U %d-@s\dZddlmZmZddlZGdddejjZGdddejjZGdd d ejjZ dS) zImplements Conv-TasNet with building blocks of it. Based on https://github.com/naplab/Conv-TasNet/tree/e66d82a8f956a69749ec8a4ae382217faa097c5c )OptionalTupleNcsTeZdZdZd eeeeeedfdd Zeje e ejejfddd Z Z S) ConvBlocka[1D Convolutional block. Args: io_channels (int): The number of input/output channels, hidden_channels (int): The number of channels in the internal layers, . kernel_size (int): The convolution kernel size of the middle layer,

. padding (int): Padding value of the convolution in the middle layer. dilation (int, optional): Dilation value of the convolution in the middle layer. no_redisual (bool, optional): Disable residual block/output. Note: This implementation corresponds to the "non-causal" setting in the paper. F) io_channelshidden_channels kernel_sizepaddingdilation no_residualc sttjtjj||ddtjtjjd|ddtjj||||||dtjtjjd|dd|_|rtdntjj||dd|_ tjj||dd|_ dS)Nr in_channels out_channelsr:0yE>Z num_groups num_channelsZeps)r rrr r groups) super__init__torchnnZ SequentialConv1dPReLU GroupNorm conv_layersres_outskip_out)selfrrrr r r  __class__A/tmp/pip-unpacked-wheel-lbdmvq91/torchaudio/models/conv_tasnet.pyrs* zConvBlock.__init__inputreturncCs6||}|jdkrd}n ||}||}||fS)N)rrr)rr#Zfeatureresidualrr r r!forward<s     zConvBlock.forward)rF) __name__ __module__ __qualname____doc__intboolrrTensorrrr& __classcell__r r rr!r s"rc sHeZdZdZeeeeeeeedfdd ZejejdddZ Z S) MaskGeneratoraTCN (Temporal Convolution Network) Separation Module Generates masks for separation. Args: input_dim (int): Input feature dimension, . num_sources (int): The number of sources to separate. kernel_size (int): The convolution kernel size of conv blocks,

. num_featrs (int): Input/output feature dimenstion of conv blocks, . num_hidden (int): Intermediate feature dimention of conv blocks, num_layers (int): The number of conv blocks in one stack, . num_stacks (int): The number of conv block stacks, . msk_activate (str): The activation function of the mask output. Note: This implementation corresponds to the "non-causal" setting in the paper.  input_dim num_sourcesr num_feats num_hidden num_layers num_stacks msk_activatec s6t||_||_tjjd|dd|_tjj||dd|_ d|_ tj g|_ t |D]t} t |D]f} d| } |j t|||| | | |dko| |dkd|j | dkr| dkr|n |d| 7_ qfqZtj|_tjj|||dd|_|dkr tj|_n&|d kr$tj|_ntd |dS) Nrrrr r)rrrr r r sigmoidZreluzUnsupported activation )rrr1r2rrr input_normr input_convZreceptive_fieldZ ModuleListrrangeappendrr output_prelu output_convZSigmoid mask_activateZReLU ValueError) rr1r2rr3r4r5r6r7slZmultirr r!rYs@    .   zMaskGenerator.__init__r"cCs|jd}||}||}d}|jD](}||\}}|dk rH||}||}q(||}||}||}|||j|j dS)zGenerate separation mask. Args: input (torch.Tensor): 3D Tensor with shape [batch, features, frames] Returns: Tensor: shape [batch, num_sources, features, frames] rgN) shaper:r;rr>r?r@viewr2r1)rr# batch_sizefeatsoutputZlayerr%skipr r r!r&s         zMaskGenerator.forward) r'r(r)r*r+strrrr-r&r.r r rr!r/Fs1r/c sheZdZdZdeeeeeeeeed fd d Zeje ejefd d dZ ejejd ddZ Z S) ConvTasNetuConv-TasNet: a fully-convolutional time-domain audio separation network *Conv-TasNet: Surpassing Ideal Time–Frequency Magnitude Masking for Speech Separation* [:footcite:`Luo_2019`]. Args: num_sources (int, optional): The number of sources to split. enc_kernel_size (int, optional): The convolution kernel size of the encoder/decoder, . enc_num_feats (int, optional): The feature dimensions passed to mask generator, . msk_kernel_size (int, optional): The convolution kernel size of the mask generator,

. msk_num_feats (int, optional): The input/output feature dimension of conv block in the mask generator, . msk_num_hidden_feats (int, optional): The internal feature dimension of conv block of the mask generator, . msk_num_layers (int, optional): The number of layers in one conv block of the mask generator, . msk_num_stacks (int, optional): The numbr of conv blocks of the mask generator, . msk_activate (str, optional): The activation function of the mask output (Default: ``sigmoid``). Note: This implementation corresponds to the "non-causal" setting in the paper. r8r9) r2enc_kernel_size enc_num_featsmsk_kernel_size msk_num_featsmsk_num_hidden_featsmsk_num_layersmsk_num_stacksr7c st||_||_||_|d|_tjjd|||j|jdd|_ t |||||||| d|_ tjj |d||j|jdd|_ dS)Nr8rF)r rrZstrider Zbiasr0)rrr2rSrR enc_striderrrencoderr/mask_generatorZConvTranspose1ddecoder) rr2rRrSrTrUrVrWrXr7rr r!rs>   zConvTasNet.__init__r"c Cs||j\}}}|jd}|||j}||||j}|dkrF|dfS|j|}tj||||j|jd} t|| gd|fS)aPad input Tensor so that the end of the input tensor corresponds with 1. (if kernel size is odd) the center of the last convolution kernel or 2. (if kernel size is even) the end of the first half of the last convolution kernel Assumption: The resulting Tensor will be padded with the size of stride (== kernel_width // 2) on the both ends in Conv1D |<--- k_1 --->| | | |<-- k_n-1 -->| | | | |<--- k_n --->| | | | | | | | | | | | v v v | |<---->|<--- input signal --->|<--->|<---->| stride PAD stride Args: input (torch.Tensor): 3D Tensor with shape (batch_size, channels==1, frames) Returns: Tensor: Padded Tensor int: Number of paddings performed r8r)dtypedevice)rErRrYrzerosr]r^cat) rr#rGrZ num_framesZis_oddZ num_stridesZnum_remainingsZ num_paddingspadr r r!_align_num_frames_with_stridess   z)ConvTasNet._align_num_frames_with_stridesc Cs|jdks|jddkr(td|j||\}}|jd|jd}}||}|||d}|||j|j d}| |}|||j|} |dkr| dd| f} | S) aPerform source separation. Generate audio source waveforms. Args: input (torch.Tensor): 3D Tensor with shape [batch, channel==1, frames] Returns: Tensor: 3D Tensor with shape [batch, channel==num_sources, frames] rOrz7Expected 3D tensor (batch, channel==1, frames). Found: rr8rD.N) ndimrErArbrZr[Z unsqueezerFr2rSr\) rr#ZpaddedZnum_padsrGZnum_padded_framesrHZmaskeddecodedrIr r r!r&s    zConvTasNet.forward) r8rMrNrOrPrNrQrOr9) r'r(r)r*r+rKrrr-rrbr&r.r r rr!rLs.0+rL) r*typingrrrrModulerr/rLr r r r!s ;\