o %ݫi&@sfdZddlZddlmZddlmZddlZddejfddZGdd d ej Z Gd d d ej Z dS) zLGenerator and discriminator used in MetricGAN-U Authors: * Szu-Wei Fu 2020 N)nn) spectral_normTcKsN|dur|}|||fi|}|rt|}tjj|jddtj|j|S)zDCreate a layer with spectral norm, xavier uniform init and zero biasNg?)gain)rrinitxavier_uniform_weightzeros_bias)in_sizeout_size spec_norm layer_typekwargslayerrX/home/ubuntu/.local/lib/python3.10/site-packages/speechbrain/lobes/models/MetricGAN_U.pyxavier_init_layersrcs4eZdZdZ     d fdd Zd d ZZS) EnhancementGeneratoraSimple LSTM for enhancement with custom initialization. Arguments --------- input_size : int Size of the input tensor's last dimension. hidden_size : int Number of neurons to use in the LSTM layers. num_layers : int Number of layers to use in the LSTM. lin_dim: int Number of neurons in the last two linear layers. dropout : int Fraction of neurons to drop during training. Example ------- >>> inputs = torch.rand([10, 100, 40]) >>> model = EnhancementGenerator(input_size=40, hidden_size=50) >>> outputs = model(inputs, lengths=torch.ones([10])) >>> outputs.shape torch.Size([10, 100, 40]) ,rcsttjdd|_tjjj||||dd|_ |j D]$\}}d|vr.tj |qd|vr9tj |qd|vrCtj |qt|d|d d |_t||d d |_t|_dS) N333333?negative_slopeT) input_size hidden_size num_layersdropout bidirectionalr weight_ih weight_hhrF)r )super__init__r LeakyReLU activationsbnnetRNNLSTMblstmnamed_parametersrrr orthogonal_rlinear1linear2Sigmoidsigmoid)selfrrrlin_dimrnameparam __class__rrr#9s.   zEnhancementGenerator.__init__cCs>|j||d\}}||}||}||}||}|S):Processes the input tensor x and returns an output tensor.)lengths)r*r-r%r.r0)r1xr8out_rrrforward^s     zEnhancementGenerator.forward)rrrrr)__name__ __module__ __qualname____doc__r#r< __classcell__rrr5rr s%rcs6eZdZdZddejddffdd Zdd ZZS) MetricDiscriminatoraMetric estimator for enhancement training. Consists of: * four 2d conv layers * channel averaging * three linear layers Arguments --------- kernel_size : tuple The dimensions of the 2-d kernel used for convolution. base_channels : int Number of channels used in each conv layer. activation : Callable Function to apply between layers. lin_dim1: int Dimensionality of the first linear layer. lin_dim2: int Dimensionality of the second linear layer. Example ------- >>> inputs = torch.rand([1, 1, 100, 257]) >>> model = MetricDiscriminator() >>> outputs = model(inputs) >>> outputs.shape torch.Size([1, 1]) )rC2 cst|dd|_tjddd|_td|tj|d|_t|tj|d|_ t|tj|d|_ t|tj|d|_ t||d|_ t||d|_ t|dd|_dS) Nrrg{Gz?) num_featuresmomentum)r kernel_size)r )r r )r"r#r%r BatchNorm2dBNrConv2dconv1conv2conv3conv4Linear1Linear2Linear3)r1rJ base_channelsr%lin_dim1lin_dim2r5rrr#s$   zMetricDiscriminator.__init__cCs||}||}||}||}||}||}||}||}t|d}||}||}||}||}| |}|S)r7)r) rNr%rOrPrQtorchmeanrRrSrT)r1r9r:rrrr<s              zMetricDiscriminator.forward) r=r>r?r@rr$r#r<rArrr5rrBks rB) r@rYrtorch.nn.utilsr speechbrainr&LinearrModulerrBrrrrs   K