hb 8SSKrSSKrSSKJrJrJrJr SSKrSSKJr SSK J r "SS\RR5r "SS\RR5r"S S \RR5r"S S \RR5r"S S\RR5r"SS\RR5r"SS\R5r"SS\R5rS\R*S\S\S\R*4SjrSrS'S\R*S\S\S\S\R*4 SjjrS(S \R*S\S!\S\S\R*4 S"jjrS#\\S\4S$jrS#\\S\4S%jrS#\\S\4S&jrg))N)AnyDictListOptional)nn) functionalc ^\rSrSrSrS S\S\S\S\4U4Sjjjr\ S\ R4S j5r S \ RS\ R4S jr S rU=r$)_ScaledEmbedding#a.Make continuous embeddings and boost learning rate Args: num_embeddings (int): number of embeddings embedding_dim (int): embedding dimensions scale (float, optional): amount to scale learning rate (Default: 10.0) smooth (bool, optional): choose to apply smoothing (Default: ``False``) num_embeddings embedding_dimscalesmoothc>[TU]5 [R"X5UlU(a[ R "URRRSS9nU[ R"SUS-5R5SS2S4- nXPRRRSS&URR=RU-slX0l g)Nrdim) super__init__r Embedding embeddingtorchcumsumweightdataarangesqrtr)selfr r rrr __class__s R/var/www/auris/envauris/lib/python3.13/site-packages/torchaudio/models/_hdemucs.pyr_ScaledEmbedding.__init__-s nD \\$.."7"7"<"^[TU]5 U c0n Sn US:XaU4Sjn U (aUS-OSn [RnXplX0lX@lX`lXlU(aUS/nUS/nU S/n [RnU"XX4U 5Ul U "U5Ul UR(aO[R"5Ul [R"5Ul[R"5UlgU"USU-SSU --SU 5Ul U "SU-5Ul[!U40U D6Ulg)Nc,[R"5$r%rIdentityds r %_HEncLayer.__init__..m BKKMr" group_normc2>[R"TU5$r%r GroupNormrMrAs r rNrOo [! r?r@rArBrCrDrErFrGnorm_fnpad_valklassrs ` r r_HEncLayer.__init__\s   H)  $**3*3*3X,,x /E,QVQ]Q],,r"r;c^\rSrSrSrSS\S\S\S\S\S\S \S \S \S \S \\ \\ 4S\4U4Sjjjr S\ RS\\ R4SjrSrU=r$) _HDecLayera{Decoder layer. This used both by the time and the frequency branches. Args: chin (int): number of input channels. chout (int): number of output channels. last (bool, optional): whether current layer is final layer (Default: ``False``) kernel_size (int, optional): Kernel size for encoder (Default: 8) stride (int): Stride for encoder layer (Default: 4) norm_groups (int, optional): number of groups for group norm. (Default: 1) empty (bool, optional): used to make a layer with just the first conv. this is used before merging the time and freq. branches. (Default: ``False``) freq (bool, optional): boolean for whether conv layer is for frequency (Default: ``True``) norm_type (str, optional): Norm type, either ``group_norm `` or ``none`` (Default: ``group_norm``) context (int, optional): context size for the 1x1 conv. (Default: 1) dconv_kw (Dict[str, Any] or None, optional): dictionary of kwargs for the DConv class. (Default: ``None``) pad (bool, optional): true to pad the input. Padding is done so that the output size is always the input size / stride. (Default: ``True``) r=r>lastr?r@rArBrCrDrErFrGc >^[TU]5 U c0n Sn U S:XaU4Sjn U (aXE- S-S:wa [S5eXE- S-n OSn XlX0lXlXlXplXPlX@l [Rn[RnU(a(US/nUS/n[Rn[RnU"XXE5UlU "U5UlUR(a5[R""5Ul[R""5UlgU"USU-SSU --SU 5UlU "SU-5Ulg)Nc,[R"5$r%rJrLs r rN%_HDecLayer.