P/jv dZddlmZddlZddlZddlmZddlZddlm Z ddl m Z ddl mZmZdd lmZdd lmZmZdd ej*d ej,eefd edefdZGdde j6ZGdde j6ZGdde j6ZGdde j6ZGdde j6Z y)zB This code contains the spectrogram and Hybrid version of Demucs. )deepcopyN)wiener)nn) functional)DConvrescale_module) capture_init)spectroispectroxpaddingsmodevaluecv|}|jd}|\}}|dk(rLt||}||kr;||z dz} t|| } | | z } || z || z f}tj|| | f}tj||||} | jd||z|zk(sJ| d|||zf|k(j sJ| S)zTiny wrapper around F.pad, just to allow for reflect padding on small input. If this is the case, we insert extra 0 padding to the right before the reflection happen.reflectr.)shapemaxminFpadall) r rrrx0length padding_left padding_rightmax_pad extra_padextra_pad_rightextra_pad_leftouts Q/root/repos/pocs/video-evals/.venv/lib/python3.12/site-packages/demucs/hdemucs.pypad1dr$s B WWR[F"*L- ylM2 W &(1,I!-;O&8N$~5}7VWHa./:;A %%8T5 )C 99R=F\1MA AA A \<88 9R ? D D FF F JcJeZdZdZ ddededeffd ZedZdZ xZ S) ScaledEmbeddingzt Boost learning rate for embeddings (with `scale`). Also, can make embeddings continuous with `smooth`. num_embeddings embedding_dimscalect|tj|||_|rt j |jjjd}|t jd|dzj|jdddfz }||jjjdd|jjxj|zc_||_ y)Nrdimr) super__init__r Embedding embeddingtorchcumsumweightdataarangetosqrtr*)selfr(r)r*smoothr4 __class__s r#r/zScaledEmbedding.__init__0s nmD \\$.."7"7"<"eZdZddddddddiddf fd Zd dZxZS) HEncLayerrFTrct|d}| rfd}| r|dz} nd} tj}||_||_||_||_| |_| |_ |r|dg}|dg}| dg} tj}|||||| |_ |jry|||_ d|_ | r*||d|zdd| zzd| |_ |d|z|_d|_|rt!|fi| |_yy)a#Encoder layer. This used both by the time and the frequency branch. Args: chin: number of input channels. chout: number of output channels. norm_groups: number of groups for group norm. empty: used to make a layer with just the first conv. this is used before merging the time and freq. branches. freq: this is acting on frequencies. dconv: insert DConv residual branches. norm: use GroupNorm. context: context size for the 1x1 conv. dconv_kw: list of kwargs for the DConv class. pad: pad the input. Padding is done so that the output size is always the input size / stride. rewrite: add 1x1 conv at the end of the layer. c*tjSr=rIdentityds r#z$HEncLayer.__init__..[ BKKMr%c0tj|Sr=r GroupNormrR norm_groupss r#rSz$HEncLayer.__init__..] [! <r%rLrrN)r.r/rConv1dfreq kernel_sizestrideemptynormrConv2dconvnorm1rewritenorm2dconvr)r9chinchoutr^r_rYr`r]rgracontextdconv_kwrrenorm_fnklassr;s ` r#r/zHEncLayer.__init__Fs( ) 12Aq-5<<-/05==3E3E3J3J1bRSUV3WWY2aELLM12CA:a$7ell]a%7$7:;CA:aELLA$5$69:C C ! [- \ii!$yy&&+C 99J9 r%)NN)r@rArBrCr/r?rGrHs@r#rrs$4:r%rc @eZdZdddddddddidddf fd ZdZxZS) HDecLayerFrKrLrTct|d}| rfd}| r|dz} nd} | |_||_||_||_||_||_||_| |_ ||_ tj}tj}|r(|dg}|dg}tj}tj}||||||_|||_|j ryd|_|rK|r||d|zdd| zzd| |_n||d|zddd| zzgdd| g|_|d|z|_d|_| rt+|fi| |_yy)zW Same as HEncLayer but for decoder. See `HEncLayer` for documentation. c*tjSr=rOrQs r#rSz$HDecLayer.__init__..rTr%c0tj|Sr=rVrXs r#rSz$HDecLayer.__init__.. rZr%rLrrNr[)r.r/rrr]rhr`r_r^rarrr\ConvTranspose1drbConvTranspose2drrfrerdrgr)r9rhrirr^r_rYr`r]rgrarjrkrrrerlrmklass_trr;s ` r#r/zHDecLayer.__init__s^ )