o %ݫi`%@sdZddlZddlmZddlmZddlmZmZddl m Z ddl m Z ddl mZdd lmZGd d d e ZGd d d ejjZGdddejjZdS)zuThe SpeechBrain implementation of ContextNet by https://arxiv.org/pdf/2005.03191.pdf Authors * Jianyuan Zhong 2020 N)Dropout)Swish)Conv1dDepthwiseSeparableConv1d) Sequential)Linear) BatchNorm1d) AdaptivePoolcsBeZdZdZdddddddddd eejjedffd d Z Z S) ContextNetaaThis class implements the ContextNet. Reference paper: https://arxiv.org/pdf/2005.03191.pdf Arguments --------- input_shape : tuple Expected shape of the inputs. out_channels : int Number of output channels of this model (default 640). conv_channels : Optional (list[int]) Number of output channels for each of the contextnet block. If not provided, it will be initialized as the default setting of above mentioned paper. kernel_size : int Kernel size of convolution layers (default 3). strides: Optional (list[int]) Striding factor for each context block. This stride is applied at the last convolution layer at each context block. If not provided, it will be initialize as the default setting of above paper. num_blocks : int Number of context block (default 21). num_layers : int Number of depthwise convolution layers for each context block (default 5). inner_dim : int Inner dimension of bottle-neck network of the SE Module (default 12). alpha : float The factor to scale the output channel of the network (default 1). beta : float Beta to scale the Swish activation (default 1). dropout : float Dropout (default 0.15). activation : torch class Activation function for each context block (default Swish). se_activation : torch class Activation function for SE Module (default torch.nn.Sigmoid). norm : torch class Normalization to regularize the model (default BatchNorm1d). residuals : Optional (list[bool]) Whether to apply residual connection at each context block (default None). Example ------- >>> inp = torch.randn([8, 48, 40]) >>> block = ContextNet(input_shape=inp.shape, num_blocks=14) >>> out = block(inp) >>> out.shape torch.Size([8, 6, 640]) iN 333333?cs`tj|d|durgdgddgd}|dur,dg|}d|d<d|d<d|d <|dur5d g|}|jt|d |d d |j|dd t| trV|j| | dd n|j| dd t|D]#}t||| }|jt||||||| | | | |||d|d qb|jt||dd |j|dd t| tr|j| | dd dS|j| dd dS)N input_shape i r Tr conv_start) layer_name norm_start act_startblock_) out_channels kernel_size num_layers inner_dimstridebetadropout activation se_activationnormresidualrconv_endnorm_endact_end) super__init__appendr isinstancerrangeintContextNetBlock)selfrr conv_channelsrstrides num_blocksr r!alphar#r$r%r&r' residualsichannels __class__W/home/ubuntu/.local/lib/python3.10/site-packages/speechbrain/lobes/models/ContextNet.pyr-Cs^     zContextNet.__init__) __name__ __module__ __qualname____doc__rtorchnnSigmoidrr- __classcell__r=r=r;r>r s"2r cs2eZdZdZejjeffdd ZddZ Z S)SEmodulea|This class implements the Squeeze-and-Excitation module. Arguments --------- input_shape : tuple Expected shape of the inputs. inner_dim : int Inner dimension of bottle-neck network of the SE Module (default 12). activation : torch class Activation function for SE Module (default torch.nn.Sigmoid). norm : torch class Normalization to regularize the model (default BatchNorm1d). Example ------- >>> inp = torch.randn([8, 120, 40]) >>> net = SEmodule(input_shape=inp.shape, inner_dim=64) >>> out = net(inp) >>> out.shape torch.Size([8, 120, 40]) cst||_||_||_|\}}}t|d|_|jjt|ddd|j|j|j|t d|_ tt |d|jd|t |j|d||_ dS)Nrr)rrr") input_size n_neurons) r,r-r!r'r%rconvr.rr avg_poolr bottleneck)r3rr!r%r'bztchnr;r=r>r-s$      zSEmodule.__init__cCs@|j\}}}||}||}||}|d|d}||S):Processes the input tensor x and returns an output tensor.r)shaperKrLrMrepeat)r3xrNrOrPavgcontextr=r=r>forwards    zSEmodule.forward) r?r@rArBrCrDrErr-rWrFr=r=r;r>rGs rGcsDeZdZdZdddeejjedffdd Z ddZ d d Z Z S) r2a&This class implements a block in ContextNet. Arguments --------- out_channels : int Number of output channels of this model (default 640). kernel_size : int Kernel size of convolution layers (default 3). num_layers : int Number of depthwise convolution layers for this context block (default 5). inner_dim : int Inner dimension of bottle-neck network of the SE Module (default 12). input_shape : tuple Expected shape of the inputs. stride : int Striding factor for this context block (default 1). beta : float Beta to scale the Swish activation (default 1). dropout : float Dropout (default 0.15). activation : torch class Activation function for this context block (default Swish). se_activation : torch class Activation function for SE Module (default torch.nn.Sigmoid). norm : torch class Normalization to regularize the model (default BatchNorm1d). residual : bool Whether to apply residual connection at this context block (default None). Example ------- >>> inp = torch.randn([8, 120, 40]) >>> block = ContextNetBlock(256, 3, 5, 12, input_shape=inp.shape, stride=2) >>> out = block(inp) >>> out.shape torch.Size([8, 60, 256]) rrTc st| |_t|d|_t|D]} |jjt||| |dkr"|ndd|j| qt|j || | d|_ t ||_ d|_ | rYt|d|_ |j jt|d|d|j | t| trd| ||_n| |_|dS)Nrr)r")rr!r%r'r )rr")r,r-r(rConvsr0r.rrGget_output_shapeSErdrop reduced_covrr/rr% _reset_params)r3rrr r!rr"r#r$r%r&r'r(r9r;r=r>r-s:         zContextNetBlock.__init__cCs<||}||}|jr|||}||}||S)rQ)rXrZr\r%r[)r3rToutr=r=r>rW$s    zContextNetBlock.forwardcCs,|D]}|dkrtjj|qdS)Nr) parametersdimrCrDinitkaiming_normal_)r3pr=r=r>r]-s  zContextNetBlock._reset_params) r?r@rArBrrCrDrErr-rWr]rFr=r=r;r>r2s-2 r2)rBrCtorch.nnrspeechbrain.nnet.activationsrspeechbrain.nnet.CNNrrspeechbrain.nnet.containersrspeechbrain.nnet.linearrspeechbrain.nnet.normalizationrspeechbrain.nnet.poolingr r rDModulerGr2r=r=r=r>s      z>