o pit@sdddlmZddlZddlmZddlmmZdejdejdejfddZ Gdd d ej Z dS) )OptionalN sequencesweightsreturncCs|jdd}|jddd}tj||dd|}t||d}t|jdd}tj||dd|||d}t|}tj||gddS)aHelper function to compute statistics pooling Assumes that weights are already interpolated to match the number of frames in sequences and that they encode the activation of only one speaker. Parameters ---------- sequences : (batch, features, frames) torch.Tensor Sequences of features. weights : (batch, frames) torch.Tensor (Already interpolated) weights. Returns ------- output : (batch, 2 * features) torch.Tensor Concatenation of mean and (unbiased) standard deviation. dimg:0yE>) unsqueezesumtorchsquaresqrtcat)rrv1meandx2v2varstdrX/home/ubuntu/.local/lib/python3.10/site-packages/pyannote/audio/models/blocks/pooling.py_pools " rc@s4eZdZdZ ddejdeejdejfddZdS) StatsPoolzStatistics pooling Compute temporal mean and (unbiased) standard deviation and returns their concatenation. Reference --------- https://en.wikipedia.org/wiki/Weighted_arithmetic_mean Nrrrc sdurjdd}jddd}tj||gddSdkr)d}jddnd}\}}}\}}} || krEtj|d d tj fd d t |Ddd} |s^| j ddS| S) aForward pass Parameters ---------- sequences : (batch, features, frames) torch.Tensor Sequences of features. weights : (batch, frames) or (batch, speakers, frames) torch.Tensor, optional Compute weighted mean and standard deviation, using provided `weights`. Note ---- `sequences` and `weights` might use a different number of frames, in which case `weights` are interpolated linearly to reach the number of frames in `sequences`. Returns ------- output : (batch, 2 * features) or (batch, speakers, 2 * features) torch.Tensor Concatenation of mean and (unbiased) standard deviation. When `weights` are provided with the `speakers` dimension, `output` is computed for each speaker separately and returned as (batch, speakers, 2 * channel)-shaped tensor. Nrr)r correctionr FTnearest)sizemodec s(g|]}tdd|ddfqSN)r).0speakerrrrr xsz%StatsPool.forward..) rrr rrr rF interpolatestackrangesqueeze) selfrrrrhas_speaker_dimension_ num_frames num_speakers num_weightsoutputrr"rforwardLs*    zStatsPool.forwardr)__name__ __module__ __qualname____doc__r Tensorrr0rrrrr@s r) typingrr torch.nnnntorch.nn.functional functionalr$r5rModulerrrrrs  "