o ߥi.@sddZddlZddlmZGdddejZGdddejZGdddejZGd d d ejZdS) zM This implementation is adapted from https://github.com/wenet-e2e/wespeaker. Nc(eZdZdZfddZddZZS)TAPzG Temporal average pooling, only first-order mean is considered c tt|dSN)superr__init__selfkwargs __class__]/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/audio/sv/pooling_layers.pyr z TAP.__init__cCs|jdd}|jdd}|S)Ndim start_dim)meanflatten)r x pooling_meanr r rforwards  z TAP.forward__name__ __module__ __qualname____doc__rr __classcell__r r r rr rcr)TSDPzR Temporal standard deviation pooling, only second-order std is considered c rr)rr"rrr r rrrz TSDP.__init__cCs(ttj|ddd}|jdd}|SNrrg:0yE>rr)torchsqrtvarr)r r pooling_stdr r rrs z TSDP.forwardrr r r rr"r!r"cr)TSTPz Temporal statistics pooling, concatenate mean and std, which is used in x-vector Comment: simple concatenation can not make full use of both statistics c rr)rr(rrr r rr-rz TSTP.__init__cCsP|jdd}ttj|ddd}|jdd}|jdd}t||fd}|Sr#)rr$r%r&rcat)r rrr'statsr r rr0s   z TSTP.forwardrr r r rr(&s r(cs*eZdZdZdfdd ZddZZS) ASTPzx Attentive statistics pooling: Channel- and context-dependent statistics pooling, first used in ECAPA_TDNN. FcsXtt|||_|rtj|d|dd|_n tj||dd|_tj||dd|_dS)Nr) kernel_size)rr+rglobal_context_attnnConv1dlinear1linear2)r in_dimbottleneck_dimr/r r rr@s  z ASTP.__init__c Cst|jdkr||jd|jd|jd|jd}t|jdks%J|jrNtj|ddd|}ttj|dddd |}tj |||fdd }n|}t | |}tj | |dd }tj||dd }tj||ddd |d}t|jd d }tj ||gdd S) z x: a 3-dimensional tensor in tdnn-based architecture (B,F,T) or a 4-dimensional tensor in resnet architecture (B,C,F,T) 0-dim: batch-dimension, last-dim: time-dimension (frame-dimension) rrr-rT)rkeepdimg|=r)min)lenshapereshaper/r$r expand_asr%r&r)tanhr2softmaxr3sumclamp) r r context_mean context_stdx_inalpharr&stdr r rrRs(*z ASTP.forward)r,Frr r r rr+;sr+) rr$torch.nnr0Modulerr"r(r+r r r rs