o %ݫi6@sdZddlZddlZddlmZddlmmZddlm Z ddl m Z dddZ ddd ZGd d d ejZGd d d ejZGdddejZGdddejZGdddejjZdS)zNResNet PreActivated for speaker verification Authors * Mickael Rouvier 2022 N)Linear) BatchNorm1dcCstj||d|dddS)z#2D convolution with kernel_size = 3rF) kernel_sizestridepaddingbiasnnConv2d in_planes out_planesrrS/home/ubuntu/.local/lib/python3.10/site-packages/speechbrain/lobes/models/ResNet.pyconv3x3srcCstj||d|ddS)z#2D convolution with kernel_size = 1rFrrr r r rrrconv1x1s rcs0eZdZdZdejffdd ZddZZS)SEBlockaAn implementation of Squeeze-and-Excitation Block. Arguments --------- channels : int The number of channels. reduction : int The reduction factor of channels. activation : Callable The function to apply between layers. Example ------- >>> inp_tensor = torch.rand([1, 64, 80, 40]) >>> se_layer = SEBlock(64) >>> out_tensor = se_layer(inp_tensor) >>> out_tensor.shape torch.Size([1, 64, 80, 40]) rcsNtt|td|_tt||||t|||t|_ dS)Nr) superr__init__r AdaptiveAvgPool2davg_pool SequentialrSigmoidfc)selfchannels reduction activation __class__rrr;s  zSEBlock.__init__cCs@|\}}}}||||}||||dd}||S)^Intermediate step. Processes the input tensor x and returns an output tensor. r)sizerviewr)rxbc_yrrrforwardGszSEBlock.forward __name__ __module__ __qualname____doc__r ReLUrr+ __classcell__rrr!rr&s rcs2eZdZdZddejffdd ZddZZS) BasicBlockaAn implementation of ResNet Block. Arguments --------- in_channels : int Number of input channels. out_channels : int The number of output channels. stride : int Factor that reduce the spatial dimensionality downsample : torch function A function for downsample the identity of block when stride != 1 activation : torch class A class for constructing the activation layers. Example ------- >>> inp_tensor = torch.rand([1, 64, 80, 40]) >>> layer = BasicBlock(64, 64, stride=1) >>> out_tensor = layer(inp_tensor) >>> out_tensor.shape torch.Size([1, 64, 80, 40]) rNcsptt|||_t||_t||||_t||_ t|||_ t||_ t |||_ ||_||_dSN)rr3rr r BatchNorm2dbn1rconv1bn2conv2bn3rconv3 downsampler)r in_channels out_channelsrr<r r!rrrjs      zBasicBlock.__init__cCs~|}||}||}||}||}||}||}||}||}||}|jdur9||}||7}|Sr#N)r6r r7r8r9r:r;r<rr&residualoutrrrr+s           zBasicBlock.forwardr,rrr!rr3Qsr3cs4eZdZdZdddejffdd ZddZZS) SEBasicBlockaAn implementation of Squeeze-and-Excitation ResNet Block. Arguments --------- in_channels : int Number of input channels. out_channels : int The number of output channels. reduction : int The reduction factor of channels. stride : int Factor that reduce the spatial dimensionality downsample : torch function A function for downsample the identity of block when stride != 1 activation : torch class A class for constructing the activation layers. Example ------- >>> inp_tensor = torch.rand([1, 64, 80, 40]) >>> layer = SEBasicBlock(64, 64, stride=1) >>> out_tensor = layer(inp_tensor) >>> out_tensor.shape torch.Size([1, 64, 80, 40]) rNcs|tt|||_t||_t||||_t||_ t|||_ t||_ t |||_ ||_||_t|||_dSr4)rrCrr r r5r6rr7r8r9r:rr;r<rrse)rr=r>rrr<r r!rrrs      zSEBasicBlock.__init__cCs|}||}||}||}||}||}||}||}||}||}||}|jdur>||}||7}|Sr?) r6r r7r8r9r:r;rDr<r@rrrr+s            zSEBasicBlock.forwardr,rrr!rrCsrCcs^eZdZdZddejjgdgdgddffdd Zdd d Zdd dZ dddZ Z S)ResNetaAn implementation of ResNet Arguments --------- input_size : int Expected size of the input dimension. device : str Device used, e.g., "cpu" or "cuda". activation : torch class A class for constructing the activation layers. channels : list of ints List of number of channels used per stage. block_sizes : list of ints List of number of groups created per stage. strides : list of ints List of stride per stage. lin_neurons : int Number of neurons in linear layers. Example ------- >>> input_feats = torch.rand([2, 400, 80]) >>> compute_embedding = ResNet(lin_neurons=256) >>> outputs = compute_embedding(input_feats) >>> outputs.shape torch.Size([2, 256]) Pcpu)rHrI)rr)rrLrLrIc stt|dks Jt|dksJt|dksJt||d|d|d|d}tjd|dddddd|_t|d|_ ||_ |j |d|d|d|dd|_ |j |d|d|d|dd|_ |j|d|d|d|dd|_|j|d|d|d|dd|_tjd||d |_ttj|d |d dd ttd tjd |d |dd tjdd |_td||d ||_tj||_|D]*} t| tjrtjj| j d ddqt| tjr tj!