P/jrBddlZddlZddlZddlZddlmZddlmZddl m Z ddl m Z m Z ddlmZGdd ej Zd Zd ZGd d ej ZGddej ZGddej Zy)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. ct|| |dzdk(sJ||_tjd||||_tj d|z||_||_y)NrT) bidirectional num_layers hidden_size input_size) super__init__ max_stepsrLSTMlstmLinearlinearskip)selfdimlayersrr __class__s P/root/repos/pocs/video-evals/.venv/lib/python3.12/site-packages/demucs/demucs.pyrzBLSTM.__init__s`  IMQ$666"GG$6s_bc iiC-  c 6|j\}}}|}d}|jc||jkDrT|j}|dz}t|||} | jd} d}| jddddj d||}|jddd}|j |d}|j |}|jddd}|rg} |j |d|} dz} t D]m} | dk(r | j| dd| ddd| f(| | dz k(r| j| dd| dd| dfO| j| dd| dd| | fotj| d} | dd|f} | }|jr||z}|S) NFrTrr.) shaperrpermutereshaperrrangeappendtorchcatr)rxBCTyframedwidthstrideframesnframesoutlimitks rforwardz BLSTM.forward"s''1a  >> %!dnn*<NNEaZFAuf-Fll1oGFq!Q*222q%@A IIaA  IIaLO KKN IIaA  CYYq"a/FaKE7^ >6JJvaAww&678'A+%JJvaAuvo67JJvaAueV|&;<=  >))C$Cc2A2g,CA 99AAr )rNF)__name__ __module__ __qualname____doc__rr8 __classcell__rs@rr r s !r r c|jjj}||z dz}|jxj|zc_|j |jxj|zc_yy)zTRescale initial weight scale. It is unclear why it helps but it certainly does. ?N)weightstddetachdatabias)conv referencerBscales r rescale_convrIFsa ++//  " " $C 9_s "EKK yy %r c|jD]W}t|tjtjtj tj fsLt||YyN)modules isinstancerConv1dConvTranspose1dConv2dConvTranspose2drI)modulerGsubs rrescale_modulerTPsI~~) cBIIr'9'9299bFXFXY Z i ()r c BeZdZdZ ddedededeffd ZdZxZS) 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`. channelscompressdepthinitc (t|| dzdk(sJ||_||_t ||_|dkD} d} |rd} t ||z }| rtj}ntj}tjg|_ t|j D]}| rd|znd}|| dzz}tj||| ||| ||tj|d|zd| d|ztjdt||g}|r|j!dt#|||| r|j!dt%|dd d tj&|}|jj)|y ) 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 BKKMr c.tjd|S)Nrr GroupNormr_s rraz DConv.__init__..{s Q 2r )dilationpaddingr")headsndecayT)rrrN)rrrWrXabsrYintrGELUReLU ModuleListrr'rNGLUr insert LocalStater Sequentialr()rrWrXrYrZnormattnrgrhrgelukerneldilatenorm_fnhiddenactr`rerfmodslayerrs rrzDConv.__init__]sl( zQ    Z * 2GX() ''C''CmmB' tzz" &A!'qAvQH&A+.G (FFXwW &!h,2H %rvvay8T* D Az&fMN AuVA4PQMM4(E KK  u % &r c>|jD] }|||z}|SrK)r)rr+r|s rr8z DConv.forwards'[[ EE!H A r ) r r-C6?TFr r FTr"T) r9r:r;r<rkfloatrr8r=r>s@rrVrVVs= Z^LP"&8&8&8&#8&QV8&tr rVc <eZdZdZddedededeffd ZdZxZS) rqaLocal 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. rWrgnfreqsrhct|||zdk(s J||f||_||_||_t j ||d|_t j ||d|_t j ||d|_ |rt j |||zd|_ |rt j |||zd|_ |jjxjdzc_ |jjJd|jjjddt j |||zz|d|_y)Nrrg{Gz?)rrrgrrhrrNcontentquerykey query_freqs query_decayrArDrEproj)rrWrgrrhrs rrzLocalState.__init__s" %1$7x&77$   yy8Q7 YYx15 99Xx3 !yy56>1ED  !yy56>1ED     # # ( (D 0 (##((4 44,.D   ! ! & &q )IIh71E r c|j\}}}|j}tj||j|j }|dddf|dddfz }|j |j||d|}|j|j||d|} tjd| |} | | jddzz} |jrtjd|jdz|j|j } tjdtjz|z| jdddz } |j|j||d||jdzz } | tjd| | z } |jrtjd|jdz|j|j }|j!