__init__..rPr"rQc2>[R"TU5$r%rSrUs r rNrrVr"rXrz#Kernel size and stride do not alignr)rrrlrGrrCr=rBr@r?rrYConvTranspose1drZConvTranspose2dconv_trr^rKr]r\)rr=r>rr?r@rArBrCrDrErFrGrarcklass_trrs ` r r_HDecLayer.__init__s*   H)  $USURURU-24nURSU:wa [S5eUR(d[ R"U5nX4$) aForward pass for decoding layer. Size depends on whether frequency or time Args: x (torch.Tensor): tensor input of shape `(B, C, F, T)` for frequency and shape `(B, C, T)` for time skip (torch.Tensor, optional): on first layer, separate frequency and time branches using param (default: ``None``) length (int): Size of tensor for output Returns: (Tensor, Tensor): Tensor output tensor after decoder layer of shape `(B, C, F * stride, T)` for frequency domain except last frequency layer shape is `(B, C, kernel_size, T)`. Shape is `(B, C, stride * T)` for time domain. Tensor contains the output just before final transposed convolution, which is used when the freq. and time branch separate. Otherwise, does not matter. Shape is `(B, C, F, T)` for frequency and `(B, C, T)` for time. rhrgrrNz%Skip must be none when empty is true..z'Last index of z must be equal to length)rCrrirjr=rBrkrpr\r]rlr^rrGrrm) rr'rlengthrqrrrtrvrws r r+_HDecLayer.forwards. 99AggGA!q))R+AzzAdjja1q9AA !HII JJt||A ' 99xxc488txxi/23#txx$((V"3334Awwr{f$ !JKKyyq At r") r=rrBrCr?rr\r^rGr]r@) FryrWrFTrQrNTrzr9s@r r}r}s,%-10+0+0+ 0+  0+  0+0+0+0+0+0+4S>*0+0+0+d..Xell-C..r"r}c+*^\rSrSrSrS&S\\S\S\S\S\S\S \S \S \ S \S \S\S\S\S\S\S\S\S\S\S\4*U4Sjjjr Sr S'Sjr S(S\ RS\S\S\S\4 S jjrS!rS"rS#\ R4S$jrS%rU=r$))HDemucsi-a Hybrid Demucs model from *Hybrid Spectrogram and Waveform Source Separation* :cite:`defossez2021hybrid`. See Also: * :class:`torchaudio.pipelines.SourceSeparationBundle`: Source separation pipeline with pre-trained models. Args: sources (List[str]): list of source names. List can contain the following source options: [``"bass"``, ``"drums"``, ``"other"``, ``"mixture"``, ``"vocals"``]. audio_channels (int, optional): input/output audio channels. (Default: 2) channels (int, optional): initial number of hidden channels. (Default: 48) growth (int, optional): increase the number of hidden channels by this factor at each layer. (Default: 2) nfft (int, optional): number of fft bins. Note that changing this requires careful computation of various shape parameters and will not work out of the box for hybrid models. (Default: 4096) depth (int, optional): number of layers in encoder and decoder (Default: 6) freq_emb (float, optional): add frequency embedding after the first frequency layer if > 0, the actual value controls the weight of the embedding. (Default: 0.2) emb_scale (int, optional): equivalent