]i VSSKrSSKrSSKJr SSKJrJrJrJrJ r J r SSK r SSK r SSK r SSKrS\S\4SjrSMS\S\S\\S \\4S jjrS \4S jr"S S\5rSNS\ R0S\S\S\ R04SjjrS\ R0S\ R04SjrS\ R0S\ R04Sjr"SS\ R8R:R<5rS\R@S\RBS\\\\RB44Sjr""SS\5r#"SS\5r$"SS\5r%"S S!\5r&"S"S#\5r'\(S$:XGaU\R@"S%S&9r)\)RUS'\S(/S)QS*S+9 \)RUS,\S-S.9 \)RUS/S0S1S29 \)RUS3\S4S59 \)RUS6\+S7S59 \)RUS8\S9S:S;9 \)RUS<\S=S>S;9 \)RUS?\+S@S59 \)RY5ur-r.\ R^"\-R`5 \""\)\-5ur1r2r-\3"SA\ Rh"55 Sr5\R"\6"\7"\1555Hr8\1\8ur9r:\5\9RvSB\1Rx- - r5\-Rz(dM=\ Rz"SC\"\85-SD-\9R|\1Rx5 \ Rz"SC\"\85-SE-\:R|\1Rx5 M \3"SF\5SG- 5 \3"SH\7"\155 \3"SI\7"\255 \-R~\1l@\ R8R:R\1\-RSJSKSL9rC\R"\C5Hur9r:\3"\9Rv5 M gg)ON)Path)OptionalUnionTupleListAnyCallablepathreturnc[R"5S:Xa [S5e0n[R"[ U55nUR US'UR US'URUS'USUS- US'U$)zLoad audio metadata this is a backend_independent wrapper around torchaudio.info Args: path: Path of filename Returns: Dict: Metadata with `samplerate`, `samples` and `duration` in seconds soxz#Deprecated backend is not supported sampleratesampleschannelsduration) torchaudioget_audio_backend RuntimeErrorinfostr sample_rate num_frames num_channels)r rsis K/mnt/rpi/tmp/demucs-venv-sys/lib/python3.13/site-packages/openunmix/data.py load_infor s~##%.@AA D T #BDmmDODIl);;D KstartdurrcUc"U5upEXE4$Uc [U5n[X#S-5n[XS-5n[R"XUS9upEXE4$)a@Load audio file Args: path: Path of audio file start: start position in seconds, defaults on the beginning. dur: end position in seconds, defaults to `None` (full file). info: metadata object as called from `load_info`. Returns: Tensor: torch tensor waveform of shape `(num_channels, num_samples)` r)r frame_offset)rloadrint)r rrrsigraterr!s r load_audior&%so$ {OOD) y <T?DL112 5 #556 OODl[ yrlist_of_function_namesctU(a*[UVs/sHn[5SU-PM sn5$S$s snf)N _augment_cU$Naudios raug_from_str..IsUr)Composeglobals)r'augs r aug_from_strr4Es:@VW@V +"34@VWXX""Xs5cX\rSrSrSrSrS\RS\R4SjrSr g) r1LzeComposes several augmentation transforms. Args: augmentations: list of augmentations to compose. cXlgr+ transforms)selfr9s r__init__Compose.__init__Rs$rr.r c<URH nU"U5nM U$r+r8)r:r.ts r__call__Compose.__call__UsAeHE! rr8N) __name__ __module__ __qualname____firstlineno____doc__r;torchTensorr?__static_attributes__r,rrr1r1Ls' %ellu||rr1r.lowhighcFU[R"S5X!- --nX-$)z.Applies a random gain between `low` and `high`)rFrand)r.rIrJgs r _augment_gainrO[s# ejjmtz**A 9rcURSS:Xa@[R"S5R5S:a[R"US/5$U$)z;Swap channels of stereo signals with a probability of p=0.5rg?g?)shaperFtensoruniform_flipr-s r_augment_channelswaprVasC {{1~u||C099;cAzz%!