dvrtd|ƒ‚||_tjdd|_||_dS)N)rÚnoner(zUnknown reduction: %sr*)Ú reduction)ÚsuperÚ__init__Ú ValueErrorr+rrÚce_lossÚgamma)Úselfr+r0©Ú __class__r rr-2s   zFocalLoss.__init__cCs*t|jƒdkr#t|jƒdkr#|jd|jdkr"td |j|j¡ƒ‚nt|jƒdkr5t|jƒdkr5tdƒ‚td |j|j¡ƒ‚| ||¡}t|jƒdkrV|jd|jdksXJ‚|dt | ¡|j}t|jƒdkrv|jd|jdksxJ‚|j dkr|  ¡S|j dkrŠ|  ¡S|j dkr‘|St d ƒ‚) z€ Weight the loss using the models assessment of the correct answer inputs: [N, C] targets: [N] érrz4Expected inputs N,C and targets N, but got {} and {}zIThis would be a reasonable thing to implement, but we haven't done it yetrr(r*z)unknown reduction! how did this happen??) ÚlenÚshaper.rÚNotImplementedErrorr/rÚexpr0r+rr(ÚAssertionError)r1ÚinputsÚtargetsÚraw_lossÚ final_lossr r rÚforward;s$ÿ &&   zFocalLoss.forward)r(r)©Ú__name__Ú __module__Ú __qualname__Ú__doc__r-r>Ú __classcell__r r r2rr''s  r'có(eZdZdZ‡fdd„Zdd„Z‡ZS)ÚMixLosszb A mixture of SequenceLoss and CrossEntropyLoss. Loss = SequenceLoss + alpha * CELoss cs4tƒ ¡t|ƒ|_t ¡|_|dksJ‚||_dSr)r,r-rÚseq_lossrrr/Úalpha)r1r rHr2r rr-]ó     zMixLoss.__init__cCs*| ||¡}| ||¡}||j|}|S)N)rGr/rH)r1Ú seq_inputsÚ seq_targetsÚ class_inputsÚ class_targetsÚslÚcelr%r r rr>ds  zMixLoss.forwardr?r r r2rrFXs rFcrE)ÚMaxEntropySequenceLossz¡ A max entropy loss that encourage the model to have large entropy, therefore giving more diverse outputs. Loss = NLLLoss + alpha * EntropyLoss cs4tƒ ¡t |¡}d|tj<t |¡|_||_ dSr) r,r-rrrrrr ÚnllrH)r1r rHr r2r rr-qrIzMaxEntropySequenceLoss.__init__c Cs€| d¡| d¡ks J‚| ||¡}| tj¡ d¡ |¡}| ¡ |d¡}t   |¡}|  |¡  ¡| d¡}||j |}|S)z5 inputs: [N, C] targets: [N] rrg)ÚsizerQÚeqrrÚ unsqueezeÚ expand_asÚcloneÚ masked_fill_rr8ÚmulrrH) r1r:r;Únll_lossÚmaskÚ masked_inputsÚpÚent_lossr%r r rr>xs  zMaxEntropySequenceLoss.forwardr?r r r2rrPjs rP)F)rCÚloggingÚnumpyrrÚtorch.nnrÚ%stanza.models.common.seq2seq_constantÚmodelsÚcommonÚseq2seq_constantrÚ getLoggerrrr&ÚModuler'rFrPr r r rÚs   1