P/jH*dZddlmZddlZddlZddlZddlmZddl m Z ddl Z ddl mZddlmZdd lmZdd lmZmZej,eeefZGd d ej0ZGd dZdZ ddej,eefdej,ej8efdededededededej@edej8fdZ!y)z Code to apply a model to a mix. It will handle chunking with overlaps and inteprolation between chunks, as well as the "shift trick". )ThreadPoolExecutorN)nn) functional)Demucs)HDemucs)HTDemucs) center_trimDummyPoolExecutorceZdZ ddejedej ejejedej effd Ze defdZ dZ xZ S) BagOfModelsmodelsweightssegmentct|t|dkDsJ|d}|D]]}|j|jk(sJ|j|jk(sJ|j |j k(sJ|W||__|j |_|j|_|j|_tj||_ |/|Dcgc]}|jDcgc]}dc}}}||_ yt|t|k(sJ|D]%}t|t|jk(r%J||_ ycc}wcc}w)a Represents a bag of models with specific weights. You should call `apply_model` rather than calling directly the forward here for optimal performance. Args: models (list[nn.Module]): list of Demucs/HDemucs models. weights (list[list[float]]): list of weights. If None, assumed to be all ones, otherwise it should be a list of N list (N number of models), each containing S floats (S number of sources). segment (None or float): overrides the `segment` attribute of each model (this is performed inplace, be careful is you reuse the models passed). rN?) super__init__lensources samplerateaudio_channelsrr ModuleListrr) selfrrrfirstother_weight __class__s O/root/repos/pocs/video-evals/.venv/lib/python3.12/site-packages/demucs/apply.pyrzBagOfModels.__init__sH 6{Qq  (E==EMM1 11##u'7'77 77''5+?+?? ??" '  ($22**}} mmF+ ?%++e,%N#U##Ar)r c*eZdZddZedZdZy) TensorChunkNc$|jd}|dk\sJ||ksJ|||z }nt||z |}t|tr&|j|_||j z|_n||_||_||_|j|_y)Nr)shaper&r%r8tensoroffsetlengthdevice)rr<r=r> total_lengths r rzTensorChunk.__init__Ps||B' {{ $$$ >!F*F.7F fk * --DK 6==0DK DK DK mm r)cbt|jj}|j|d<|S)Nr:)listr<r;r>)rr;s r r;zTensorChunk.shapecs(T[[&&'KKb  r)cP||jz }|jjd}|dk\sJ|j|dzz }||z}t d|}t ||}||z }||z } t j|jd||f|| f} | jd|k(sJ| S)Nr:r.)r>r<r;r=maxr&Fpad) r target_lengthdeltar@startend correct_start correct_endpad_left pad_rightouts r paddedzTensorChunk.paddedis +{{((, zz eqj(m#Au ,,  5(+% eeDKK];%> >?(IAVWyy} --- r))rN)r.r/r0rr5r;rQr)r r8r8Os $& r)r8ctt|tr|St|tjsJt|S)N)r%r8thTensor)tensor_or_chunks r tensor_chunkrW|s1/;//299555?++r)r(mixshiftssplitoverlaptransition_powerprogress num_workersrr!c  | |j}ntj|}| *|dkDr|jdk(r t|} n t } ||||||| | d} t |t r0d} dgt|jz} t|j|jD]\}}tt|jj}|j|t!||fi| }|j|t#|D])\}}|dd|ddddfxx|zcc<| |xx|z cc<+| |z } ~t | tj$sJt'| j(dD]}| dd|ddddfxx| |zcc<| S|j||j+|dk\sJd|j(\}}}|rd| d<t-d |j.z}t1|}t |t2sJ|j5|d |zz}d}t'|D]G}t7j8d|}t3||||z|z }t!||fi| }||d ||z dfz }I||z}t |tj$sJ|S|red | d <tj:|t|j|||j}tj:||j}| |j<} | | dkDsJt-|j.