o %ݫiE@sdZddlZddlmZddlmmZddlmZee Z Gdddej Z Gdddej Z Gdd d ej ZGd d d ej ZGd d d ej ZGdddej ZdS)zLibrary implementing pooling. Authors * Titouan Parcollet 2020 * Mirco Ravanelli 2020 * Nauman Dawalatabad 2020 * Jianyuan Zhong 2020 * Sarthak Yadav 2022 * Ha Nguyen 2023 N) get_loggercs6eZdZdZ      d fdd Zd d ZZS) Pooling1daThis function implements 1d pooling of the input tensor. Arguments --------- pool_type : str It is the type of pooling function to use ('avg','max'). kernel_size : int It is the kernel size that defines the pooling dimension. For instance, kernel size=3 applies a 1D Pooling with a size=3. input_dims : int The count of dimensions expected in the input. pool_axis : int The axis where the pooling is applied. ceil_mode : bool When True, will use ceil instead of floor to compute the output shape. padding : int It is the number of padding elements to apply. dilation : int Controls the dilation factor of pooling. stride : int It is the stride size. Example ------- >>> pool = Pooling1d('max',3) >>> inputs = torch.rand(10, 12, 40) >>> output=pool(inputs) >>> output.shape torch.Size([10, 4, 40]) FrNc st||_|dur|}|dkr>|dkr#tjj||||d|_dS|dkr:tjjd|fd|fd|f|d|_dStd|dkrr|dkrTtjj |||||d |_dS|dkrntjj d|fd|fd|fd|f|d |_dStdtd ) Navgrstridepadding ceil_moderrzinput_dims must be 3 or 4max)rr dilationr z pool_type must be 'avg' or 'max') super__init__ pool_axistorchnn AvgPool1d pool_layer AvgPool2d ValueError MaxPool1d MaxPool2d) self pool_type kernel_size input_dimsrr r r r __class__L/home/ubuntu/.local/lib/python3.10/site-packages/speechbrain/nnet/pooling.pyr5sN     zPooling1d.__init__cCs*|d|j}||}|d|j}|S)zPerforms 1d pooling to the input tensor. Arguments --------- x : torch.Tensor It represents a tensor for a mini-batch. Returns ------- x : torch.Tensor The pooled outputs. ) transposerrrxrrr forwardos zPooling1d.forward)rrFrrN__name__ __module__ __qualname____doc__rr% __classcell__rrrr rs#:rcs4eZdZdZ     d fdd Zd d ZZS) Pooling2daThis function implements 2d pooling of the input tensor. Arguments --------- pool_type : str It is the type of pooling function to use ('avg','max'). kernel_size : int It is the kernel size that defines the pooling dimension. For instance, kernel size=3,3 performs a 2D Pooling with a 3x3 kernel. pool_axis : tuple It is a list containing the axis that will be considered during pooling. ceil_mode : bool When True, will use ceil instead of floor to compute the output shape. padding : int It is the number of padding elements to apply. dilation : int Controls the dilation factor of pooling. stride : int It is the stride size. Example ------- >>> pool = Pooling2d('max',(5,3)) >>> inputs = torch.rand(10, 15, 12) >>> output=pool(inputs) >>> output.shape torch.Size([10, 3, 4]) rFrrNcst||_||_||_||_||_||_|dur||_n||_|jdkr8t j j |j|j|j|jd|_ dSt j j |j|j|j|jd|_ dS)Nrr)rrrrrr r r rrrrrr)rrrrr r r rrrr rs.    zPooling2d.