%% rcURSS:a USS2S4nURSS:Xa[R"USSS9nU$)NrrQ.rLdim)rRrFrepeat_interleaver-s r_augment_force_stereor[isL {{1~bqb#g {{1~''qa8 Lrc \rSrSrSrSS\\\4S\S\ \S\ \ SS4 S jjr S \ S\ 4S jrS\ 4S jrS\4S jrS\4SjrSrg) UnmixDatasetuNrootr seq_durationsource_augmentationsr c[[R5R5UlX lX0lX@lgr+)rargsr` expanduserrrarb)r:r`rrarbs rr;UnmixDataset.__init__xs.O..0 &($8!rindexc[er+NotImplementedError)r:rgs r __getitem__UnmixDataset.__getitem__!!rc[er+rir:s r__len__UnmixDataset.__len__rmrc.SURR-nSRUR55/nX R 5R 5- nU/UVs/sHnSUR -U-PM sn-nSRU5$s snf)NzDataset zNumber of datapoints: {}  ) __class__rAformatrp extra_repr splitlines _repr_indentjoin)r:headbodylineliness r__repr__UnmixDataset.__repr__sDNN333*11$,,.AB !,,..TJTT# 1 11D8TJJyyKs Bcg)Nr,ros rrwUnmixDataset.extra_reprsr)r`rrarb)NN)rArBrCrDryrrrfloatrr r;r#rrkrprrwrHr,rrr]r]usL )-37 9D#I 9 9uo 9 'x0 9  9""""" # Crr]parserrdc  URS:XaURS[S9 URS[S9 UR5n[ UR 5R Ul[ UR5URURUR S.n[ UR 5R Ul[S:SSS.UD6n[S:S S 0UD6nGOnURS :XGa6URS [S S9 URS[S9 URS[SS9 URS[SS9 URS[SS9 URS[S S9 UR5nURUl[ UR5URURURS.n[!UR"5n[%S:SUSUR&URS.UD6n[%S:S SURUR(S.UD6nGO'URS:XaURS[S9 URS[S S9 URSSS S!S"9 URS[S S9 UR5n[ UR*5R Ul[ UR5UR,UR*S#.n[!UR"5n[/S:SUUR0SURS$.UD6n[/S:S S%S&.UD6nGOURS':XGa6URS[SS9 URS[S9 URS[S S9 URS(SS S)S"9 URS*SS S+S"9 UR5n[ UR*5R Ul[ UR5UR*URUR2S,.n[5UR"Vs/sHn[75S-U-PM sn5n[9S:SUUR:SURS..UD6n[9S:S S%S&.UD6nOURS/SS S0S"9 URS1[S2S9 URS[S3S4/S S59 UR5nURUR<SURURS%LUR>S6.n[!UR"5n[AS:SURBURUSS7.UD6n[AS:S S8S%S9.UD6nX4U4$s snf);zgLoads the specified dataset from commandline arguments Returns: train_dataset, validation_dataset alignedz --input-file)typez --output-file)r`ra input_file output_filetrainT)split random_chunksrvalid sourcefolderz--interferer-dirs+)rnargsz --target-dirz--ext.wavrdefaultz--nb-train-samplesz--nb-valid-samplesdz--source-augmentations)r`interferer_dirs target_dirext)rrbr nb_samplesra)rrrartrackfolder_fixz --target-filez--interferer-filesz--random-track-mix store_trueFz&Apply random track mixing augmentation)actionrhelp)r`interferer_files target_file)rrbrandom_track_mixrraN)rratrackfolder_varz--random-interferer-mixz+Apply random interferer mixing augmentationz--silence-missingzsilence missing targets)r`rrsilence_missing_targetsr))rrbrandom_interferer_mixrraz--is-wavzloads wav instead of STEMSz--samples-per-track@gain channelswap)rrr)r`is_wavsubsetstargetdownloadseed)rsamples_per_trackrarbrrL)rrrar,)"dataset add_argumentr parse_argsrrstemrr`seq_durrAlignedDatasetr#rrrr4rbSourceFolderDatasetnb_train_samplesnb_valid_samplesrrFixedSourcesTrackFolderDatasetrsilence_missingr1r2!VariableSourcesTrackFolderDatasetrrr MUSDBDatasetr)rrddataset_kwargs train_dataset valid_datasetrbr3s r load_datasetsrs ||y N5O#6  "4++,11 O LL//++   4++,11 &[WD[N[ &GWGG  '/cEN5G#v>0sDI0sCH43cJ  "oo O#33//88   ,D,E,EF+ !5,,    , ,,     * *O#60s#F 9  43cJ  "4++,11 O $ 5 5++  ,D,E,EF6 !5!22    7jWSWj[ij * *G#v>O#643cJ %>   *    "4++,11 O++88'+';';   'PTPiPi'jPi +2C(DPi'jk9 !5"&"<"<    :mVZm^lm   -  1RH43Q^H_gjk  "IIkkkk T)II   ,D,E,EF$ "44!5!    %m7aVZm^lm  --[(ks*Ucx\rSrSrSS\S\S\S\S\\S\S \S \\S \ S S4S jjr Sr Sr Sr Srg)ri>Nr`rrrrarrrbrr c H[U5R5UlX lXplXPlX`lX0lX@l[UR55Ul UR(d [S5eXl [R"UR5 g)ajA dataset of that assumes multiple track folders where each track includes and input and an output file which directly corresponds to the the input and the output of the model. This dataset is the most basic of all datasets provided here, due to the least amount of preprocessing, it is also the fastest option, however, it lacks any kind of source augmentations or custum mixing. Typical use cases: * Source Separation (Mixture -> Target) * Denoising (Noisy -> Clean) * Bandwidth Extension (Low Bandwidth -> High Bandwidth) Example ======= data/train/01/mixture.wav --> input data/train/01/vocals.wav ---> output z)Dataset is empty, please check parametersN)rrer`rrrarrrlist _get_paths tuple_pathsrrrandom) r:r`rrrrarrrbrs rr;AlignedDataset.__init__?sy@J))+  &(*$& 12JK K  DIIrc@URUup#UR(aL[U5n[U5n[USUS5n[R "SX`R - 5nOSn[X'UR S9up[X7UR S9upX4$)Nrrrr)rrrminruniformrar&) r:rg input_path output_path input_info output_inforrX_audio_Y_audios rrkAlignedDataset.__getitem__ms"&"2"25"9   ":.J#K0K:j1;z3JKHNN1h1B1B&BCEE T=N=NO  d>O>OP rc,[UR5$r+)lenrros rrpAlignedDataset.__len__}s4##$$rc#P# [URUR5n[R"UR 55HnUR 5(dM[ URUR55n[ URUR55nU(dMkU(dMtURbM[US5n[US5n[USUS5nXpR:aUSUS4v MMUSUS4v M g7f)Loads input and output tracksNrr) rr`rtqdmiterdiris_dirrglobrrrarr)r:p track_pathrrrr min_durations rrAlignedDataset._get_pathss DJJ '))AIIK0J  ""!*//$//"BC ":??43C3C#DE :++((4%.z!}%= &/ A&? '*:j+A;zCZ'[ '*;*;;",Q-Q"??<)m[^;;1sAD&!A D&2D&;A+D&) rrrr`rrrarr)rz mixture.wav vocals.wavNF@N*)rArBrCrDrrrboolr r#r;rkrprrHr,rrrr>s''(,#$37,,, ,  , uo ,,,'x0,, ,\ %rrr`rrrrrrarrrbrr c `[U5R5UlX lXlXplXPlXlXlX0l X@l URUR/-Ul UR5Ul X`lXl[ R"UR5 g)a@A dataset that assumes folders of sources, instead of track folders. This is a common format for speech and environmental sound datasets such das DCASE. For each source a variable number of tracks/sounds is available, therefore the dataset is unaligned by design. By default, for each sample, sources from random track are drawn to assemble the mixture. Example ======= train/vocals/track11.wav ----------------- train/drums/track202.wav (interferer1) ---+--> input train/bass/track007a.wav (interferer2) --/ train/vocals/track11.wav ---------------------> output N)rrer`rrrarrrbrrsource_folders get_tracks source_tracksrrr) r:r`rrrrrrarrrbrs rr;SourceFolderDataset.__init__s@J))+  &(*$8!$."22doo5FF!__.$  DIIrcP/nURHnURS:Xa[R"U5 [R"UR U5n[ U5SnUR(a$[R"SXPR- 5nO[US-URS-- S5n[XFURS9upxURU5nURU5 M [R"U5n U R!S5n U Sn X4$)NrrrrQr)rrrrchoicerrrrramaxr&rbappendrFstacksum) r:rg audio_sourcessource source_pathrrr.rstemsxys rrkSourceFolderDataset.__getitem__s ))FzzW$ E"!--(:(:6(BCK -j9H!!q(5F5F*FGHMD,=,=,BBAF!+@Q@QRHE--e4E   '%*( M* IIaL "It rcUR$r+)rros rrpSourceFolderDataset.__len__s rc[URUR5n0n[R"UR5Hn/nX- n[ UR SUR-55HTnURb3[U5nUSUR:aURU5 MAMCURU5 MV XBU'M U$)r*r) rr`rrrsortedrrrarr)r:rr source_foldertracksrsource_track_pathrs rrSourceFolderDataset.get_trackss DJJ ' !YYt':':;MF+K%+K,<,rrrrr`rrrrarrrbrrr c [U5R5UlX lXlXPlXplX`lXlX0l X@l URUR/-Ul Xl [R"UR5 [UR55Ul[#UR 5(d [%S5eg)aBA dataset that assumes audio sources to be stored in track folder where each track has a fixed number of sources. For each track the users specifies the target file-name (`target_file`) and a list of interferences files (`interferer_files`). A linear mix is performed on the fly by summing the target and the inferers up. Due to the fact that all tracks comprise the exact same set of sources, the random track mixing augmentation technique can be used, where sources from different tracks are mixed together. Setting `random_track_mix=True` results in an unaligned dataset. When random track mixing is enabled, we define an epoch as when the the target source from all tracks has been seen and only once with whatever interfering sources has randomly been drawn. This dataset is recommended to be used for small/medium size for example like the MUSDB18 or other custom source separation datasets. Example ======= train/1/vocals.wav --------------- train/1/drums.wav (interferer1) ---+--> input train/1/bass.wav -(interferer2) --/ train/1/vocals.wav -------------------> output zNo tracks foundN)rrer`rrrarrrbrr source_filesrrrrrrr) r:r`rrrrarrrbrrs rr;'FixedSourcesTrackFolderDataset.__init__sTJ))+  &( 0*$8!& 0 11T5E5E4FF  DII4??,- 4;;01 1 rcrURUSnURUSnUR(a$[R"SX0R- 5nOSn/n[ X R - X@RS9upgURU5nURU5 URHnUR(a[R"[[UR555n URU SnUR(a5URU Sn[R"SX0R- 5n[ X(- X@RS9upURU 5n URU 5 M [R"U5n U R!S5n U Sn X4$)Nr rrr)rrrrrar&rrbrrrrrangerrFrr)r:rgrrrr target_audiorr random_idxr.rrrs rrk*FixedSourcesTrackFolderDataset.__getitem__4sp[['/ {{5).9   NN1l5F5F&FGEE $Z2B2B%B%UfUfg 00> \*++F$$#]]5T[[1A+BC ![[4V< %%#';;z#:>#JL"NN1l=N=N.NOE!*"5UHYHYZHE--e4E   ', M* IIaL !Ht rc,[UR5$r+rrros rrp&FixedSourcesTrackFolderDataset.__len__Y4;;rc## [URUR5n[R"UR 55HnUR 5(dMUR Vs/sHo2U- PM nn[SU55(d[SU5 M]URbD[[[U55n[SU55nX`R:aX&S.v MMUSS.v M gs snf7f)rc3@# UHoR5v M g7fr+)exists).0sps r .bs>299;;szExclude track Nc3*# UH oSv M g7frNr,ris rrris&De}er r)rr`rrrrr allprintrarmaprr)r:rrs source_pathsinfosrs rr)FixedSourcesTrackFolderDataset.get_tracks\s DJJ '))AIIK0J  ""8<8I8IJ8I1Q8I J>>>>*J7$$0 Y !=>E#&&De&D#DL#&7&77(2QR8%/EF1JsAD!D/D=B D) rrrr`rrrarbr rrr)rArBrCrDrrrrrr r#r;rkrprrHr,rrrrs''1;&?(,#!&3F$:2:2:2 :2 s) :2 uo :2:2:2'x0:2:2:2 :2x#J Grrc\rSrSrSSSSSSSSS4 S \S \S \S \S \\S\S\S\S\\S\SS4Sjjr Sr Sr Sr Sr g)riprrrNFrcU$r+r,r-s rr/*VariableSourcesTrackFolderDataset.{rrr`rrrrarrrrbrr c [U5R5UlX lXlXPlX`lXplXlX0l X@l Xl [UR55Ulg)aXA dataset that assumes audio sources to be stored in track folder where each track has a _variable_ number of sources. The users specifies the target file-name (`target_file`) and the extension of sources to used for mixing. A linear mix is performed on the fly by summing all sources in a track folder. Since the number of sources differ per track, while target is fixed, a random track mix augmentation cannot be used. Instead, a random track can be used to load the interfering sources. Also make sure, that you do not provide the mixture file among the sources! Example ======= train/1/vocals.wav --> input target train/1/drums.wav --> input target | train/1/bass.wav --> input target --+--> input train/1/accordion.wav --> input target | train/1/marimba.wav --> input target / train/1/vocals.wav -----------------------> output N)rrer`rrrarrrbrrrrrr) r:r`rrrrarrrrbrs rr;*VariableSourcesTrackFolderDataset.__init__qs`NJ))+  &(*%:"$8!&'>$4??,- rcfURUSnUR(a6URUSn[R"SX0R- 5nOSnUR (a[R "[[UR555nURUSnUR(a6URUSn[R"SXpR- 5nOSnOUnUn[[URSUR-555n Sn U HAn XUR- :XaM[XURS9upXR#U 5- n MC [%X R- 5R'5(a;[X R- UURS9upUR#U5nX- n X4$[(R*"W R,5nX4$![a& US:aUS- OUS-nUR!U5ss $f=f)Nr rrrrrL)rrrrrarrr rrrrrrr&rrkrbrrrFzerosrR)r:rgtarget_track_pathtarget_min_duration target_startrintfr_track_pathintfr_min_duration intfr_startsourcesrrr.rrs rrk-VariableSourcesTrackFolderDataset.__getitem__s KK.v6   "&++e"4^"D !>>!-@CTCT-TULL  % %uS-='>?J#{{:6v> !!%)[[%<^%L"$nnQ0BEVEV0VW  0 &K.33C$((NCDE "K1A1AAA /%k$J[J[\ **51 1A# !$4$44 5 < < > >!$4$44"%%DA ))!,A FA t  EKK(At '  /%*QY EAI''.. /s H+H0/H0c,[UR5$r+rros rrp)VariableSourcesTrackFolderDataset.__len__rrc ## [URUR5n[R"UR 55HnUR 5(dM[X R 5R5(dUR(dMU[[URSUR-555nU(d[SU5 MURbD[[[ U55n[#SU55nXPR:aX%S.v MMUSS.v M g7f)Nrz empty track: c3*# UH oSv M g7frr,rs rr?VariableSourcesTrackFolderDataset.get_tracks..s*H%QZ=%r r!)rr`rrrrrrrrrrrr#rar$rr)r:rrr6r'rs rr,VariableSourcesTrackFolderDataset.get_trackss DJJ '))AIIK0J  "" $4$45<<>>$B^B^B^$T*//#.*I%JKG"oz: ((4 $SG%< ='**H%*H'H '*;*;;,6#UV<)3DIJ#1sAD>!7D>B"D>) rrrr`rrarrbrrr)rArBrCrDrrrrr r;rkrprrHr,rrrrps'(,#&+$3F(-1.1.1. 1.  1. uo 1.1. $1.1.'x01."&1. 1.f5n Krrc\rSrSrSSSSSSSSSSS 4 S \S \S \S \S\S\S\\S\S\\ S\S\SS4Sjjr Sr Sr Sr g)rirNFrg@rcU$r+r,r-s rr/MUSDBDataset.rrrrr`rrrrrarrbrrr c SSKnXl[R"U 5 X@lXplXlXPlX`lXlXl Xl UR"U UUUUUS.U D6Ul SUl g)aMUSDB18 torch.data.Dataset that samples from the MUSDB tracks using track and excerpts with replacement. Parameters ---------- target : str target name of the source to be separated, defaults to ``vocals``. root : str root path of MUSDB download : boolean automatically download 7s preview version of MUSDB is_wav : boolean specify if the WAV version (instead of the MP4 STEMS) are used subsets : list-like [str] subset str or list of subset. Defaults to ``train``. split : str use (stratified) track splits for validation split (``valid``), defaults to ``train``. seq_duration : float training is performed in chunks of ``seq_duration`` (in seconds, defaults to ``None`` which loads the full audio track samples_per_track : int sets the number of samples, yielded from each track per epoch. Defaults to 64 source_augmentations : list[callables] provide list of augmentation function that take a multi-channel audio file of shape (src, samples) as input and output. Defaults to no-augmentations (input = output) random_track_mix : boolean randomly mixes sources from different tracks to assemble a custom mix. This augmenation is only applied for the train subset. seed : int control randomness of dataset iterations args, kwargs : additional keyword arguments used to add further control for the musdb dataset initialization function. rN)r`rrrrr)musdbrrrrarrrrrbrDBmusr)r:rr`rrrrrarrbrrrdkwargsrBs rr;MUSDBDataset.__init__sl   D (   !2$8! 088     #rc/nSnURRXR-nURS:XGaUR(Ga[ URR S5HupVX`R:XaUnUR(a*[R"URR5nURUl [R"SURUR- 5Ul[R "UR"UR$R&[R(S9nUR+U5nUR-U5 M [R."USS9nUR1S5n UbXn X4$[3URR SR555R7S5n XU - n X4$[R "UR$R&[R(S9n [R "UR8URR$R&[R(S9n X4$)Nrr6r)dtyperXr)rDrrrra enumeratesetuprrrrchunk_durationrr chunk_startrF as_tensorr6r.Tfloat32rbrrrrkeysrgtargets) r:rgr target_indtrackkrr.rrrvocinds rrkMUSDBDataset.__getitem__Bs  )?)? ?@ :: T%6%6%6&txx~~i'@A [[(!"J(("MM$((//:E(,'8'8$$*NN1enntGXGX6X$Y! f(=(C(C(E(EU]][11%8$$U+#B(KK 15E ! A%%t dhhnnY7<<>?EEhOf %t  U]]CA dkk : @ @ B B%--XAt rcZ[URR5UR-$r+)rrDrrros rrpMUSDBDataset.__len__ts 488??#d&<&<<rss >>  CD6    % 4. @## f EQVQ]Q]    %,, 5;;##++@f.  # #f.+3+=+=f. <x'9'9 9:f.RR<\R    >  S"=%%'GD!  !3!34)6vt)D&M=$ Z99;<  YYuS/0 1Q11771: 0I0I#II 999 OOGc!f,w6]=V=V W OOGc!f,w6]=V=V W 2 )+BT+IJ %s='9: *C ,>?"&MKK$$//?? 0M -(1 agg)Kr