9i6SSKrSSKrSSKJrJr SSKrSSKJr SSKJs J r SSK J r SSK JrJr SSKJrJr SSKJrJr SSKJr SSKJr S r\R6"\R8\ R:S 9"S S \55rS r"SS\R@5r!"SS\RD5r#"SS5r$"SS\R@5r%"SS\R@5r&"SS\R@5r'g)N)AnyDict)Models)MODELS TorchModel)MossFormerModuleScaledSinuEmbedding)CumulativeLayerNormGlobalLayerNorm)Tensor)Tasks:0yE>) module_namec`^\rSrSrSrS\4U4SjjrS\S\\\ 44Sjr S Sjr S r S r U=r$) MossFormerz]Library to support MossFormer speech separation. Args: model_dir (str): the model path. model_dirc,>[TU]"U/UQ70UD6 [USUSS9Ul[ USSUSUSUSS9Ul[ USUS[US US US US US USUS5USUSS9UlUSUl g)N kernel_size out_channels)rr in_channelsstridebias)rrrrr num_blocksd_model attn_dropout group_size query_key_dimexpansion_factorcausalnormnum_spks)r"r#) super__init__EncoderencoderDecoderdecoderMossFormerMaskNet MossFormerMmask_netr#)selfrargskwargs __class__s m/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/modelscope/models/audio/separation/mossformer.pyr%MossFormer.__init__!s 4T4V4}-/1 }-}-(# ! * = ! > " |,fY.?~.|0D/8J1Kx( *J') z* inputsreturnc URU5nURU5n[R"U/UR-5nX#-n[R "[ UR5Vs/sH%nURXE5RS5PM' snSS9nURS5nURS5nXx:a[R"USSSXx- 45nU$USS2SU2SS24nU$s snf)Ndimrr) r'r,torchstackr#catranger) unsqueezesizeFpad) r-r4mix_west_masksep_hi est_sourcet_origint_ests r1forwardMossFormer.forward7s V$==' UGdmm34 YYt}}- -A UX&004-   ;;q>"  zAq!X5E+FGJ$Ayy!O4J s2,C:c&U(d URnU(d[R"S5nURR [R "[ RRUS5US9SS9 URR [R "[ RRUS5US9SS9 URR [R "[ RRUS5US9SS9 g)Ncpuz encoder.bin) map_locationT)strictz decoder.binz masknet.bin) rr:devicer'load_state_dictloadospathjoinr)r,)r- load_pathrOs r1load_check_pointMossFormer.load_check_pointNsI\\%(F $$ JJ Y 6V M %  $$ JJ Y 6V M %  %% JJ Y 6V M & r3cT[URURURS9$)N)r'r)masknet)dictr'r)r,)r-s r1as_dictMossFormer.as_dict`s$LL$,, O Or3)r)r'r,r#)NN)__name__ __module__ __qualname____firstlineno____doc__strr%r rrrIrVr[__static_attributes__ __classcell__r0s@r1rrs@ +#+,fc3h.$OOr3rcUS:Xa [XSS9$US:Xa [USS9$US:Xa[R"SUSS9$[R"U5$) z5Just a wrapper to select the normalization type. glnT)elementwise_affineclnlnrreps)r r nn GroupNorm BatchNorm1d)r"r9shapes r1 select_normrqesU u}sdCC u}"34@@ t|||As--~~c""r3ch^\rSrSrSrS S\S\S\4U4SjjjrS\R4Sjr S r U=r $) r&sa:Convolutional Encoder Layer. Args: kernel_size: Length of filters. in_channels: Number of input channels. out_channels: Number of output channels. Examples: >>> x = torch.randn(2, 1000) >>> encoder = Encoder(kernel_size=4, out_channels=64) >>> h = encoder(x) >>> h.shape # torch.Size([2, 64, 499]) rrrc x>[[U] 5 [R"UUUUS-SSS9UlX0lg)NrF)rrrrgroupsr)r$r&r%rmConv1dconv1dr)r-rrrr0s r1r%Encoder.__init__sA gt%'ii#%#!#   'r3xcURS:Xa[R"USS9nURU5n[R "U5nU$)a1Return the encoded output. Args: x: Input tensor with dimensionality [B, L]. Returns: Encoded tensor with dimensionality [B, N, T_out]. where B = Batchsize L = Number of timepoints N = Number of filters T_out = Number of timepoints at the output of the encoder rr8)rr:r>rxr@relur-rzs r1rIEncoder.forwardsA   q q)A KKN FF1Ir3)rxr)ru@r) r]r^r_r`raintr%r:r rIrcrdres@r1r&r&ssL  %&%'$% '! '" '" ' 'r3r&c8^\rSrSrSrU4SjrU4SjrSrU=r$)r(afA decoder layer that consists of ConvTranspose1d. Args: kernel_size: Length of filters. in_channels: Number of input channels. out_channels: Number of output channels. Example --------- >>> x = torch.randn(2, 100, 1000) >>> decoder = Decoder(kernel_size=4, in_channels=100, out_channels=1) >>> h = decoder(x) >>> h.shape torch.Size([2, 1003]) c.>[[U] "U0UD6 gN)r$r(r%)r-r.r/r0s r1r%Decoder.__init__s gt%t6v6r3c>UR5S;a$[SRUR55e[TU]UR5S:XaUO[ R"US55n[ R"U5R5S:Xa[ R"USS9nU$[ R"U5nU$)zReturn the decoded output. Args: x: Input tensor with dimensionality [B, N, L]. where, B = Batchsize, N = number of filters L = time points )ruz{} accept 3/4D tensor as inputrrr8) r9 RuntimeErrorformatr]r$rIr:r>squeeze)r-rzr0s r1rIDecoder.forwards 557& ?FF    GOAA5??1a3H I ==    !Q & aQ'A a Ar3) r]r^r_r`rar%rIrcrdres@r1r(r(s 7r3r(c$\rSrSrSrSrSrSrg) IdentityBlockzThis block is used when we want to have identity transformation within the Dual_path block. Example ------- >>> x = torch.randn(10, 100) >>> IB = IdentityBlock() >>> xhat = IB(x) c grr)r-r/s r1_init__IdentityBlock._init__s r3cU$rrr}s r1__call__IdentityBlock.