__init__cCs|ddd|jdd|jd|jd|jd}||}||jd|jdd|jdd|jddd}|S)zPerforms 2d pooling to the input tensor. Arguments --------- x : torch.Tensor It represents a tensor for a mini-batch. Returns ------- x : torch.Tensor The pooled outputs. r!rr) unsqueezer"rsqueezerr#rrr r%s      zPooling2d.forward)r-FrrNr&rrrr r,s"&r,cs6eZdZdZd fdd Zd ddZd d d ZZS)StatisticsPoolingaThis class implements a statistic pooling layer. It returns the mean and/or std of input tensor. Arguments --------- return_mean : bool If True, the average pooling will be returned. return_std : bool If True, the standard deviation will be returned. Example ------- >>> inp_tensor = torch.rand([5, 100, 50]) >>> sp_layer = StatisticsPooling() >>> out_tensor = sp_layer(inp_tensor) >>> out_tensor.shape torch.Size([5, 1, 100]) Tcs8td|_||_||_|js|jstddSdS)Ngh㈵>zZboth of statistics are equal to False consider enabling mean and/or std statistic pooling)rreps return_mean return_stdr)rr4r5rrr rs  zStatisticsPooling.__init__Nc Cs`|dur|jr |jdd}|jr|jdd}nTg}g}t|jdD]8}tt|||jd}|jrF| tj||d|dfdd|jrZ| tj||d|dfddq"|jrct |}|jrkt |}|jr~|j | |j d}|}||7}|jr||j}|jr|jrtj||fdd}|d}|S|jr|d}|S|jr|d}|S)atCalculates mean and std for a batch (input tensor). Arguments --------- x : torch.Tensor It represents a tensor for a mini-batch. lengths : torch.Tensor The lengths of the samples in the input. Returns ------- pooled_stats : torch.Tensor The mean and std for the input. Nrdimr.device)r4meanr5stdrangeshapeintrroundappendstack_get_gauss_noisesizer9r3catr0) rr$lengthsr:r;snt_id actual_sizegnoise pooled_statsrrr r%!sJ  "       zStatisticsPooling.forwardcpucCs@tj||d}|t|8}|t|}|jd|d}|S)auReturns a tensor of epsilon Gaussian noise. Arguments --------- shape_of_tensor : torch.Tensor It represents the size of tensor for generating Gaussian noise. device : str Device on which to perform computations. Returns ------- gnoise : torch.Tensor The Gaussian noise. r8i )rrandnminr r3)rshape_of_tensorr9rHrrr rBZs z"StatisticsPooling._get_gauss_noise)TT)N)rJ)r'r(r)r*rr%rBr+rrrr r2s  9r2c(eZdZdZfddZddZZS) AdaptivePoola6This class implements the adaptive average pooling. Arguments --------- output_size : int The size of the output. Example ------- >>> pool = AdaptivePool(1) >>> inp = torch.randn([8, 120, 40]) >>> output = pool(inp) >>> output.shape torch.Size([8, 1, 40]) cstt|tpt|tpt|t}|sJdt|ts$t|tr.t|dks.Jdt|tr;t||_ dSt ||_ dS)Nz&output size must be int, list or tupler.z-len of output size must not be greater than 2) rr isinstancer>tuplelistlenrAdaptiveAvgPool1dpoolAdaptiveAvgPool2d)r output_size conditionrrr rs    zAdaptivePool.__init__cCsX|jdkr||ddddddS|jdkr*||ddddddddSdS)aPerforms adaptive pooling to the input tensor. Arguments --------- x : torch.Tensor It represents a tensor for a mini-batch. Returns ------- x : torch.Tensor The pooled outputs. rrr.rr N)ndimrVpermuter#rrr r%s  "zAdaptivePool.forwardr&rrrr rPqs rPcsFeZdZdZ      dfdd Zd d Zd d ZddZZS)GaussianLowpassPoolinga This class implements a learnable Gaussian lowpass pooling from Neil Zeghidour, Olivier Teboul, F{'e}lix de Chaumont Quitry & Marco Tagliasacchi, "LEAF: A LEARNABLE FRONTEND FOR AUDIO CLASSIFICATION", in Proc. of ICLR 2021 (https://arxiv.org/abs/2101.08596) Arguments --------- in_channels : int The number of input channels. kernel_size: int Kernel size of the