]i0SSKJrJr SSKrSSKrSSKJr SSKJs J r SSKJ r SSKJ r J r JrJr SSKJr SSKJrJr "SS \R*5r"S S \R*5rg) )OptionalMappingN)Tensor)LSTM BatchNorm1dLinear Parameter)wiener)make_filterbanks ComplexNormc^\rSrSrSrSS\S\S\S\S\S\\RS \\RS \\4U4S jjjr S r S \ S\ 4Sjr SrU=r$) OpenUnmix a(OpenUnmix Core spectrogram based separation module. Args: nb_bins (int): Number of input time-frequency bins (Default: `4096`). nb_channels (int): Number of input audio channels (Default: `2`). hidden_size (int): Size for bottleneck layers (Default: `512`). nb_layers (int): Number of Bi-LSTM layers (Default: `3`). unidirectional (bool): Use causal model useful for realtime purpose. (Default `False`) input_mean (ndarray or None): global data mean of shape `(nb_bins, )`. Defaults to zeros(nb_bins) input_scale (ndarray or None): global data mean of shape `(nb_bins, )`. Defaults to ones(nb_bins) max_bin (int or None): Internal frequency bin threshold to reduce high frequency content. Defaults to `None` which results in `nb_bins` nb_bins nb_channels hidden_size nb_layersunidirectional input_mean input_scalemax_binc >[[U] 5 XlU(aXlOURUlX0l[ URU-USS9Ul[U5Ul U(aUn OUS-n [UU UU(+SUS:aSOSS9Ul US-n [ XSS9Ul [U5Ul [ UURU-SS9Ul[URU-5UlUb3[ R""USUR*5R%5nO [ R&"UR5nUb5[ R""S USUR- 5R%5nO [ R("UR5n[+U5Ul[+U5Ul[+[ R("UR5R%55Ul[+[ R("UR5R%55Ulg) NF)biasr g?r) input_sizer num_layers bidirectional batch_firstdropout) in_features out_featuresrg?)superr__init__nb_output_binsrrrfc1rbn1rlstmfc2bn2fc3bn3torch from_numpyfloatzerosonesr rr output_scale output_mean) selfrrrrrrrrlstm_hidden_sizefc2_hiddensize __class__s L/mnt/rpi/tmp/demucs-venv-sys/lib/python3.13/site-packages/openunmix/model.pyr$OpenUnmix.__init__ s i')% "L..DL&$,,4kN{+ * *a/ "( ,,$q=Ca   %qnUZ[{+#,,{:  t22[@A  !)):n +E*EFLLNJT\\2J  "**3^t||1L+LMSSUK**T\\2K#J/$[1%ejj1D1D&E&K&K&MN$UZZ0C0C%D%J%J%LMc`UR5H nSUlM UR5 gNF parameters requires_gradevalr4ps r8freezeOpenUnmix.freezed%"A#AO# r:xreturncURSSSS5nURRup#pEUR5R 5nUSSUR 24nXR -nXR-nURURSX@R -55nURU5nURX#UR5n[R"U5nURU5n[R"XS/S5nUR!URSURS55nUR#U5n[$R&"U5nUR)U5nUR+U5nURX#X@R,5nXR.-nXR0- n[$R&"U5U-nURSSSS5$)z Args: x: input spectrogram of shape `(nb_samples, nb_channels, nb_bins, nb_frames)` Returns: Tensor: filtered spectrogram of shape `(nb_samples, nb_channels, nb_bins, nb_frames)` rr r.N)permutedatashapedetachclonerrrr&reshaper'rr-tanhr(catr)r*Frelur+r,r%r2r3)r4rF nb_frames nb_samplesrrmixlstm_outs r8forwardOpenUnmix.forwardks IIaAq !67ffll3 {hhj  c>T\\>! "      HHQYYr;#=> ? HHQK IIiT-=-= > JJqM99Q< IIq1+& + HHQYYr1772;/ 0 HHQK FF1I HHQK HHQK IIi[:M:M N     FF1IOyyAq!$$r:)r'r*r,r&r)r+rrrr(rr%r3r2)rirIFNNN)__name__ __module__ __qualname____firstlineno____doc__intboolrnpndarrayr$rCrrY__static_attributes__ __classcell__r7s@r8rr s($+/,0!