0;jiz1dZddlmZmZddlZGddejjZGddeZGdd eZ Gd d eZ Gd d eZ GddeZ GddeZ GddeZdS)zSubsampling layer definition.)TupleUnionNcVeZdZfdZdeeejfdedejfdZxZ S)BaseSubsamplingcdtd|_d|_dS)Nr)super__init__ right_contextsubsampling_rate)self __class__s i/root/voice-cloning/.venv/lib/python3.11/site-packages/chatterbox/models/s3gen/transformer/subsampling.pyr zBaseSubsampling.__init__s/  !offsetsizereturnc8|j||SN)pos_encposition_encoding)r rrs rrz!BaseSubsampling.position_encodings|--fd;;;r) __name__ __module__ __qualname__r rinttorchTensorr __classcell__rs@rrrsq""""" <c5<.?(@< #<(- <<<<<<<)epsrrN) r r rr' SequentialLinear LayerNormDropoutoutrr r r*s rr zLinearNoSubsampling.__init__Os 8&& HOOD$ ' ' H  t  . . H  \ * *   %  !rrr+r,rrch||}|||\}}|||fSr.rArr/s rr1zLinearNoSubsampling.forward\7& HHQKK\\!V,, 7'6!!rr2r3rs@rr8r8Es "S " "5 " % " " " " " "",- "" <" "c5<'( " u|U\5<7 8 """"""""rr8c eZdZdZdedededejjffd Z d dej d ej d e eej fd e ej ej ej ffd Z xZS)Conv1dSubsampling2a!Convolutional 1D subsampling (to 1/2 length). It is designed for Whisper, ref: https://github.com/openai/whisper/blob/main/whisper/model.py Args: idim (int): Input dimension. odim (int): Output dimension. dropout_rate (float): Dropout rate. r"r#r$r%c ttjtj||ddtjtj||dddtj|_||_d|_ d|_ dS)z'Construct an Conv1dSubsampling2 object.r) kernel_sizepadding)rIstriderJN) r r rr'r=Conv1dGELUconvrr r r*s rr zConv1dSubsampling2.__init__s H'' HOOD$AqO A A HMMOO HOOD$AaO K K HMMOO    % !"rrr+r,rrc|d}|dd}||}|dd}|||\}}|||dddd|dzdzddffS)aSubsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1, time). Returns: torch.Tensor: Subsampled tensor (#batch, time', odim), where time' = time // 2. torch.Tensor: Subsampled mask (#batch, 1, time'), where time' = time // 2. torch.Tensor: positional encoding rrKN)r transposerPr)r r+r,rtimer0s rr1zConv1dSubsampling2.forwards(vvayy KK1   IIaLL KK1  \\!V,, 7'6!!!QQQQ(9(9"9:::rr2r3rs@rrFrFts  S5 %*,- ;; <; ;c5<'( ; u|U\5<7 8 ;;;;;;;;rrFc eZdZdZdedededejjffd Z d dej d ej d e eej fd e ej ej ej ffd Z xZS)Conv2dSubsampling4zConvolutional 2D subsampling (to 1/4 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_rate (float): Dropout rate. r"r#r$r%c Rttjtjd|ddtjtj||ddtj|_tjtj||dz dzdz dzz||_ ||_ d|_ d|_ dS)z'Construct an Conv2dSubsampling4 object.rrHrKrMN) r r rr'r=Conv2dReLUrPr>rArr r r*s rr zConv2dSubsampling4.__init__s H'' HOOAtQ * * HMMOO HOOD$1 - - HMMOO    8&& HOODdQh1_q%8Q$>? F FHH$ !"rrr+r,rrc|d}||}|\}}}}||dd||||z}|||\}}|||dddddddfdddddddffS)aSubsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1, time). Returns: torch.Tensor: Subsampled tensor (#batch, time', odim), where time' = time // 4. torch.Tensor: Subsampled mask (#batch, 1, time'), where time' = time // 4. torch.Tensor: positional encoding rrKN) unsqueezerPrrArR