__call__sr3rN)r]r^r_r`rarrrcrr3r1rrs r3rc^^\rSrSrSrSU4SjjrS\R4SjrSr U=r $)r+aThis class implements the transformer encoder. Args: num_blocks : int Number of mossformer blocks to include. d_model : int The dimension of the input embedding. attn_dropout : float Dropout for the self-attention (Optional). group_size: int the chunk size query_key_dim: int the attention vector dimension expansion_factor: int the expansion factor for the linear projection in conv module causal: bool true for causal / false for non causal Example ------- >>> import torch >>> x = torch.rand((8, 60, 512)) #B, S, N >>> net = MossFormerM(num_blocks=8, d_model=512) >>> output, _ = net(x) >>> output.shape torch.Size([8, 60, 512]) c >[T U]5 [UUUUUUUS9UlSSKnUR R RUSS9Ulg)N)r9depthrrr r!rrgư>rk) r$r%r mossformerM speechbrainnnet normalization LayerNormr") r-rrrrrr r!sbr0s r1r%MossFormerM.__init__sX +!'-%' !GG))33G3F r3srccJURU5nURU5nU$)z Args: src: Tensor shape [B, S, N], where, B = Batchsize, S = time points N = number of filters The sequence to the encoder layer (required). rr")r-routputs r1rIMossFormerM.forwards'!!#&6" r3r)Ng?g@F r]r^r_r`rar%r:r rIrcrdres@r1r+r+s7<!""$G* 5<<  r3r+cV^\rSrSrSrSU4SjjrS\R4SjrSr U=r $)ComputeAttentioni%a Computation block for dual-path processing. Args: att_mdl : torch.nn.module Model to process within the chunks. out_channels : int Dimensionality of attention model. norm : str Normalization type. skip_connection : bool Skip connection around the attention module. Example --------- >>> att_block = MossFormerM(num_blocks=8, d_model=512) >>> comp_att = ComputeAttention(att_block, 512) >>> x = torch.randn(10, 64, 512) >>> x = comp_att(x) >>> x.shape torch.Size([10, 64, 512]) cx>[[U] 5 XlX@lX0lUb[ X2S5Ulgg)Nr)r$rr%att_mdlskip_connectionr"rqatt_norm)r-rrr"rr0s r1r%ComputeAttention.__init__<s> .0 .  'A>DM r3rzcURSSS5R5nURU5nURSSS5R5nURbUR U5nUR (aX!-nUnU$)zReturns the output tensor. Args: x: Input tensor of dimension [B, S, N]. Returns: out: Output tensor of dimension [B, S, N]. where, B = Batchsize, N = number of filters S = time points rrur)permute contiguousrr"rr)r-rzatt_outouts r1rIComputeAttention.forwardMs~))Aq!$//1,,w'//!Q*557 99 mmG,G   kG r3)rrr"r)rjTrres@r1rr%s*4 ?"r3rcZ^\rSrSrSrSU4SjjrS\R4SjrSr U=r $)r*ika The dual path model which is the basis for dualpathrnn, sepformer, dptnet. Args: in_channels : int Number of channels at the output of the encoder. out_channels : int Number of channels that would be inputted to the intra and inter blocks. att_model : torch.nn.module Attention model to process the input sequence. norm : str Normalization type. num_spks : int Number of sources (speakers). skip_connection : bool Skip connection around attention module. use_global_pos_enc : bool Global positional encodings. Example --------- >>> mossformer_block = MossFormerM(num_blocks=8, d_model=512) >>> mossformer_masknet = MossFormerMaskNet(64, 64, att_model, num_spks=2) >>> x = torch.randn(10, 64, 2000) >>> x = mossformer_masknet(x) >>> x.shape torch.Size([2, 10, 64, 2000]) c >[[U] 5 XPl[ XAS5Ul[ R"XSSS9UlXpl UR(a[U5Ul [R"[UUUUS95Ul[ R"X"U-SS9Ul[ R"X!SSS9Ul[ R$"5Ul[ R("5Ul[ R,"[ R"X"S5[ R."55Ul[ R,"[ R"X"S5[ R2"55Ulg)NrrF)r)r)r)r$r*r%r#rqr"rmrwconv1d_encoderuse_global_pos_encr pos_enccopydeepcopyrmdl conv1d_out conv1_decoderPReLUpreluReLU activation SequentialTanhrSigmoid output_gate) r-rr att_modelr"r#rrr0s r1r%MossFormerMaskNet.__init__s /1 15  ii qu6"4  " ".|rs  &0+*)+  ??AHOHOAHOV #3bii3l'b  'T">"))>BCryyCLp pr3