to scaling the embedding learning rate (Default: 10) emb_smooth (bool, optional): initialize the embedding with a smooth one (with respect to frequencies). (Default: ``True``) kernel_size (int, optional): kernel_size for encoder and decoder layers. (Default: 8) time_stride (int, optional): stride for the final time layer, after the merge. (Default: 2) stride (int, optional): stride for encoder and decoder layers. (Default: 4) context (int, optional): context for 1x1 conv in the decoder. (Default: 4) context_enc (int, optional): context for 1x1 conv in the encoder. (Default: 0) norm_starts (int, optional): layer at which group norm starts being used. decoder layers are numbered in reverse order. (Default: 4) norm_groups (int, optional): number of groups for group norm. (Default: 4) dconv_depth (int, optional): depth of residual DConv branch. (Default: 2) dconv_comp (int, optional): compression of DConv branch. (Default: 4) dconv_attn (int, optional): adds attention layers in DConv branch starting at this layer. (Default: 4) dconv_lstm (int, optional): adds a LSTM layer in DConv branch starting at this layer. (Default: 4) dconv_init (float, optional): initial scale for the DConv branch LayerScale. (Default: 1e-4) sourcesaudio_channelschannelsgrowthnfftdepthfreq_emb emb_scale emb_smoothr? time_strider@rE context_enc norm_startsrA dconv_depth dconv_comp dconv_attn dconv_lstm dconv_initc >[T+U]5 X`lXPlX lXlXlXlXlX0l URS-Ul SUl [R"5Ul[R"5Ul[R"5Ul[R"5UlUnUS-nUnUnURS-n[%UR5GHnUU:nUU:nUU:aSOSnUS:nU n U n!U(dUS:wa ['S5eU S-n!U n Sn"Sn#U(a UU ::aUn!Sn"Sn#U!U UU"UUUUUUUS .S .n$[)U$5n%S U%S 'U U%S 'U U%S'SU%S'[)U$5n&U#(a[+UU5nUn[-UU4SU0U$D6n'U(a@U#SLaUS:Xa SU%S'SU%S '[-UU4UU#S.U%D6n(UR R/U(5 URR/U'5 US :Xa'UR[1UR 5-nUS-n[3UU4US :HU S.U&D6n)U(a0[3UU4U#US :HU S.U%D6n*UR"R5S U*5 URR5S U)5 UnUn[7UU-5n[7UU-5nU(aUU ::aSnOUU -nUS :XdGMU(dGM[9UUXS9Ul XplGM [=U5 g)NrWrXrQnonerz$When freq is false, freqs must be 1.TF)lstmattnrcompressinit)r?r@rCrGrDrArFrrCr?r@rGrE)rErB)rrE)rBrrE)rr)rrrrrrr?rEr@r hop_lengthrr ModuleList freq_encoder freq_decoder time_encoder time_decoderrangerldictmaxr;appendlenr}insertr2r freq_emb_scale_rescale_module),rrrrrrrrrrr?rr@rErrrArrrrrr=chin_zr>chout_zfreqsindexrrrDrCstrikerrG last_freqkwkwtkw_decenctencdectdecrs, r rHDemucs.__init__Qs60   , &   ))q. MMOMMOMMOMMO Q4::&EJ&DJ&D(-(< &I19DDCA:$%KLL!Ao"CI,  #&*  ( *& Br(CCK!,C "CMCJ"XFeW-VWHkHRHC$$%CM)*C &!$[{)[WZ[!!((.    $ $S )z**S->>WfY5A:wYRXYC!%hYUaZY`hdgh!!((D1    $ $Q ,DF'E&7*+GK'Ef$Ezhh 0z c &.#W'Z r"cURnURnUnX#S-:wa [S5e[[R "UR SU- 55nUS-S-nURXXeU--UR S- SS9n[XU5SSS2SS24nUR SUS-:wa [S 5eUSSSU-24nU$) NrWzHop length must be nfft // 4rgrXrhreflect)mode.zESpectrogram's last dimension must be 4 + input size divided by stride) rrrlr2mathceilri_pad1d_spectro)rr'hlrx0rurGrws r _spec HDemucs._specs __yy  ?;< < 1772;+, -Agk KK2g  ;)K L Qb !