o %ݫi(@sdZddlZddlZddlmmZddZdddZ dd d Z dd d Z d dZ ddZ dddZ dddZddZddZdddZdS) zThis library implements different operations needed by complex- valued architectures. This work is inspired by: "Deep Complex Networks" from Trabelsi C. et al. Authors * Titouan Parcollet 2020 NcCsHt|dvrtdtt|d}|ddkr"tdt|dS)zCheck the complex-valued shape for a linear layer. Arguments --------- input_shape : tuple Expected shape of the input. >zCComplex linear accepts only input of dimension 2 or 3. input.dim = rzWComplex torch.Tensors must have an even number of hidden dimensions. input.size()[1] = N)len Exceptionstrinputdim) input_shape nb_hiddenr [/home/ubuntu/.local/lib/python3.10/site-packages/speechbrain/nnet/complex_networks/c_ops.pycheck_complex_inputs   rlinearrcCsp|dkr*|d}|dkr|dd|dS|dkr(|dd|dSdS||}||d|dS)aFReturns the real components of the complex-valued input. Arguments --------- input : torch.Tensor Input tensor. input_type : str, (convolution, linear) (default "linear") channels_axis : int. Default 1. Returns ------- The real component of complex-valued inputs. rrrrrrNsizer narrowr input_type channels_axisr nb_featmapsr r rget_real&s      rcCs||dkr.|d}|dkr|d|d|dS|dkr,|d|d|dSdS||}|||d|dS)aOReturns the imaginary components of the complex-valued input. Arguments --------- input : torch.Tensor Input tensor. input_type : str (convolution, linear) (default "linear") channels_axis : int Default 1. Returns ------- The imaginary components of complex-valued inputs. rrrrrNrrr r rget_imagFs    rcCsTt|||}t|||}|dkrtj|| gddS|dkr(tj|| g|dSdS)aLReturns the conjugate (z = r - xi) of the input complex numbers. Arguments --------- input : torch.Tensor Input tensor input_type : str, (convolution, linear) (default "linear") channels_axis : int. Default 1. Returns ------- The conjugate of the input complex numbers. rrr convolutionN)rrtorchcat)r rr input_imag input_realr r r get_conjugatefs  r cCstj|| gdd}tj||gdd}tj||gdd}|dkr2|jr,t|||St||St||}|jr?||S|S)a Applies a complex linear transformation to the incoming data. Arguments --------- input : torch.Tensor Complex input tensor to be transformed. real_weight : torch.Parameter Real part of the quaternion weight matrix of this layer. imag_weight : torch.Parameter First imaginary part of the quaternion weight matrix of this layer. bias : torch.Parameter Returns ------- Output after complex linear transformation is applied. rrrr)rrr requires_gradaddmmmmmatmul)r real_weight imag_weightbiascat_realcat_imag cat_complexoutputr r rcomplex_linear_op~s   r,c Cs\tj|| gdd}tj||gdd} tj|| gdd} |r"tj} ntj} | || ||||S)a>Applies a complex convolution to the incoming data. Arguments --------- input : torch.Tensor Complex input tensor to be transformed. real_weight : torch.Parameter Real part of the quaternion weight matrix of this layer. imag_weight : torch.Parameter First imaginary part of the quaternion weight matrix of this layer. bias : torch.Parameter stride : int Stride factor of the convolutional filters. padding : int Amount of padding. See torch.nn documentation for more information. dilation : int Dilation factor of the convolutional filters. conv1d : bool If true, a 1D convolution operation will be applied. Otherwise, a 2D convolution is called. Returns ------- Output after complex convolution is applied. rrr)rrFconv1dconv2d) r r%r&r'stridepaddingdilationr.r(r)r*convfuncr r rcomplex_conv_opsr4glorotc Cs|dur ||f}nt|tur||ft|f}n ||fg|R}t|}tjdd|}tjdd|}td|D]#}t||d||dd} ||| <||| <q<| |}| |}||fS)aReturns a matrix of unitary complex numbers. Arguments --------- in_features : int Number of real values of the input layer (quaternion // 4). out_features : int Number of real values of the output layer (quaternion // 4). kernel_size : int Kernel_size for convolutional layers (ex: (3,3)). criterion : str (glorot, he) (default "glorot"). Returns ------- Matrix of unitary complex numbers. Ng?rrg-C6?) typeinttuplenpprodrandomuniformrangesqrtreshape) in_features out_features kernel_size criterion kernel_shapenumber_of_weightsv_rv_iinormr r r unitary_inits   "  rKc Cs|durt|}||}||}n|}|}|dkr!d||}nd|}|dur.||f}nt|tur>||ft|f}n ||fg|R}tjj||d} tjtj tj|} | t | } | t | } | | fS)aReturns a matrix of complex numbers initialized as described in: "Deep Complex Networks", Trabelsi C. et al. Arguments --------- in_features : int Number of real values of the input layer (quaternion // 4). out_features : int Number of real values of the output layer (quaternion // 4). kernel_size : int Kernel_size for convolutional layers (ex: (3,3)). criterion: str (glorot, he) (default "glorot") Returns ------- Matrix of initialized complex numbers. Nr5r6)scaler) r:r;r7r8r9r<rayleighr=picossin) rArBrCrDreceptive_fieldfan_outfan_insrmodulusphase weight_real weight_imagr r r complex_inits&    rYcCsT||d|dd|\}}t|t|}}||j|_||j|_dS)aApplies the weight initialization function given to the parameters. Arguments --------- real_weight: torch.Parameters imag_weight: torch.Parameters init_func: function (unitary_init, complex_init) criterion: str (glorot, he) rrNrr from_numpytype_asdata)r%r& init_funcrDabr r r affect_init*s rac Cs^|d}|d}|||||d\}}t|t|}}||j|_||j|_dS)a^Applies the weight initialization function given to the parameters. This is specifically written for convolutional layers. Arguments --------- real_weight: torch.Parameters imag_weight: torch.Parameters kernel_size: int init_func: function (unitary_init, complex_init) criterion: str (glorot, he) rr)rCrDNrZ) r%r&rCr^rD in_channels out_channelsr_r`r r raffect_conv_init<s   rdFcCs*t|}|}t|D]}|||}q |S)zD Performs `torch.mean` over multiple dimensions of `input`. )sortedreversedmean)r axeskeepdimmaxisr r r multi_mean[s  rl)rr)Nr5)F)__doc__numpyr:rtorch.nn.functionalnn functionalr-rrrr r,r4rKrYrardrlr r r rs"   %) . 2