contiguousviewr r r+r,rbctfr0s rr1zConv2dSubsampling4.forwards( KKNN IIaLLVVXX 1a HHQ[[A&&113388Aq1uEE F F\\!V,, 7'6!!!QQQ1*-aaaADqDj999rr2r3rs@rrUrUsS5 %.,- :: <: :c5<'( : u|U\5<7 8 ::::::::rrUc eZdZdZdedededejjffd Z d dej d ej d e eej fd e ej ej ej ffd Z xZS)Conv2dSubsampling6zConvolutional 2D subsampling (to 1/6 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_rate (float): Dropout rate. pos_enc (torch.nn.Module): Custom position encoding layer. r"r#r$r%c ttjtjd|ddtjtj||ddtj|_tj||dz dzdz dzz||_ ||_ d|_ d|_ dS)z'Construct an Conv2dSubsampling6 object.rrHrKrW N r r rr'r=rXrYrPr>linearrr r r*s rr zConv2dSubsampling6.__init__s H'' HOOAtQ * * HMMOO HOOD$1 - - HMMOO    hoodqQ/Bq.H&I&*,, $ !rrr+r,rrc|d}||}|\}}}}||dd||||z}|||\}}|||dddddddfdddddddffS)aSubsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1, time). Returns: torch.Tensor: Subsampled tensor (#batch, time', odim), where time' = time // 6. torch.Tensor: Subsampled mask (#batch, 1, time'), where time' = time // 6. torch.Tensor: positional encoding rrKNrMrHr[rPrrirRr\r]rr^s rr1zConv2dSubsampling6.forwards$ KKNN IIaLLVVXX 1a KK Aq))4466;;Aq!a%HH I I\\!V,, 7'6!!!QQQ1*-aaaADqDj999rr2r3rs@rrdrds S  5 %      *,- :: <: :c5<'( : u|U\5<7 8 ::::::::rrdc eZdZdZdedededejjffd Z d dej d ej d e eej fd e ej ej ej ffd Z xZS)Conv2dSubsampling8zConvolutional 2D subsampling (to 1/8 length). Args: idim (int): Input dimension. odim (int): Output dimension. dropout_rate (float): Dropout rate. r"r#r$r%c ttjtjd|ddtjtj||ddtjtj||ddtj|_tj||dz dzdz dzdz dzz||_ ||_ d|_ d|_ dS)z'Construct an Conv2dSubsampling8 object.rrHrKNrhr*s rr zConv2dSubsampling8.__init__$s H'' HOOAtQ * * HMMOO HOOD$1 - - HMMOO HOOD$1 - - HMMOO    hoo taxAo)a/!39 :DBB $ !rrr+r,rrc|d}||}|\}}}}||dd||||z}|||\}}|||dddddddfdddddddfdddddddffS)aSubsample x. Args: x (torch.Tensor): Input tensor (#batch, time, idim). x_mask (torch.Tensor): Input mask (#batch, 1, time). Returns: torch.Tensor: Subsampled tensor (#batch, time', odim), where time' = time // 8. torch.Tensor: Subsampled mask (#batch, 1, time'), where time' = time // 8. torch.Tensor: positional encoding rrKNrkr^s rr1zConv2dSubsampling8.forward7s& KKNN IIaLLVVXX 1a KK Aq))4466;;Aq!a%HH I I\\!V,, 7'6!!!QQQ1*-aaaADqDj9!!!QQQ1*EEErr2r3rs@rrmrms S  5 %      .,- FF <F Fc5<'( F u|U\5<7 8 FFFFFFFFrrmc eZdZdZdedededejjffd Z d dej d ej d e eej fd e ej ej ej ffd Z xZS)LegacyLinearNoSubsamplingr9r"r#r$r%cttjtj||tj|dtj|tj|_ ||_ d|_ d|_ dSr;) r r rr'r=r>r?r@rYrArr r r*s rr z"LegacyLinearNoSubsampling.__init__\s 8&& HOOD$ ' ' H  t  . . H  \ * * HMMOO    %  !rrr+r,rrch||}|||\}}|||fSr.rCr/s rr1z!LegacyLinearNoSubsampling.forwardjrDrr2r3rs@rrsrsRs "S " "5 " % " " " " " "$,- "" <" "c5<'( " u|U\5<7 8 """"""""rrs)r4typingrrrr'r6rr!r8rFrUrdrmrsrrrxs $#  < < < <