o %ݫi;@sdZddlZddlmZddlmmZddlmZddl m Z m Z m Z m Z ddlmZeeZGdddejjZGdd d ejZdS) zfLibrary implementing complex-valued convolutional neural networks. Authors * Titouan Parcollet 2020 N)get_padding_elem)affect_conv_initcomplex_conv_op complex_init unitary_init) get_loggercsReZdZdZ        dfdd Zd d Zd d ZddZddZZ S)CConv1daThis function implements complex-valued 1d convolution. Arguments --------- out_channels : int Number of output channels. Please note that these are complex-valued neurons. If 256 channels are specified, the output dimension will be 512. kernel_size : int Kernel size of the convolutional filters. input_shape : tuple The expected shape of the input tensor. stride : int, optional Stride factor of the convolutional filters (default 1). dilation : int, optional Dilation factor of the convolutional filters (default 1). padding : str, optional (same, valid, causal). If "valid", no padding is performed. If "same" and stride is 1, output shape is same as input shape. "causal" results in causal (dilated) convolutions. (default "same") groups : int, optional This option specifies the convolutional groups. See torch.nn documentation for more information (default 1). bias : bool, optional If True, the additive bias b is adopted (default True). padding_mode : str, optional This flag specifies the type of padding. See torch.nn documentation for more information (default "reflect"). init_criterion : str, optional (glorot, he). This parameter controls the initialization criterion of the weights. It is combined with weights_init to build the initialization method of the complex-valued weights. (default "glorot") weight_init : str, optional (complex, unitary). This parameter defines the initialization procedure of the complex-valued weights. "complex" will generate random complex-valued weights following the init_criterion and the complex polar form. "unitary" will normalize the weights to lie on the unit circle. (default "complex") More details in: "Deep Complex Networks", Trabelsi C. et al. Example ------- >>> inp_tensor = torch.rand([10, 16, 30]) >>> cnn_1d = CConv1d( ... input_shape=inp_tensor.shape, out_channels=12, kernel_size=5 ... ) >>> out_tensor = cnn_1d(inp_tensor) >>> out_tensor.shape torch.Size([10, 16, 24]) sameTreflectglorotcomplexc st||_||_||_||_||_||_||_| |_ d|_ | |_ | |_ | |d|_|\|_|_tjtj|j|_tjtj|j|_|jrdtjtd|j|_|jjdnd|_ttd|j |_t|j|j|j|j|j dSNFr)r unitary)super__init__ out_channels kernel_sizestridedilationpaddinggroupsbias padding_mode unsqueezeinit_criterion weight_init _check_input in_channels_get_kernel_and_weight_shapek_shapew_shapetorchnn ParameterTensor real_weight imag_weightbdatafill_rrwinitr selfrr input_shaperrrrrrrr __class__[/home/ubuntu/.local/lib/python3.10/site-packages/speechbrain/nnet/complex_networks/c_CNN.pyrMs< zCConv1d.__init__c Cs|dd}|jdkr|||j|j|j}n#|jdkr-|jd|j}t||df}n |jdkr3ntd|jt ||j |j |j |jd|jdd }|dd}|S) a-Returns the output of the convolution. Arguments --------- x : torch.Tensor (batch, time, channel). Input to convolve. 3d or 4d tensors are expected. Returns ------- wx : torch.Tensor The convolved outputs. r r causalrvalid4Padding must be 'same', 'valid' or 'causal'. Got %s.Trrrconv1d) transposer_manage_paddingrrrFpad ValueErrorrr'r(r))r.xnum_padwxr2r2r3forwards6     zCConv1d.forwardcCs2|jd}t||||}tj|t||jd}|S)aThis function performs zero-padding on the time axis such that their lengths is unchanged after the convolution. Arguments --------- x : torch.Tensor Input tensor. kernel_size : int Kernel size. dilation : int Dilation. stride : int Stride. Returns ------- x : torch.Tensor The padded outputs. r4mode)shaperr<r=tupler)r.r?rrrL_inrr2r2r3r;s zCConv1d._manage_paddingcCsnt|dkr |d}ntd||jddkrtd|j|ddkr5tdt|jdt|j|S):Checks the input and returns the number of input channels.rz%ComplexConv1d expects 3d inputs. Got r4The field kernel size must be an odd number. Got %s.HComplex torch.Tensors must have dimensions divisible by 