#ssA+ . 772;"q& de e c1q2v:o r"cURn[R"U/SQ5n[R"USS/5nUS-S-nU[[R "X#- 55-SU--n[ XUS9nUSXDU-24nU$)N)rrrrrXrh)r.)rrkrGr2rr _ispectro)rrwrrrGrur's r _ispecHDemucs._ispecs __ EE!\ " EE!aV Agk #dii ,- -C 7 aB ' c3v%% &r"r' padding_left padding_rightrvaluecURSnUS:Xa.[X#5nXg::a[R"USXv- S-45n[R"XU4XE5$)zWrapper around F.pad, in order for reflect padding when num_frames is shorter than max_pad. Add extra zero padding around in order for padding to not break.rgrrr)rirrkrG)rr'rrrrrmax_pads r rHDemucs._pad1ds\ 9 ,6G EE!a!1A!567uuQ}5tCCr"cURup#pE[R"U5RSSSSS5nUR X#S-XE5nU$)NrrrWrXrh)rir view_as_realrnro)rrwrqrrrsrtms r _magnitudeHDemucs._magnitudesKgg b   q ! ) )!Q1a 8 IIaQ &r"cURup#pEnURX#SSXV5RSSSSSS5n[R"UR 55nU$)NrgrXrrrWrh)rirjrnrview_as_complex contiguous)rrrqSrrrsrtr*s r _mask HDemucs._masksVaQffQ2q"(00Aq!QB##CNN$45 r"inputcURS:wa[SUR35eURSUR:wa[SURSS35eUnURSnUR U5nUR U5nUnURupgpUR SSS 9n URSSS 9n X*- S U -- nUn U R S SS 9n U RS SS 9nX- S U-- n /n/n/n/n[UR5GH=unnURURS5 S nU[UR5:aZURU RS5 URUnU"U 5n UR(dURU 5 OU nU"UU5nUS :XaURbv[R "URSUR"S9nURU5R%5S S S 2S S 2S 4R'U5nX R(U--nURU5 GM@ [R*"U5n[R*"U5n [UR,5HunnUR/S5nU"UUUR/S55unnUR0[UR25- nUU:dM]UR2UU- nUR/S5nUR(aFURSS:wa[SUR35eUS S 2S S 2S 4nU"US U5un nMUR/S5nU"U UU5un nM [U5S :wa [5S5e[U5S :wa [5S5e[U5S :wa [5S5e[UR65n UR9UU SX5nX+S S 2S 4-U S S 2S 4-nUR;U5n!UR=U!U5nU R9UU SU5n XS S 2S 4-U S S 2S 4-n X-nU$)zHDemucs forward call Args: input (torch.Tensor): input mixed tensor of shape `(batch_size, channel, num_frames)` Returns: Tensor output tensor split into sources of shape `(batch_size, num_sources, channel, num_frames)` rhzDExpected 3D tensor with dimensions (batch, channel, frames). Found: rzZThe channel dimension of input Tensor must match `audio_channels` of HDemucs model. Found:.rg)rrXrhT)rkeepdimgh㈵>)rrXNr)devicerXz0If tdec empty is True, pre shape does not match zsaved is not emptyzlengths_t is not emptyzsaved_t is not empty)ndimrlrirrrmeanstd enumeraterrrrrBrrrrt expand_asr zeros_likerpoprrAssertionErrorrrjrr)"rrr'rrwmagrqrrFqrtrrxtmeantstdtsavedsaved_tlengths lengths_tidxencodererfrsembdecoderpreoffsetrlength_t_rzouts" r r+HDemucs.forward s@ ::?cdidodocpqr r ;;q>T00 0Q(+   JJu ooa  gg bvv)Tv2ee 4e0 X$* %FD1vv&$v/jTD[ )! $T%6%67KC NN1772; 'FS**++  ".((-"XzzNN2& Fq&!AaxDMM5ll1772;qxx@mmC(**,T1a-=>HHK++c11 LLO/82   Q    a %T%6%67KC99R=DAtW[[_5FAsZZ#d&7&7"88Ff}((v6$==,::yy|q((+[\_\e\e[f)ghhaAg,C dH5EB";;r?D T84EB!8$ u:? !56 6 y>Q  !9: : w<1  !78 8   FF1aR # AtG tAtG} ,zz!} KKf % WWQ2v & q$w- %4. 