o iw @s`ddlZddlZddlmZmZmZddlmZGdddeZ ddZ Gdd d ej j Z dS) N)OptionalTupleUnion pad_to_lencs eZdZdZfddZZS)TooShortUttErrorzRaised when the utt is too short for subsampling. Args: message (str): Message for error catch actual_size (int): the short size that cannot pass the subsampling limit (int): the limit size for subsampling cst|||_||_dS)z/Construct a TooShortUttError for error handler.N)super__init__ actual_sizelimit)selfmessager r  __class__[/home/ubuntu/.local/lib/python3.10/site-packages/funasr/models/rwkv_bat/rwkv_subsampling.pyr s  zTooShortUttError.__init__)__name__ __module__ __qualname____doc__r __classcell__rrrrr s rcCs\t|tr |dkr dSt|tr|dkrdSt|tr!|dkr!dSt|tr,|dkr,dSd S) z4Check if the utterance is too short for subsampling.)Tr)Tr )Tr)Tr)F) isinstanceConv2dSubsampling2Conv2dSubsamplingConv2dSubsampling6Conv2dSubsampling8)inssizerrrcheck_short_uttsr#c seZdZdZ   ddedeeefdeded eed df fd d Zd e j dee j dee j d ee j e j ffddZ de j d e j fddZ ded efddZ ZS) RWKVConvInputzStreaming ConvInput module definition. Args: input_size: Input size. conv_size: Convolution size. subsampling_factor: Subsampling factor. output_size: Block output dimension. rN input_size conv_sizesubsampling_factorconv_kernel_size output_sizereturnc st|dkr|\}}}tjtjjd||d|dddtjtjj|||ddg|dddtjtjj|||d|dddtjtjj|||ddg|dddtjtjj|||d|dddtjtjj|||ddg|dddtj |_||dd} d|_d|_ |j |_ n|\}}}t |d} tjtjjd||d|dddtjtjj|||| dg|dddtjtjj|||d|dddtjtjj|||ddg|dddtjtjj|||d|dddtjtjj|||d|dddtj |_||dd} ||_|j |_ | |_ d|_ |durAtj| ||_||_dSd|_| |_dS)zConstruct a ConvInput object.)stridepaddingrN)rr torchnn SequentialConv2dReLUconvr(stride_1create_new_vgg_maskcreate_new_maskintmin_frame_lengthLinearoutputr*) r r&r'r(r)r* conv_size1 conv_size2 conv_size3 output_projkernel_1rrrr 3s       /        /   zRWKVConvInput.__init__xmask chunk_sizec st|dur||}t|dd}|\}}}|d}|dur^t||jt t |||jt fdd|}t |}t j|dd}||j}|||d||j|}||}|\} } } }|dur|dd|d| |ddd|ddf}n|dd|d| |}|jdur||}||ddd|fddd|dffS) a'Encode input sequences. Args: x: ConvInput input sequences. (B, T, D_feats) mask: Mask of input sequences. (B, 1, T) Returns: x: ConvInput output sequences. (B, sub(T), D_out) mask: Mask of output sequences. (B, 1, sub(T)) Nrr,cs t|dS)Nr,r)inputsmax_input_lengthrrs z'RWKVConvInput.forward..)dimr-r)r8maxeqsumr" unsqueezer9r(mathceilfloatmaplistr0stackviewr5 transpose contiguousr<) r rBrCrDolensbtfN_chunks_crrFrforwards4   6  .zRWKVConvInput.forwardcCs8|jdkr|dddddfdddd|jfS|S)zCreate a new mask for VGG output sequences. Args: mask: Mask of input sequences. (B, T) Returns: mask: Mask of output sequences. (B, sub(T)) r,Nr-)r(r6)r rCrrrr7s *z!RWKVConvInput.create_new_vgg_maskr"cCs ||jS)zReturn the original size before subsampling for a given size. Args: size: Number of frames after subsampling. Returns: : Number of frames before subsampling. )r()r r"rrrget_size_before_subsamplings z)RWKVConvInput.get_size_before_subsampling)r%rN)rrrrr9rrrr r0Tensorr^r7r_rrrrrr$*s<    + r$)rNr0typingrrr*funasr.models.transformer.utils.nets_utilsr Exceptionrr#r1Moduler$rrrrs