VhP dZddlmZddlZ ddlmZn #e$rdZYnwxYwddlm Z ddl m Z m Z ddlmZdejfd Zdejfd Zdejd efd Zdejd efd ZeeZeZneZeZdejdejfdZ ddejdejdeejdededef dZGddejjZdS)ab Implementation of a FFT based 1D convolution in PyTorch. While FFT is used in CUDNN for small kernel sizes, it is not the case for long ones, e.g. 512. This module implements efficient FFT based convolutions for such convolutions. A typical application is for evaluationg FIR filters with a long receptive field, typically evaluated with a stride of 1. )OptionalN) functional)pad_tounfold) simple_reprxcVtj|d}tj|SNdim)new_fftrffttorch view_as_real)r zs T/var/www/html/movieo_spanner_bot/venv/lib/python3.11/site-packages/julius/fftconv.py _new_rfftrs' QBA  a  c,tj|dS)Nr)rr)r s r _old_rfftrs :a  rlengthc6tj|d|f}|S)Nr) signal_sizes)rirfft)r rresults r _old_irfftr"s [AVI 6 6 6F MrcXtj|}tj||dSr )rview_as_complexrr)r rs r _new_irfftr!'s) a  A =F + + ++rabc Bd}tjtj||d|dtj||d|dztj||d|dtj||d|dz gdS)z Given a and b two tensors of dimension 4 with the last dimension being the real and imaginary part, returns a multiplied by the conjugate of b, the multiplication being with respect to the second dimension. zbcft,dct->bdft).r).rr r )rstackeinsum)r"r#ops r_compl_mul_conjugater(4s B ; R6AfI..b!F)QvY1W1WW R6AfI..b!F)QvY1W1WW    rinputweightbiasstridepadding block_ratioctj|||f}|j\}}}|j\} } } || krtd| d|d|dkrtdt t | |z|} | | z dz} t || }t|}t|| | }t|}t||}t|| }|dd| dzf}| || d}|ddd|f}|| z |zdz}|dd|f}|||dddfz }|S) a Same as `torch.nn.functional.conv1d` but using FFT for the convolution. Please check PyTorch documentation for more information. Args: input (Tensor): input signal of shape `[B, C, T]`. weight (Tensor): weight of the convolution `[D, C, K]` with `D` the number of output channels. bias (Tensor or None): if not None, bias term for the convolution. stride (int): stride of convolution. padding (int): padding to apply to the input. block_ratio (float): can be tuned for speed. The input is splitted in chunks with a size of `int(block_ratio * kernel_size)`. Shape: - Inputs: `input` is `[B, C, T]`, `weight` is `[D, C, K]` and bias is `[D]`. - Output: `(*, T)` ..note:: This function is faster than `torch.nn.functional.conv1d` only in specific cases. Typically, the kernel size should be of the order of 256 to see any real gain, for a stride of 1. ..Warning:: Dilation and groups are not supported at the moment. This function might use more memory than the default Conv1d implementation. z5Input should be at least as large as the kernel size z, but it is only z samples long.rz#Block ratio must be greater than 1..Nr ) Fpadshape RuntimeErrorminintr_rfftrr(_irfftreshape)r*r+r,r-r.r/batchchannelsr out_channels_ kernel_size block_size fold_strideweight_zframesframes_zout_zout target_lengths r fft_conv1drGGsB E%'7+ , ,E#kE8V#)< L![ DS^DD-3DDDEE EQ@AAA#kK788&AAJ{*Q.K FJ ' 'FV}}HE:{ 3 3FV}}H 8 4 4E  # #C c$[L1$$$ %C ++e\2 . .C c88V8m Ck)f4q8M c>M>! "C  tAAAtG} Jrc \eZdZdZ ddedededed ed ef fd Zd ejfd Z dZ xZ S) FFTConv1da Same as `torch.nn.Conv1d` but based on `fft_conv1d`. Please check PyTorch documentation for more information. Args: in_channels (int): number of input channels. out_channels (int): number of output channels. kernel_size (int): kernel size of convolution. stride (int): stride of convolution. padding (int): padding to apply to the input. bias (bool): if True, use a bias term. ..note:: This module is faster than `torch.nn.Conv1d` only in specific cases. Typically, `kernel_size` should be of the order of 256 to see any real gain, for a stride of 1. ..warning:: Dilation and groups are not supported at the moment. This module might use more memory than the default Conv1d implementation. >>> fftconv = FFTConv1d(12, 24, 128, 4) >>> x = torch.randn(4, 12, 1024) >>> print(list(fftconv(x).shape)) [4, 24, 225] rrT in_channelsr<r>r-r.r,ct||_||_||_||_||_tj ||||}|j |_ |j |_ dS)N)r,) super__init__rJr<r>r-r.rnnConv1dr+r,) selfrJr<r>r-r.r,conv __class__s rrMzFFTConv1d.__init__sp &(&  x{L+DQQk I rr*cPt||j|j|j|jS)N)rGr+r,r-r.)rPr*s rforwardzFFTConv1d.forwards* 4; 4; FF Frc6t|d|jduiS)Nr,) overrides)rr,)rPs r__repr__zFFTConv1d.__repr__s!4FDIT4I+JKKKKr)rrT) __name__ __module__ __qualname____doc__r6boolrMrTensorrTrW __classcell__)rRs@rrIrIs6BF  C s   +. :>      FU\FFFFLLLLLLLrrI)Nrrr))r[typingrr torch.fftfftr ImportErrortorch.nnrr1corerrutilsrr]rrr6rr!r7r8r(floatrGrNModulerIrrris GGG$$$$$$ !!!!! %, ,%,,,,,,  ? E FF E FELU\*NO@@|@%*\@u|$@58@GJ@@@@@F-L-L-L-L-L-L-L-L-L-Ls