0 Fr")rrrErrrrrrr?rrr@rr)rX0rXg? TryrXrWrrrWrWrXrWrWrW-C6?r%)zerog)r-r.r/r0r1rr{r2r3r4rrrrr6rrrr+r7r8r9s@r rr-s!L  -~c~~ ~  ~  ~~~~~~~~~~ !~"#~$%~&'~()~*+~,-~~@0D DCDDSVDhmDoU\\oor"rct^\rSrSrSrSS\S\S\S\S\S\S \S \S \S \4U4S jjjr Sr Sr U=r $)r`i}am New residual branches in each encoder layer. This alternates dilated convolutions, potentially with LSTMs and attention. Also before entering each residual branch, dimension is projected on a smaller subspace, e.g. of dim `channels // compress`. Args: channels (int): input/output channels for residual branch. compress (float, optional): amount of channel compression inside the branch. (default: 4) depth (int, optional): number of layers in the residual branch. Each layer has its own projection, and potentially LSTM and attention.(default: 2) init (float, optional): initial scale for LayerNorm. (default: 1e-4) norm_type (bool, optional): Norm type, either ``group_norm `` or ``none`` (Default: ``group_norm``) attn (bool, optional): use LocalAttention. (Default: ``False``) heads (int, optional): number of heads for the LocalAttention. (default: 4) ndecay (int, optional): number of decay controls in the LocalAttention. (default: 4) lstm (bool, optional): use LSTM. (Default: ``False``) kernel_size (int, optional): kernel size for the (dilated) convolutions. (default: 3) rrrrrDrheadsndecayrr?c 8>[TU]5 U S-S:Xa [S5eXlX l[ U5UlUS:n Sn US:XaSn [X- 5n [Rn[R"/5Ul [UR 5HnU (a [SU5OSnUU S--n[R"XU UUS9U "U 5U"5[R"U SU-S5U "SU-5[R"S5[!X5/nU(aUR#S [%XUS 95 U (aUR#S ['U SS S 95 [R("U6nURR+U5 M g) NrXrz(Kernel size should not be divisible by 2c,[R"5$r%rJrLs r rN!_DConv.__init__..rPr"rQc0[R"SU5$)NrrSrLs r rNrs Q 2r"r)dilationpaddingrh)rrT)layersr)rrrlrrabsrr2rGELUrrrpowrYGLU _LayerScaler _LocalState_BLSTM Sequentialr)rrrrrrDrrrrr?dilaterahiddenactrMrrmodslayerrs r r_DConv.__init__s^  ?a GH H  Z *  $2GX()ggmmB' tzz"A$*s1ayH+"23G (K(T[\ &!h,2H %q H+D A{6vNO AvfQTBCMM4(E KK  u %##r"c@URH nX"U5-nM U$)zDConv forward call Args: x (torch.Tensor): input tensor for convolution Returns: Tensor Output after being run through layers. )r)rr'r$s r r+_DConv.forwards$[[EE!H A!r")rrrr) rWrXr rQFrWrWFrh) r-r.r/r0r1r2r3r{r4rr+r7r8r9s@r r`r`}s.%1&1&1& 1&  1&  1&1&1&1&1&1&1&f  r"r`cv^\rSrSrSrS S\S\4U4SjjjrS\RS\R4Sjr S r U=r $) riaE BiLSTM with same hidden units as input dim. If `max_steps` is not None, input will be splitting in overlapping chunks and the LSTM applied separately on each chunk. Args: dim (int): dimensions at LSTM layer. layers (int, optional): number of LSTM layers. (default: 1) skip (bool, optional): (default: ``False``) rrc>[TU]5 SUl[R"SX!US9Ul[R "SU-U5UlX0lg)NT) bidirectional num_layers hidden_size input_sizerX) rr max_stepsrLSTMrLinearlinearr)rrrrrs r r_BLSTM.