2. input.size()[] = )lenr>rstr channels_axis nb_channelsr.r/rr2r2r3rs,   zCConv1d._check_inputcCs$|j}|j|jft|f}||fS)BReturns the kernel size and weight shape for convolutional layers.)rrrrFr.ksr"r2r2r3r sz$CConv1d._get_kernel_and_weight_shaper r r r Tr r r ) __name__ __module__ __qualname____doc__rrBr;rr __classcell__r2r2r0r3rs:70 rcsTeZdZdZ        dfdd Zdd d Zd dZddZddZZ S)CConv2daThis function implements complex-valued 1d convolution. Arguments --------- out_channels : int Number of output channels. Please note that these are complex-valued neurons. If 256 channels are specified, the output dimension will be 512. kernel_size : int Kernel size of the convolutional filters. input_shape : tuple The expected shape of the input. stride : int, optional Stride factor of the convolutional filters (default 1). dilation : int, optional Dilation factor of the convolutional filters (default 1). padding : str, optional (same, valid, causal). If "valid", no padding is performed. If "same" and stride is 1, output shape is same as input shape. "causal" results in causal (dilated) convolutions. (default "same") groups : int, optional This option specifies the convolutional groups (default 1). See torch.nn documentation for more information. bias : bool, optional If True, the additive bias b is adopted (default True). padding_mode : str, optional This flag specifies the type of padding (default "reflect"). See torch.nn documentation for more information. init_criterion : str , optional (glorot, he). This parameter controls the initialization criterion of the weights (default "glorot"). It is combined with weights_init to build the initialization method of the complex-valued weights. weight_init : str, optional (complex, unitary). This parameter defines the initialization procedure of the complex-valued weights (default complex). "complex" will generate random complex-valued weights following the init_criterion and the complex polar form. "unitary" will normalize the weights to lie on the unit circle. More details in: "Deep Complex Networks", Trabelsi C. et al. Example ------- >>> inp_tensor = torch.rand([10, 16, 30, 30]) >>> cnn_2d = CConv2d( ... input_shape=inp_tensor.shape, out_channels=12, kernel_size=5 ... ) >>> out_tensor = cnn_2d(inp_tensor) >>> out_tensor.shape torch.Size([10, 16, 30, 24]) r r Tr r r c sLt||_||_||_||_||_||_||_| |_ d|_ | |_ | |_ t |jtr3|j|jg|_t |jtr@|j|jg|_t |jtrM|j|jg|_||d|_|\|_|_tjtj|j|_tjtj|j|_|jrtjtd|j|_|jjdnd|_ttd|j |_t |j|j|j|j|j dSr)!rrrrrrrrrrrrr isinstanceintrrr r!r"r#r$r%r&r'r(r)r*r+rrr,rr-r0r2r3r/sH    zCConv2d.__init__Fc Cs|r|||dd}|jdkr|||j|j|j}n#|jdkr4|jd|j}t||df}n |jdkr:nt d|jt ||j |j |j |jd|jdd }|dd}|S) aReturns the output of the convolution. Arguments --------- x : torch.Tensor (batch, time, feature, channels). Input to convolve. 3d or 4d tensors are expected. init_params : bool Whether to initialize the parameters in this pass. Returns ------- x : torch.Tensor The output of the convolution. r r4r r5rr6r7Fr8) init_paramsr:rr;rrrr<r=r>rr'r(r))r.r?r^r@rAr2r2r3rBps:      zCConv2d.forwardcCs2|jd|jdf}|j|jfg|R}||fS)rRrr )rrrrSr2r2r3r sz$CConv2d._get_kernel_and_weight_shapec Csb|jd}t||d|d|d}t||d|d|d}||}tjj|t||jd}|S)aThis function performs zero-padding on the time and frequency axes such that their lengths is unchanged after the convolution. Arguments --------- x : torch.Tensor Input tensor. kernel_size : int Kernel size. dilation : int Dilation. stride: int Stride. Returns ------- x : torch.Tensor The padded tensor. r4rC)rErr$ functionalr=rFr) r.r?rrrrG padding_time padding_freqrr2r2r3r;s zCConv2d._manage_paddingcCst|dkr d|_d}nt|dkr|d}ntd||jdddks/|jdddkr6td|j|ddkrLtd t|jd t|j|S) rHrITr zExpected 3d or 4d inputs. Got rrrJrKrL)rMrr>rrNrOrPrQr2r2r3rs.    $ zCConv2d._check_inputrU)F) rVrWrXrYrrBr r;rrZr2r2r0r3r[s: A8&r[)rYr#torch.nnr$torch.nn.functionalr`r<speechbrain.nnet.CNNr'speechbrain.nnet.complex_networks.c_opsrrrrspeechbrain.utils.loggerrrVloggerModulerr[r2r2r2r3s   c