| j dtj!| j"dqdS)NrJrrrLrF)r )rrH)rdimfan_outrelu)mode nonlinearity)#rrlenmathceilr r r7r5r6 activation1_make_layer_selayer1layer2 _make_layerlayer3layer4torchr norm_statsrConv1dr1Softmax attentionrfc_embed norm_embedmodules isinstanceinitkaiming_normal_weight constant_r ) r input_sizedevicer r block_sizesstrides lin_neurons input_outmr!rrrsT "   zResNet.__init__rc Cs|d}|dks ||krttj||d|ddt|}g}|t||d||td|D] }|t||dq-tj|S)aConstruct the squeeze-and-excitation block layer. Arguments --------- in_channels : int Number of input channels. out_channels : int The number of output channels. block_num: int Number of ResNet blocks for the network. stride : int Factor that reduce the spatial dimensionality. Default is 1 Returns ------- se_block : nn.Sequential Squeeze-and-excitation block NrFr)r rr r5appendrCrangerr=r> block_numrr<layersirrrrXCs&  zResNet._make_layer_sec Csxd}|dks ||krttj||d|ddt|}g}|t||||td|D] }|t||q,tj|S)a Construct the ResNet block layer. Arguments --------- in_channels : int Number of input channels. out_channels : int The number of output channels. block_num: int Number of ResNet blocks for the network. stride : int Factor that reduce the spatial dimensionality. Default is 1 Returns ------- block : nn.Sequential ResNet block NrFr)r rr r5rrr3rsrtrrrr[ls"  zResNet._make_layerNcCs|d}||}||}||}||}||}||}||}|dd}| dd}| |}t j ||dd}t t j |d|dd|djdd}t j||gdd}||}||}||}|S)aMReturns the embedding vector. Arguments --------- x : torch.Tensor Tensor of shape (batch, time, channel). lengths : torch.Tensor Corresponding relative lengths of the inputs. Returns ------- x : torch.Tensor The embedding vector. rrLrrNgh㈵>)min) unsqueezer7r6rWrYrZr\r] transposeflattenrbr^sumsqrtclampcatr_rcrd)rr&lengthswmusgrrrr+s$           ,   zResNet.forwardrr4) r-r.r/r0r^r r1rrXr[r+r2rrr!rrEs ; )(rEcs2eZdZdZ    d fdd Zdd ZZS) ClassifieraThis class implements the cosine similarity on the top of features. Arguments --------- input_size : int Expected size of the inputs. device : str Device used, e.g., "cpu" or "cuda". lin_blocks : int Number of linear layers. lin_neurons : int Number of neurons in linear layers. out_neurons : int Number of classes. Example ------- >>> classify = Classifier(input_size=2, lin_neurons=2, out_neurons=2) >>> outputs = torch.tensor([ [1., -1.], [-9., 1.], [0.9, 0.1], [0.1, 0.9] ]) >>> outputs = outputs.unsqueeze(1) >>> cos = classify(outputs) >>> (cos < -1.0).long().sum() tensor(0) >>> (cos > 1.0).long().sum() tensor(0) rGrrIcsntt|_t|D]}|jt|dt||dg|}qt t j |||d|_ tj |j dS)N)rk)rk n_neurons)rl)rrr ModuleListblocksrsextend _BatchNorm1dr Parameterr^ FloatTensorrirgxavier_uniform_)rrkrl lin_blocksro out_neurons block_indexr!rrrs    zClassifier.__init__cCs>|jD]}||}qtt|dt|j}|dS)zReturns the output probabilities over speakers. Arguments --------- x : torch.Tensor Torch tensor. Returns ------- x : torch.Tensor Output probabilities over speakers. r)rFlinear normalizesqueezeriry)rr&layerrrrr+s   zClassifier.forward)rGrrIr)r-r.r/r0rr+r2rrr!rrsrr)r0rUr^torch.nnr torch.nn.functional functionalrspeechbrain.nnet.linearrspeechbrain.nnet.normalizationrrrrModulerr3rCrErrrrrs     +IQT