__init__sE GG$6_bc iiC-  r"r'r#c XURup#nUnSnSnSnSn URbaX@R:aRURnUS-n[XU5n U RSn SnU RSSSS5R SX75nURSSS5nUR U5SnUR U5nURSSS5nU(a/n UR USX75n US-n [U 5Hpn U S:Xa!U RU SS2U SS2SU *245 M*XS- :Xa U RU SS2U SS2U S245 MRU RU SS2U SS2X*245 Mr [R"U S5n U S SU24n U nUR(aX-nU$) zBLSTM forward call Args: x (torch.Tensor): input tensor for BLSTM shape is `(batch_size, dim, time_steps)` Returns: Tensor Output after being run through bidirectional LSTM. Shape is `(batch_size, dim, time_steps)` FrNrXTrrhrg.) rir/_unfoldrnrorr2rrrcatr)rr'rqrrrtrvframedwidthr@nframesframesr*limitks r r+_BLSTM.forwards''a  >> %!nn*<NNEaZFQv.Fll1oGFq!Q*222q@A IIaA  IIaLO KKN IIaA  CYYq"a/FaKE7^6JJvaAww&678A+%JJvaAuvo67JJvaAuV|&;<= $))C$Cc2A2g,CA 99Ar")r2rr/r)rF) r-r.r/r0r1r2r4rrr6r+r7r8r9s@r rrs@C4..%,,..r"rcz^\rSrSrSrS S\S\S\4U4SjjjrS\RS\R4S jr S r U=r $) rizLocal state allows to have attention based only on data (no positional embedding), but while setting a constraint on the time window (e.g. decaying penalty term). Also a failed experiments with trying to provide some frequency based attention. rrrc|>[[U] 5 X-S:wa [S5eX lX0l[ R"XS5Ul[ R"XS5Ul [ R"XS5Ul [ R"XU-S5Ul U(anURR=RS-sl URRc [S5eSURRRSS&[ R"XS--US5Ulg)z Args: channels (int): Size of Conv1d layers. heads (int, optional): (default: 4) ndecay (int, optional): (default: 4) rz$Channels must be divisible by heads.rg{Gz?Nzbias must not be None.r)rrrrlrrrrYcontentquerykey query_decayrrbiasproj)rrrrrs r r_LocalState.__init__s k4)+  q CD D  yyQ7 YYx15 99X399Xv~qA     # # ( (D 0 ($$, !9::,.D   ! ! & &q )IIh2Ha@ r"r'r#cURup#nURn[R"XARUR S9nUSS2S4USSS24- nUR U5RX%SU5nURU5RX%SU5n [R"SX5n U [R"U RS5-n UR(a[R"SURS-URUR S9n URU5RX%SU5n [R"U 5S- n U RSSS5*UR5-[R"UR5- n U [R"SX5- n U R![R""XJR[R$S9S5 [R&"U SS 9nUR)U5RX%SU5n[R"S X5nUR+USU5nXR-U5-$) zLocalState forward call Args: x (torch.Tensor): input tensor for LocalState Returns: Tensor Output after being run through LocalState layer. )rdtypeNrgzbhct,bhcs->bhtsrXrzfts,bhfs->bhtsirzbhts,bhct->bhcs)rirrrrrHrArjrBeinsumrrrrCsigmoidr masked_fill_eyer4softmaxr@rorE)rr'rqrrrtrindexesdeltaquerieskeysdotsdecaysdecay_q decay_kernelweightsr@results r r+_LocalState.forward6s''a ,,qA4 747#33**Q-$$Qr15xx{"a0||-t=  $**Q-(( ;;\\!T[[1_QXXQWWUF&&q)..qQ?GmmG,q0G"KKAq11EIIK?$))DKKBXXL ELL!1<I ID %))AkkLdS--!,,,q/&&qQ7/B2q)99V$$$r")r@rrBrrErArC)rWrW) r-r.r/r0r1r2rrr6r+r7r8r9s@r rrsL AASAcAA2#%#%%,,#%#%r"rcv^\rSrSrSrS S\S\4U4SjjjrS\RS\R4Sjr S r U=r $) ri\zLayer scale from [Touvron et al 2021] (https://arxiv.org/pdf/2103.17239.pdf). This rescales diagonally residual outputs close to 0 initially, then learnt. rrc>[TU]5 [R"[R "USS95UlX R RSS&g)zk Args: channels (int): Size of rescaling init (float, optional): Scale to default to (default: 0) T) requires_gradN)rrr Parameterrzerosrr)rrrrs r r_LayerScale.