gaussian lowpass filters. stride : int Stride factor of the convolutional filters. When the stride factor > 1, a decimation in time is performed. initialization_constant : float The constant used for initialization, default 0.4 padding : str (same, valid). If "valid", no padding is performed. If "same" and stride is 1, output shape is the same as the input shape. padding_mode : str This flag specifies the type of padding. See torch.nn documentation for more information. bias : bool If True, the additive bias b is adopted. skip_transpose : bool If False, uses batch x time x channel convention of speechbrain. If True, uses batch x channel x time convention. Example ------- >>> inp_tensor = torch.rand([10, 8000, 40]) >>> low_pass_pooling = GaussianLowpassPooling( ... 40, kernel_size=401, stride=160, ... ) >>> # parameters corresponding to a window of 25 ms and stride 10 ms at 16000 kHz >>> out_tensor = low_pass_pooling(inp_tensor) >>> out_tensor.shape torch.Size([10, 50, 40]) r皙?sameconstantTFc srt||_||_||_||_||_||_t t dd|df||_ |r4t j t ||_ dSd|_ dSNr)rrrrr padding_mode in_channelsskip_transposer Parameterronesweights_bias) rrbrrinitialization_constantr rabiasrcrrr rs  zGaussianLowpassPooling.__init__cCsz|j}tj|d|dd}tjd||j|jd}t|d|ddf}|d|d}|d|d}td||dS) Ng@g?)rMr r)dtyper9rgr.)rrclamparangerjr9reshapeexp)rsigma filter_sizet numerator denominatorrrr _get_impulse_responsessz-GaussianLowpassPooling._get_impulse_responsescCs|js |dd}||j}|d|j|j}|ddd}|jdkr,| ||j}n |jdkr2nt d|jt j |||j |jd|jd}|jsP|dd}|S) zPerforms GaussianLowpass Pooling. Arguments --------- x : torch.Tensor 3D tensor in input [batch,time,channels]. Returns ------- outputs : torch.Tensor The pooled outputs. rr!r.rr^validz'Padding must be 'same' or 'valid'. Got )rirr groups)rcr"rtrfrmrrbr[r _manage_paddingrFconv1drgr)rr$kerneloutputsrrr r%s.     zGaussianLowpassPooling.forwardcCs(dd}||}tj|||jdd}|S)NcSs@|f}ddlm}ddlm}||dd|dddD}|S)zGet number of elements to pad.r)reduce)__add__cSs0g|]}|d|d|dd|dfqS)r.rr).0krrr ,s"zUGaussianLowpassPooling._manage_padding..get_padding_value..Nr!) functoolsr|operatorr})r kernel_sizesr|r} conv_paddingrrr get_padding_value$s   zAGaussianLowpassPooling._manage_padding..get_padding_valuer)modevalue)rxpadra)rr$rr pad_valuerrr rw sz&GaussianLowpassPooling._manage_padding)rr]r^r_TF) r'r(r)r*rrtr%rwr+rrrr r\s- (r\crO)AttentionPoolingaPThis function implements a self-attention pooling (https://arxiv.org/abs/2008.01077). Arguments --------- input_dim: int The dimension of the input torch.Tensor Example ------- >>> inp_tensor = torch.rand([4, 40]) >>> pool = AttentionPooling(input_dim=40) >>> out_tensor = pool(inp_tensor) cs$t||_tj|d|_dSr`)rr input_dimrrLinearattn_pooling_w)rrrrr rGs zAttentionPooling.__init__cCsB||d}tjjj|ddd}tj||dd}|S)zReturns the output the adapter. Arguments --------- x : torch.Tensor Input tensor. Returns ------- out : torch.Tensor The pooled outputs. r!r6r) rr1floatrr functionalsoftmaxr0sum)rr$outrrr r%Os zAttentionPooling.forwardr&rrrr r8s r)r*rtorch.nnrtorch.nn.functionalrrxspeechbrain.utils.loggerrr'loggerModulerr,r2rPr\rrrrr s  swr9