o i @s\ddlmZddlmZddlZddlmZddlmZddl m Z Gdddej j Z dS) )Tuple)wpe_one_iterationN) ComplexTensor) MaskEstimator) make_pad_maskcseZdZ          dd ed ed ededededededededeffdd Zdede j de ee j effddZ Z S)DNN_WPEblstmp,@TFwtypewidimwlayerswunitswprojs dropout_ratetapsdelay use_dnn_mask iterations normalizationc sTt| |_||_||_| |_| |_d|_|jr(t||||||dd|_ dSdS)NTr)nmask) super__init__rrrrr inverse_powerrmask_est) selfrrrrrrrrrrr __class__R/home/ubuntu/.local/lib/python3.10/site-packages/funasr/frontends/utils/dnn_wpe.pyr s  zDNN_WPE.__init__datailensreturncCs|dddd}}d}t|jD]K}|jd|jd}|dkr>|jr>|||\\}}|jr:||jddd}||}|j d d}t | ||j |j |jd }|t||jdq|dddd}|duro|dd }|||fS) aMThe forward function Notation: B: Batch C: Channel T: Time or Sequence length F: Freq or Some dimension of the feature vector Args: data: (B, C, T, F) ilens: (B,) Returns: data: (B, C, T, F) ilens: (B,) rr rN)dim).N)rrr)permuterangerrealimagrrrsummeanr contiguousrrr masked_fill_r transpose)r r%r&enhancedmaskipower_r#r#r$forward)s,   zDNN_WPE.forward) rr r r r r rr TrF)__name__ __module__ __qualname__strintfloatboolrrtorch LongTensorrr; __classcell__r#r#r!r$r sT    r)typingr pytorch_wperrCtorch_complex.tensorr%funasr.frontends.utils.mask_estimatorr*funasr.models.transformer.utils.nets_utilsrnnModulerr#r#r#r$s