o i @sZdZddlZddlZddlmZddlmmZeeej j Z Gdddej Z dS)zLightweight Convolution Module.N)nncs.eZdZdZ  dfdd ZddZZS)LightweightConvolutionaLightweight Convolution layer. This implementation is based on https://github.com/pytorch/fairseq/tree/master/fairseq Args: wshare (int): the number of kernel of convolution n_feat (int): the number of features dropout_rate (float): dropout_rate kernel_size (int): kernel size (length) use_kernel_mask (bool): Use causal mask or not for convolution kernel use_bias (bool): Use bias term or not. Fc stt|||dksJ||_||_||_||_t|d|_t ||d|_ t |||_ t |_t t|jd|dd|_||_|jrVt t||_t|jt|d}t|jt|dd}tj||fddd|_dS)z(Construct Lightweight Convolution layer.rdimN)superr__init__wshareuse_kernel_mask dropout_rate kernel_sizeint padding_sizerLinearlinear1linear2GLUact ParametertorchTensoruniform_weightuse_biasbiaszerosonescat unsqueeze kernel_mask) selfr n_featr rr r kernel_mask0 kernel_mask1 __class__]/home/ubuntu/.local/lib/python3.10/site-packages/funasr/models/transformer/utils/lightconv.pyr s"   zLightweightConvolution.__init__c Cs|}|\}}}|j} ||}||}|ddd| |}tj|j |j |j d} |j rD|j |j|_ | |j dktd} tj| dd} tj|| |j|jd|||}|jrh||jddd}|dd}|d ur|j s|dd }||d kd}||}|S) a*Forward of 'Lightweight Convolution'. This function takes query, key and value but uses only query. This is just for compatibility with self-attention layer (attention.py) Args: query (torch.Tensor): (batch, time1, d_model) input tensor key (torch.Tensor): (batch, time2, d_model) NOT USED value (torch.Tensor): (batch, time2, d_model) NOT USED mask (torch.Tensor): (batch, time1, time2) mask Return: x (torch.Tensor): (batch, time1, d_model) output rrr)traininggz-infr)paddinggroupsNr)sizer rr transpose contiguousviewFdropoutrr r*r r!todevice masked_fillfloatsoftmaxconv1drrrr) r"querykeyvaluemaskxBTCHrr(r(r)forward?s(      zLightweightConvolution.forward)FF)__name__ __module__ __qualname____doc__r rC __classcell__r(r(r&r)r s #r)rGnumpyrrtorch.nn.functional functionalr2r7finfofloat32min MIN_VALUEModulerr(r(r(r)s