__init__as; \\%++hd"KL ! r"r'r#c.URSS2S4U-$)zLayerScale forward call Args: x (torch.Tensor): input tensor for LayerScale Returns: Tensor Output after rescaling tensor. Nr)rr's r r+_LayerScale.forwardkszz!T'"Q&&r"r`)r) r-r.r/r0r1r2r3rrr6r+r7r8r9s@r rr\s@""E"" ' '%,, ' 'r"rar?r@r#c[URSS5n[URS5n[R"XB- 5nUS- U-U-n[ R "USXd- /S9n[UR55Vs/sHopRU5PM nnUSS:wa [S5eUSSUS/-nURU5 URU5 URX85$s snf)zGiven input of size [*OT, T], output Tensor of size [*OT, F, K] with K the kernel size, by extracting frames with the given stride. This will pad the input so that `F = ceil(T / K)`. see https://github.com/pytorch/pytorch/issues/60466 Nrgrr)rrGzData should be contiguous.) listrir2rrrkrGrrr@rlr as_strided) rbr?r@rirn_frames tgt_lengthrstridess r r5r5xs " E  Fyy)HQ,&(;6J AAz234A(-aeeg7xx}G7r{a566crlfa[(G LL LL << '' 8sC5cUR5Hn[U[R[R[R [R 45(dMRURR5R5nUS- S-nUR=RU-sl URcMUR=RU-sl M g)zA Rescales initial weight scale for all models within the module. g?g?N) modules isinstancerrYrrZrrrdetachrrD)modulesubrrs r rrs~~ cBIIr'9'9299bFXFXY Z Z**.."))+C3Y3&E JJOOu $Oxx# &  r"r'n_fftrrGc t[URSS5n[URS5nURSU5n[R "UUSU--U[R "U5RU5USSSSS9 nURupxn URX/5 URU5$)NrgrTr)window win_length normalizedcenterreturn_complexpad_mode) rdrir2rorstft hann_windowtoextendrj) r'rorrGotherrrwrrframes r rrs " E  F "fA   S  '**1- AggOAe LL% 66%=r"rwrc [URSS5n[URS5n[URS5nSU-S- nURSXV5nUSU--n[R "UUU[R "U5RUR5USUSS9n U RupURU5 U RU5$)NrrgrXrT)rqrrrsrrt) rdrir2rjristftrxryrealr) rwrrrGr{rr:rorrr'rs r rrs " E  E  F IME r5!A1s7#J     ,//7 AIA LL 66%=r"rc[USSS9$)zBuilds low nfft (1024) version of :class:`HDemucs`, suitable for sample rates around 8 kHz. Args: sources (List[str]): See :py:func:`HDemucs`. Returns: HDemucs: HDemucs model. irrrrrrs r hdemucs_lowr 7Q 77r"c[USSS9$)aBuilds medium nfft (2048) version of :class:`HDemucs`, suitable for sample rates of 16-32 kHz. .. note:: Medium HDemucs has not been tested against the original Hybrid Demucs as this nfft and depth configuration is not compatible with the original implementation in https://github.com/facebookresearch/demucs Args: sources (List[str]): See :py:func:`HDemucs`. Returns: HDemucs: HDemucs model. rr rrrs r hdemucs_mediumrs 7Q 77r"c[USSS9$)zBuilds medium nfft (4096) version of :class:`HDemucs`, suitable for sample rates of 44.1-48 kHz. Args: sources (List[str]): See :py:func:`HDemucs`. Returns: HDemucs: HDemucs model. r r rrrs r hdemucs_highrrr")irr)rrr)rtypingtprrrrrrtorch.nnrrkModuler r;r}rr`rrrr6r2r5rrrr{rrrr"r rs4 ,, $#uxx#Lkk\sslMehhooM` TUXX__Tn@UXX__@FB%"))B%J'"))'8(u||(#(s(u||(( ' SC#V[VbVb(3C#V[VbVb. 8c 8w 88DI8'8& 8$s) 8 8r"