o }oi@sddlZddlmmZddlmZddlmZmZddlm Z m Z m Z m Z m Z mZdgZGdddeZGdddeZdS) N)nn)Loss typecheck) LabelsType LengthsType LogprobsTypeLossType NeuralTypeSpectrogramTypeMLMLosscsbeZdZeddZeddZeddZ  dd ed effd d Z e dddZ Z S)r c CsZtdtddtdttdtttdtddttdtddtdtdddS)z(Input types definitions for Contrastive.BDTToptional)r rrr rr ) spec_masksdecoder_outputstargetsdecoder_lengthstarget_lengthsmasks)r r rrtuplerselfr^/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/asr/losses/ssl_losses/mlm.py input_typess  zMLMLoss.input_typescCsdttdiS)z]Output types definitions for Contrastive. loss: NeuralType(None) loss) elements_type)r rrrrr output_types&szMLMLoss.output_typescCsdS)NTrrrrr needs_labels.szMLMLoss.needs_labels皙?combine_time_stepsmask_thresholdcs$tt|_||_||_dSN)super__init__rNLLLossnll_lossr%r&)rr%r& __class__rrr)2s   zMLMLoss.__init__Nc Cs|dur|}|dd}||jd|jd|jd}|d|jk}||}t|d|jd|jdf}||}|||} t | } | S)Nr#r) transposereshapeshaper%meanr&Fpadr+torch) rrrrrrrout_masked_onlytargets_masked_onlyrrrrforward<s     zMLMLoss.forward)r#r$)NNNN) __name__ __module__ __qualname__propertyrr!r"intfloatr)rr9 __classcell__rrr,rr s"    c sXeZdZdZeddZ    ddeded ed effd d Z e dddZ Z S) MultiMLMLossa; Masked language model loss for multiple decoders, where cross-entropy loss is applied separately on each decoder. This loss can be used with `nemo.collections.asr.modules.ssl_modules.MultiSoftmaxDecoder` to train a model with multiple targets per frame. Reference: https://arxiv.org/abs/2202.01855 c Csz|jr|jdkrtdt}tdt}n tdt}tdt}tdt||ttdtdd ttdtdd d S) Nr#)r rCr)r rrBH)r rrCr r Tr)rrrrr)squeeze_single num_decodersr rrr rr)rrrrrrr[s    zMultiMLMLoss.input_typesr#r$Fr%r&rErDcs&t||_||_t|||_dSr')r(r)rErDr mlm_loss)rr%r&rErDr,rrr)ks zMultiMLMLoss.__init__Nc Cs|jr|jdkr|j|||||dSd}t|jD]#}||j||dddddd|f|dddd|f||d7}q||jS)Nr#)rrrrrg)rDrErFrange)rrrrrrrirrrr9ws$ zMultiMLMLoss.forward)r#r$r#F)NN) r:r;r<__doc__r=rr>r?boolr)rr9r@rrr,rrATs&  rA)r6torch.nn.functionalr functionalr4 nemo.corerrnemo.core.neural_typesrrrrr r __all__r rArrrrs  ;