miZdZddlZddlmZmZddlZddlmZddl m Z ddl m Z ddl mZGd d ejj ZGd d ejj Z dd ej&deededededeef dZ dd ej&dededededeef dZy)z$ FIR windowed sinc lowpass filters. N)SequenceOptional) functional)sinc) fft_conv1d simple_reprc VeZdZdZ d deededededeef fd Z dZ d Z xZ S) LowPassFiltersa Bank of low pass filters. Note that a high pass or band pass filter can easily be implemented by substracting a same signal processed with low pass filters with different frequencies (see `julius.bands.SplitBands` for instance). This uses a windowed sinc filter, very similar to the one used in `julius.resample`. However, because we do not change the sample rate here, this filter can be much more efficiently implemented using the FFT convolution from `julius.fftconv`. Args: cutoffs (list[float]): list of cutoff frequencies, in [0, 0.5] expressed as `f/f_s` where f_s is the samplerate and `f` is the cutoff frequency. The upper limit is 0.5, because a signal sampled at `f_s` contains only frequencies under `f_s / 2`. stride (int): how much to decimate the output. Keep in mind that decimation of the output is only acceptable if the cutoff frequency is under `1/ (2 * stride)` of the original sampling rate. pad (bool): if True, appropriately pad the input with zero over the edge. If `stride=1`, the output will have the same length as the input. zeros (float): Number of zero crossings to keep. Controls the receptive field of the Finite Impulse Response filter. For lowpass filters with low cutoff frequency, e.g. 40Hz at 44.1kHz, it is a bad idea to set this to a high value. This is likely appropriate for most use. Lower values will result in a faster filter, but with a slower attenuation around the cutoff frequency. fft (bool or None): if True, uses `julius.fftconv` rather than PyTorch convolutions. If False, uses PyTorch convolutions. If None, either one will be chosen automatically depending on the effective filter size. ..warning:: All the filters will use the same filter size, aligned on the lowest frequency provided. If you combine a lot of filters with very diverse frequencies, it might be more efficient to split them over multiple modules with similar frequencies. ..note:: A lowpass with a cutoff frequency of 0 is defined as the null function by convention here. This allows for a highpass with a cutoff of 0 to be equal to identity, as defined in `julius.filters.HighPassFilters`. Shape: - Input: `[*, T]` - Output: `[F, *, T']`, with `T'=T` if `pad` is True and `stride` is 1, and `F` is the numer of cutoff frequencies. >>> lowpass = LowPassFilters([1/4]) >>> x = torch.randn(4, 12, 21, 1024) >>> list(lowpass(x).shape) [1, 4, 12, 21, 1024] cutoffsstridepadzerosfftc t |t||_t |jdkr t dt |jdkDr t d||_||_||_ t|t |jDcgc] }|dkDs | c}z dz |_ ||jdkD}||_ tjd|jzdzd }tj|j |jdz}g} |D]l} | dk(rtj |} n>d| z|zt#d| zt$j&z|zz} | | j)z} | j+| n|j-d tj.| dddfycc}w) Nrz(Minimum cutoff must be larger than zero.g?z'A cutoff above 0.5 does not make sense. rF)periodicfilters)super__init__listr min ValueErrormaxrrrint half_sizertorch hann_windowarange zeros_likermathpisumappendregister_bufferstack) selfr rrrrcwindowtimercutofffilter_ __class__s H/var/www/stems/demucs_env/lib/python3.12/site-packages/julius/lowpass.pyrzLowPassFilters.__init__Hs G} t|| q GH H t|| s "FG G  UST\\)KQU!)K%LLqPQ ;..2%C""1t~~#5#9EJ||T^^OT^^a-?@ $F{**40f*v-QZ$''5ID5P0QQ7;;=( NN7 # $ Y G(>> lowpass = LowPassFilter(1/4, stride=2) >>> x = torch.randn(4, 124) >>> list(lowpass(x).shape) [4, 62] r-rrrrcLt|t|g|||||_yrB)rrr _lowpasses)r)r-rrrrr/s r0rzLowPassFilter.__init__s$ (&63sKr@c4|jjdSNr)rSr rCs r0r-zLowPassFilter.cutoffs&&q))r@c.|jjSrB)rSrrCs r0rzLowPassFilter.strides%%%r@c.|jjSrB)rSrrCs r0rzLowPassFilter.pad"""r@c.|jjSrB)rSrrCs r0rzLowPassFilter.zeross$$$r@c.|jjSrB)rSrrCs r0rzLowPassFilter.fftrXr@c*|j|dSrU)rS)r)r=s r0r?zLowPassFilter.forwardsu%a((r@ct|SrBr rCs r0rDzLowPassFilter.__repr__rEr@rF)rHrIrJrKrLrrMrrpropertyr-rrrrr?rDrNrOs@r0rQrQvs DH9=LuLcLDLL(0L **&&##%%##)!r@rQr=r rrrrcJt|||||j||S)z[ Functional version of `LowPassFilters`, refer to this class for more information. )r to)r=r rrrrs r0lowpass_filtersr`s* F>'63s ; > >u Ee LLr@r-c*t||g||||dS)z Same as `lowpass_filters` but with a single cutoff frequency. Output will not have a dimension inserted in the front. r)r`)r=r-rrrrs r0lowpass_filterrbs  56(FC DQ GGr@rF)rKr#typingrrrtorch.nnrr7corerfftconvrutilsr nnModuler rQTensorrLrrMr`rbr@r0rls % $a!UXX__a!H,!EHHOO,!`26<@M5<<M8E?MM*.M M+3D>M15;?H%,,HHH)-HH*24.Hr@