%BNBNBN BN  BN  BNRZZ(BNbjj)BN#BNBNH;%;%F;%;%r:rc^\rSrSrSrSS\\\R4S\ S\ S\ S\ S\ S \ S \ S \ \ S \4U4S jjjr SrS\S\4SjrSS\S\ \S\4SjjrSrU=r$) Separatora Separator class to encapsulate all the stereo filtering as a torch Module, to enable end-to-end learning. Args: targets (dict of str: nn.Module): dictionary of target models the spectrogram models to be used by the Separator. niter (int): Number of EM steps for refining initial estimates in a post-processing stage. Zeroed if only one target is estimated. defaults to `1`. residual (bool): adds an additional residual target, obtained by subtracting the other estimated targets from the mixture, before any potential EM post-processing. Defaults to `False`. wiener_win_len (int or None): The size of the excerpts (number of frames) on which to apply filtering independently. This means assuming time varying stereo models and localization of sources. None means not batching but using the whole signal. It comes at the price of a much larger memory usage. filterbank (str): filterbank implementation method. Supported are `['torch', 'asteroid']`. `torch` is about 30% faster compared to `asteroid` on large FFT sizes such as 4096. However, asteroids stft can be exported to onnx, which makes is practical for deployment. target_modelsnitersoftmaskresidual sample_raten_fftn_hoprwiener_win_len filterbankc f>[[U] 5 X lX@lX0lXl[UUSU US9uUlUl [US:HS9Ul [R"U5Ul[UR5UlUR#S[$R&"U55 g)NT)rprqcentermethodror )monoro)r#rir$rlrnrmrrr stftistftr complexnormnn ModuleDictrklen nb_targetsregister_bufferr- as_tensor) r4rkrlrmrnrorprqrrrrsr7s r8r$Separator.__init__s i')   , 0# !  4:'K1,<= ]]=9d001 ]EOOK,HIr:c`UR5H nSUlM UR5 gr<r=rAs r8rCSeparator.freezerEr:audiorGc URnURSnURU5nURU5n[R "URU4-UR URS9n[URR55H2unupU "UR5R55n XSU4'M4 URSSSSS5nURSSSSS5nUR(aUS- nUS:XaURS:a [!S5eURSn [R "URU4-UR URS9n [#U5Hn SnUR$(a UR$nOU nX:dM,[R&"U[)XU-55n[+US 5S-n[-XmU4XMU4URUR.URS 9XU4'X:aMrM U RSS SSSS5R15n UR3XRSS 9nU$) aPerforming the separation on audio input Args: audio (Tensor): [shape=(nb_samples, nb_channels, nb_timesteps)] mixture audio waveform Returns: Tensor: stacked tensor of separated waveforms shape `(nb_samples, nb_targets, nb_channels, nb_timesteps)` r)dtypedevice.rIrr zoCannot use EM if only one target is estimated.Provide two targets or create an additional one with `--residual`rJ)rmrn)length)r~rMrxrzr-r0rr enumeraterkitemsrNrOrKrnrl Exceptionrangerrarangeminrar rm contiguousry)r4r nb_sourcesrVmix_stftX spectrogramsj target_name target_moduletarget_spectrogramrU targets_stftsampleposrr cur_frame estimatess r8rYSeparator.forwards<__ [[^ 99U#   X &{{177j]#:%++VWV^V^_ /89K9K9Q9Q9S/T +A+ !.qxxz/?/?/A!B #5a 0U$++Aq!Q: ##Aq!Q2 == !OJ ?tzzA~(  !&&q) {{8>>ZM#A]e]l]lm J'FC""!%!4!4!*/!LLc)>=Q.RS )B-(1,28 !23Y./JJ!]]!]] 3 Y./ / (&$++Aq!Q1=HHJ JJ|KKNJC r:raggregate_dictc0n[UR5HupEUSS2US4X5'M UR(a USS2SS4US'Ub;0nUH1n[R"S5Xg'X'HnXgX5-Xg'M M3 UnU$)zConvert estimates as stacked tensor to dictionary Args: estimates (Tensor): separated targets of shape (nb_samples, nb_targets, nb_channels, nb_timesteps) aggregate_dict (dict or None) Returns: (dict of str: Tensor): N.rJrng)rrkrnr-tensor)r4rrestimates_dictktarget new_estimateskeys r8to_dictSeparator.to_dict?s"4#5#56IA%.q!Sy%9N "7 ==)21b#:)>N: &  %M%%*\\#%6 ",1F)6);n>T)TM&2&+Nr:) rzryr~rlrnrmrxrkrr) rFFg@r[iri,r-)N)r\r]r^r_r`rstrr{Modulerarbr/rr$rCrrYdictrrerfrgs@r8riris<$(+!$JsBII~.$J$J $J  $J  $J$J$J$J! $J$J$JLKVKKZ$SWr:ri)typingrrnumpyrcr-torch.nnr{torch.nn.functional functionalrSrrrrr filteringr transformsr r rrrir:r8rsK$ 995Y% Y%xp pr: