VhrB ddlZddlZddlZddlZddlmZddlmZddl m Z ddl m Z m Z ddlmZGdd ejZd Zd ZGd d ejZGddejZGddejZdS)N)nn) functional) capture_init) center_trimunfold) LayerScalec*eZdZdZdfd ZdZxZS)BLSTMz 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. rNFct| |dzdksJ||_tjd||||_tjd|z||_||_dS)NrT) bidirectional num_layers hidden_size input_size) super__init__ max_stepsrLSTMlstmLinearlinearskip)selfdimlayersrr __class__s S/var/www/html/movieo_spanner_bot/venv/lib/python3.11/site-packages/demucs/demucs.pyrzBLSTM.__init__sw  IMQ$6$6$66"G$6s_bccc iC--  c ~|j\}}}|}d}|jd||jkrY|j}|dz}t|||} | jd} d}| ddddd||}|ddd}||d}||}|ddd}|rg} ||d||} |dz} t| D]} | dkr'| | dd| ddd| f/| | dz kr&| | dd| dd| df^| | dd| dd| | ftj | d} | dd|f} | }|j r||z}|S) NFrTrr.) shaperrpermutereshaperrrangeappendtorchcatr)rxBCTyframedwidthstrideframesnframesoutlimitks rforwardz BLSTM.forward"s'1a  > %!dn*<*<NEaZFAuf--Fl1oGFq!Q**222q%@@A IIaA   IIaLLO KKNN IIaA    CYYq"a//FaKE7^^ > >66JJvaaaAAAww&678888'A+%%JJvaaaAAAuvvo67777JJvaaaAAAueV|&;<====)C$$Cc2A2g,CA 9 AAr )rNF)__name__ __module__ __qualname____doc__rr8 __classcell__rs@rr r sV !!!!!!!r r c|j}||z dz}|jxj|zc_|j|jxj|zc_dSdS)zTRescale initial weight scale. It is unclear why it helps but it certainly does. ?N)weightstddetachdatabias)conv referencerBscales r rescale_convrIFsk +//   " " $ $C 9_s "EK y %r c|D]N}t|tjtjtjtjfrt||OdSN)modules isinstancerConv1dConvTranspose1dConv2dConvTranspose2drI)modulerGsubs rrescale_modulerTPs[~~)) cBIr'929bFXY Z Z ) i ( ( ())r c BeZdZdZ dded ed ed effd Zd ZxZS)DConva  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`. r r-C6?TFr"channelscompressdepthinitc xt| dzdksJ||_||_t ||_|dk} d} |rd} t ||z }| r tj}n tj }tj g|_ t|jD]}| rd|znd}|| dzz}tj ||| ||| ||tj |d|zd| d|ztjdt||g}|r&|dt#|||| r'|dt%|dd d tj|}|j |d S) a Args: channels: input/output channels for residual branch. compress: amount of channel compression inside the branch. depth: number of layers in the residual branch. Each layer has its own projection, and potentially LSTM and attention. init: initial scale for LayerNorm. norm: use GroupNorm. attn: use LocalAttention. heads: number of heads for the LocalAttention. ndecay: number of decay controls in the LocalAttention. lstm: use LSTM. gelu: Use GELU activation. kernel: kernel size for the (dilated) convolutions. dilate: if true, use dilation, increasing with the depth. rrrc(tjSrKrIdentityds rz DConv.__init__..ys BKMMr c,tjd|S)Nrr GroupNormr`s rrbz DConv.__init__..{s Q 2 2r )dilationpaddingr")headsndecayT)rrrN)rrrXrYabsrZintrGELUReLU ModuleListrr'rNGLUr insert LocalStater Sequentialr()rrXrYrZr[normattnrhrirgelukerneldilatenorm_fnhiddenactrarfrgmodslayerrs rrzDConv.__init__]s( zQ    ZZ *)  322GX())  'CC'CmB'' tz"" & &A!'.qAvvQH&A+.G (FFXwWWW &!h,22H %%rvayy8T** D O Az&fMMMNNN R AuVA4PPPQQQM4(E K  u % % % % & &r c6|jD]}|||z}|SrK)r)rr+r}s rr8z DConv.forwards+[  EEE!HH AAr ) r rrWTFr r FTr"T) r9r:r;r<rlfloatrr8r=r>s@rrVrVVs Z^LP"&8&8&8&8&#8&QV8&8&8&8&8&8&tr rVc <eZdZdZd dedededeffd Zd ZxZS) rraLocal 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. r rrXrhnfreqsric|t||zdks J||f||_||_||_t j||d|_t j||d|_t j||d|_ |rt j|||zd|_ |r\t j|||zd|_ |j j xj dzc_ |j jJd|j jj dd<t j|||zz|d|_dS)Nrrg{Gz?)rrrhrrirrNcontentquerykey query_freqs query_decayrArDrEproj)rrXrhrrirs rrzLocalState.__init__s: %1$$$x&7$$$   y8Q77 Yx155 9Xx33  F!y56>1EED   /!y56>1EED    # ( (D 0 ( (#(444,.D  ! &qqq )Ih71EE r c|j\}}}|j}tj||j|j}|dddf|dddfz }||||d|}||||d|} tj d| |} | | jddzz} |j rtjd|j dz|j|j} tj dtj z|z| dddz } ||||d||j dzz } | tj d| | z } |jrtjd|jdz|j|j}||||d|}tj|dz }|ddd |z|jdzz }| tj d||z } | tj|| jtjdtj| d }||||d|}tj d ||}|j r-tj d || }tj||gd}||d|}|||zS) N)devicedtyper#zbhct,bhcs->bhtsrr@rzfts,bhfs->bhtsirzbhts,bhct->bhcszbhts,fts->bhfs)r$rhr)arangerrrviewreinsumrcosmathpirrirsigmoidrk masked_fill_eyeboolsoftmaxrr*r&r)rr+r,r-r.rhindexesdeltaquerieskeysdotsperiods freq_kernelfreq_qdecaysdecay_q decay_kernelweightsrresulttime_sigs rr8zLocalState.forwards'1a ,qAAA4 747#33**Q--$$Qr155xx{{5"a00|-tW==  1 s"" ; Hl1dkAoahagVVVG)AK%$7',,r1a:P:P$PQQK%%a((--aA>>PSASSF EL!1;GG GD ; J\!T[1_QXQWUUUF&&q))..q%Q??GmG,,q0G#[[Q222UYY[[@4;PSCSSL EL!1<II ID %)AdkLLLdSSS-!,,,,,q//&&q%Q77/'BB ; 6|$4g{KKHY1155F2q))499V$$$$r )r rr )r9r:r;r<rlrr8r=r>s@rrrrrs FFFSFcFsFFFFFF&"%"%"%"%"%"%"%r rrcxeZdZe dfd ZdZdZdfd ZxZS)Demucsr@@Trr rrW皙?D(c t||_||_||_| |_| |_||_||_||_ ||_ ||_ ||_ tj|_tj|_tj|_| rtjd}d}ntj}d}| r tj}n tj}|}d}t+|D]}d} || krfd} g}!|!tj|||| | ||gz }!||k}"||k}#|dzr|!t/|||||"|#gz }!|r+|!tj|||zd| ||z|gz }!|jtj|!g}$|dkr|}%nt5|j|z}%|r3|$tj|||zd| zdz| | ||z|gz }$|dzr|$t/|||||"|#gz }$|$tj||%|| |gz }$|dkr|$| |%|gz }$|jdtj|$|}t;||z}|}|rt=|||_nd |_|rtA|| d Sd S) ab Args: sources (list[str]): list of source names audio_channels (int): stereo or mono channels (int): first convolution channels depth (int): number of encoder/decoder layers growth (float): multiply (resp divide) number of channels by that for each layer of the encoder (resp decoder) depth (int): number of layers in the encoder and in the decoder. rewrite (bool): add 1x1 convolution to each layer. lstm_layers (int): number of lstm layers, 0 = no lstm. Deactivated by default, as this is now replaced by the smaller and faster small LSTMs in the DConv branches. kernel_size (int): kernel size for convolutions stride (int): stride for convolutions context (int): kernel size of the convolution in the decoder before the transposed convolution. If > 1, will provide some context from neighboring time steps. gelu: use GELU activation function. glu (bool): use glu instead of ReLU for the 1x1 rewrite conv. norm_starts: layer at which group norm starts being used. decoder layers are numbered in reverse order. norm_groups: number of groups for group norm. dconv_mode: if 1: dconv in encoder only, 2: decoder only, 3: both. dconv_depth: depth of residual DConv branch. dconv_comp: compression of DConv branch. dconv_attn: adds attention layers in DConv branch starting at this layer. dconv_lstm: adds a LSTM layer in DConv branch starting at this layer. dconv_init: initial scale for the DConv branch LayerScale. normalize (bool): normalizes the input audio on the fly, and scales back the output by the same amount. resample (bool): upsample x2 the input and downsample /2 the output. rescale (float): rescale initial weights of convolutions to get their standard deviation closer to `rescale`. samplerate (int): stored as meta information for easing future evaluations of the model. segment (float): duration of the chunks of audio to ideally evaluate the model on. This is used by `demucs.apply.apply_model`. rrrrc(tjSrKr^r`s rrbz!Demucs.__init__..Fs  r c.tj|SrKrd)ra norm_groupss rrbz!Demucs.__init__..HsBLa$@$@r )rZr[rYrur)rgN)rG)!rraudio_channelssources kernel_sizecontextr2rZresamplerX normalize sampleratesegmentrroencoderdecoder skip_scalesrprnrmr'rNrVr(rslenrOrqrlr rrT)'rrrrXgrowthrZrewrite lstm_layersrr2rrvglu norm_startsr dconv_mode dconv_depth dconv_comp dconv_attn dconv_lstm dconv_initrrrescalerr activationch_scaleact2 in_channelsrgindexryencoderurdecode out_channelsrs' ` rrzDemucs.__init__sX , &       "$ } } =??  AJHHJH  7DD7D$ 5\\( .( .E--G ##@@@@F  +xfEE!! F J&DJ&DA~ M5:*44dLLLMM >Ih8(;Q??GHx/00*>> L   v 6 7 7 7Fqyy* "4<00>A  >Ih8(;Q[1_V]^^^GHx/00*>>A~ M5:*44dLLLMM r)(L"FG===> >Fqyy77<00$$&&99 L  2= : : :"K6H,--HH  h 44DIIDI  4 47 3 3 3 3 3 3 4 4r ch|jr|dz}t|jD]9}tj||jz |jz dz}td|}:t|jD]}|dz |jz|jz}|jrtj|dz }t|S)aX Return the nearest valid length to use with the model so that there is no time steps left over in a convolution, e.g. for all layers, size of the input - kernel_size % stride = 0. Note that input are automatically padded if necessary to ensure that the output has the same length as the input. rr) rr'rZrceilrr2maxrl)rlength_idxs r valid_lengthzDemucs.valid_lengthxs =  aKFtz"" $ $AY)9 9T[HIIAMFF^^FF$$ C CCqjDK/$2BBFF = +Yvz**F6{{r c|}|jd}|jrQ|dd}|dd}|dd}||z d|zz }nd}d}|||z }t j||dz||dzz f}|jrtj |dd}g}|j D]"} | |}| |#|j r| |}|j D]5} |d} t| |} | || z}6|jrtj |dd}||z|z}t||}||dt%|j|j|d}|S)Nr#rT)rkeepdimgh㈵>rr)r$rmeanrBrFpadrjulius resample_fracrr(rrpoprrsizerrr) rmixr+rmonorrBrsavedrrrs rr8zDemucs.forwards  > 884800D99T922D((r4(00CTdSj)AADC!!&))F2 E!eqj%%1*"45 6 6 = .$Q1--Al  Fq A LLOOOO 9  ! Al ! !F99R==DtQ''Dq4x  AA = .$Q1--A GdN 6 " " FF166!99c$,//1DaffRjj Q Qr ct|jD]@}dD];}dD]6}|d|d|}|d|d|}||vr||vr||||<7s@rrrs!" !EY4Y4Y4Y4Y4\Y4v.%%%N 6 6 6 6 6 6 6 6 6 6r r)rtypingtprr)rtorch.nnrrstatesrutilsrr transformerr Moduler rIrTrVrrrr rrs{  $$$$$$ &&&&&&&&######/////BI///d   ))) DDDDDBIDDDN;%;%;%;%;%;%;%;%|d6d6d6d6d6RYd6d6d6d6d6r