| z}t-d|z |z}t'd||} t?tA||j.z d}!tjBtjDd|d zdz|tjD||d zz dd|g}"t|"|k(sJ|"|"jGz |z}"g}#| D]@}t3|||}$| jHt ||$fi| }%|#jK|%|f||z }B|rtMjL|#|!dd}#|#D]\}%}|%jO}&|&j(d}'|d |||zfxx|"d|'|&zj|jz cc<||||zxxx|"d|'j|jz ccc|jQdkDsJ||z}t |tj$sJ|St |tRr| t-| |j.z}(n tU|dr|jW|}(n|}(t1|}t |t2sJ|j5|(j|}tjX5||}dddt tj$sJt[||S#1swY1xYw)a Apply model to a given mixture. Args: shifts (int): if > 0, will shift in time `mix` by a random amount between 0 and 0.5 sec and apply the oppositve shift to the output. This is repeated `shifts` time and all predictions are averaged. This effectively makes the model time equivariant and improves SDR by up to 0.2 points. split (bool): if True, the input will be broken down in 8 seconds extracts and predictions will be performed individually on each and concatenated. Useful for model with large memory footprint like Tasnet. progress (bool): if True, show a progress bar (requires split=True) device (torch.device, str, or None): if provided, device on which to execute the computation, otherwise `mix.device` is assumed. When `device` is different from `mix.device`, only local computations will be on `device`, while the entire tracks will be stored on `mix.device`. num_workers (int): if non zero, device is 'cpu', how many threads to use in parallel. segment (float or None): override the model segment parameter. Nrcpu)rYrZr[r\r]r?poolrgrz-transition_power < 1 leads to weird behavior.rYg?rD.FrZ)r?z.2fr:xseconds) unit_scalencolsunit valid_length).r?rTtyperr r%r rrziprrnextiter parametersto apply_model enumeraterUranger;evalintrrWr8rQrandomrandintzerosrr$formatcatarangerEsubmitappendtqdmresultr&r hasattrrgno_gradr ))r(rXrYrZr[r\r]r?r^rrakwargs estimatestotals sub_model model_weightsoriginal_model_devicerPk inst_weightbatchchannelsr> max_shift padded_mixrr=shifted shifted_out sum_weightsegment_lengthstrideoffsetsscalerfutureschunkfuture chunk_out chunk_lengthrgs) r rnrns6~6" | ?v{{e3%k2D$&D, $F%%13 EMM**(+ELL%--(H  $I}$(i.B.B.D)E$F$M$M ! LL i77C LL. /"+M": );Aq!QJ;.q [(  )  I )RYY///yq)* /A aAqj !VAY . ! / HHV JJL q Q"QQ !iiE8V xe.../ 3#{+++ZZY 67 v 9A^^Ay1F!*ffy6H66QRG%eW??K ;sI$6$778 8C  9 v #ryy))) whhuc%--0(F3::VXXfSZZ8 ?mmG"w|33!%"2"2W"<=a'k^3466*fVe&6&66>?1n&9A&=fM>Na4G#GBW]^`a6{n,,,6::<'*:: %FV^ B 4 vf~5 6&,:O:R:RSVS]S]:^ ^ 6  _ ~~!### z#ryy)))  eX &7+>w)9)99:L UN + --f5L!L3#{+++ZZ -008 ZZ\ $ #C $#ryy)))3'' $ $s + W$$W-) rTg?rFNrNN)"__doc__concurrent.futuresrrstypingr1torchrTrtorch.nnrrFr{demucsrhdemucsrhtdemucsr utilsr r Unionr3Moduler r8rWrUrrboolr$r4rnrRr)r rs 2 $ 1 (*+1A"))1Ah**Z,04AC/3DH M(rxx U 23M(XXbii45M(M((,M(M(:?M( M( ! M( 02{{5/A M( !ii M(r)