|j||d|}tj"|dz }|jddd |j%z|jdzz }| tjd||z } | j'tj(|| jtj*dtj,| d }|j/|j||d|}tjd ||}|jr/tjd | }tj0||gd}|j3|d|}||j5|zS) N)devicedtyper#zbhct,bhcs->bhtsrr@rzfts,bhfs->bhtsirzbhts,bhct->bhcszbhts,fts->bhfs)r$rgr)arangerrrviewreinsumrcosmathpirrhrsigmoidrj masked_fill_eyeboolsoftmaxrr*r&r)rr+r,r-r.rgindexesdeltaquerieskeysdotsperiods freq_kernelfreq_qdecaysdecay_q decay_kernelweightsrresulttime_sigs rr8zLocalState.forwards''1a ,,qA4 747#33**Q-$$Qr15xx{5"a0||-tW=  1 s"" ;;ll1dkkAoahhaggVG))AK%$7',,r1a:P$PQK%%a(--aA>PSASSF ELL!1;G GD ;;\\!T[[1_QXXQWWUF&&q)..q%Q?GmmG,q0G#[[Q22UYY[@4;;PSCSSL ELL!1<I ID %))AdkkLdS--!,,,q/&&q%Q7/'B ;;||$4g{KHYY115F2q)499V$$$r )r rr )r9r:r;r<rkrr8r=r>s@rrqrqs3 FFSFcFsF&"%r rqcreZdZe dfd ZdZdZdfd ZxZS)Demucsc t&|||_||_||_| |_| |_||_||_||_ ||_ ||_ ||_ tj|_tj|_tj|_| rtj$d}d}ntj&}d}| rtj(}ntj&}|}d}t+|D]}d} || k\rfd} g}!|!tj,|||| | ||gz }!||k\}"||k\}#|dzr|!t/|||||"|#gz }!|r)|!tj,|||zd| ||z|gz }!|jj1tj2|!g}$|dkDr|}%nt5|j|z}%|r1|$tj,|||zd| zdz| | ||z|gz }$|dzr|$t/|||||"|#gz }$|$tj6||%|| |gz }$|dkDr|$| |%|gz }$|j j9dtj2|$|}t;||z}|}|rt=|||_nd |_|rtA|| y y ) 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 rraz!Demucs.__init__..Fs  r c0tj|SrKrc)r` norm_groupss rraz!Demucs.__init__..HsBLLa$@r )rYrZrXrtr)rfN)rG)!rraudio_channelssources kernel_sizecontextr2rYresamplerW normalize sampleratesegmentrrnencoderdecoder skip_scalesrormrlr'rNrVr(rrlenrOrprkr rrT)'rrrrWgrowthrYrewrite lstm_layersrr2rruglu norm_startsr dconv_mode dconv_depth dconv_comp dconv_attn dconv_lstm dconv_initrrrescalerr activationch_scaleact2 in_channelsrfindexrxencodertrdecode out_channelsrs' ` rrzDemucs.__init__sX , &       "$ }} }} ==? AJHJH 77D77D$ 5\( .E-G #@F  +xfE! F J&DJ&DA~5:*44dLMMIIh8(;Q?Hx/0*>> LL   v 6 7Fqy* "4<<0>A IIh8(;Q[1_V]^Hx/0*>>A~5:*44dLMM r))(L"FG=> >Fqy7<0$&99 LL  2== :"K6H,-HQ( .T h 4DIDI  47 3 r c|jr|dz}t|jD]@}tj||j z |j z dz}td|}Bt|jD]!}|dz |j z|j z}#|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'rYrceilrr2maxrk)rlength_idxs r valid_lengthzDemucs.valid_lengthxs == aKFtzz" $AYY)9)9 9T[[HIAMFF^F $$ CCqjDKK/$2B2BBF C ==YYvz*F6{r c|}|jd}|jrE|jdd}|jdd}|jdd}||z d|zz }nd}d}|j ||z }t j ||dz||dzz f}|jrtj|dd}g}|jD]} | |}|j||jr|j|}|jD]*} |jd} t| |} | || z},|jrtj|dd}||z|z}t||}|j!|j#dt%|j&|j(|j#d}|S)Nr#rT)rkeepdimgh㈵>rr)r$rmeanrBrFpadrjulius resample_fracrr(rrpoprrsizerrr) rmixr+rmonorrBrsavedrrrs rr8zDemucs.forwards  >>88480D99T92D((r4(0CTdSj)ADC!!&)F2 EE!eqj%%1*"45 6 ==$$Q1-All Fq A LLO  99 ! All !F99R=DtQ'Dq4x A ! ==$$Q1-A GdN 6 " FF166!9c$,,/1D1DaffRj Qr ct|jD]B}dD];}dD]4}|d|d|}|d|d|}||vs||vs!|j|||<6=Dt|||y)N)rr)rErA.z.3.z.2.)strict)r'rYrrload_state_dict) rstaterrabnewoldrs rrzDemucs.load_state_dicts$ 4C+ 4+4ACqS,CCqS,Ce|5(8%*YYs^c 4 4 4 f5r )r@g@Trr rTTr r rrr r r r~TTg?iD()T) r9r:r;rrrr8rr=r>s@rrrsx!" !EY4Y4v.%N 6 6r r)rtypingtprr)rtorch.nnrrstatesrutilsrr transformerr Moduler rIrTrVrqrr rrss  $ &#/BII/d ) DBIIDN;%;%|d6RYYd6r