o ji @sddlZddlmZddlmZmZmZddlmZddlm Z ddl m Z Gdddej Z Gd d d ej ZGd d d ej ZGd ddej ZGdddej ZdS)N)nn)Conv1dBN Conv1dBNBlockResidualConv1dBNBlock)FFTransformerBlock)WNBlocks)RelativePositionTransformerc*eZdZdZfddZdddZZS)WaveNetDecoderaWaveNet based decoder with a prenet and a postnet. prenet: conv1d_1x1 postnet: 3 x [conv1d_1x1 -> relu] -> conv1d_1x1 TODO: Integrate speaker conditioning vector. Note: default wavenet parameters; params = { "num_blocks": 12, "hidden_channels":192, "kernel_size": 5, "dilation_rate": 1, "num_layers": 4, "dropout_p": 0.05 } Args: in_channels (int): number of input channels. out_channels (int): number of output channels. hidden_channels (int): number of hidden channels for prenet and postnet. params (dict): dictionary for residual convolutional blocks. c sttj||dd|_t|dfd|i||_tj|d|dtjtj||dtjtj||dtjtj||dg|_ tj |j |_ dS)Nhidden_channels c_in_channels) super__init__torchrConv1dprenetrwnReLUpostnet Sequential)self in_channels out_channelsr r params __class__V/home/kuhnn/.local/lib/python3.10/site-packages/TTS/tts/layers/feed_forward/decoder.pyr$s  zWaveNetDecoder.__init__NcCs.|||}||||}|||}|SN)rrrrxx_maskgorrrforward6szWaveNetDecoder.forwardNN__name__ __module__ __qualname____doc__rr% __classcell__rrrrr s r cr )"RelativePositionTransformerDecoderaDecoder with Relative Positional Transformer. Note: Default params params={ 'hidden_channels_ffn': 128, 'num_heads': 2, "kernel_size": 3, "dropout_p": 0.1, "num_layers": 8, "rel_attn_window_size": 4, "input_length": None } Args: in_channels (int): number of input channels. out_channels (int): number of output channels. hidden_channels (int): number of hidden channels including Transformer layers. params (dict): dictionary for residual convolutional blocks. cs4tt||dd|_t|||fi||_dSNr )rrrrrrel_pos_transformerrrrr rrrrrSs z+RelativePositionTransformerDecoder.__init__NcCs|||}|||}|Sr)rr/r rrrr%Xs z*RelativePositionTransformerDecoder.forwardr&r'rrrrr-=s r-cr )FFTransformerDecoderaDecoder with FeedForwardTransformer. Default params params={ 'hidden_channels_ffn': 1024, 'num_heads': 2, "dropout_p": 0.1, "num_layers": 6, } Args: in_channels (int): number of input channels. out_channels (int): number of output channels. hidden_channels (int): number of hidden channels including Transformer layers. params (dict): dictionary for residual convolutional blocks. cs0tt|fi||_t||d|_dSr.)rrrtransformer_blockrrr)rrrrrrrrps zFFTransformerDecoder.__init__NcCs0|durdn|}|||}|||}|Sr.)r2rr rrrr%uszFFTransformerDecoder.forwardr&r'rrrrr1^s r1cr )ResidualConv1dBNDecoderaResidual Convolutional Decoder as in the original Speedy Speech paper TODO: Integrate speaker conditioning vector. Note: Default params params = { "kernel_size": 4, "dilations": 4 * [1, 2, 4, 8] + [1], "num_conv_blocks": 2, "num_res_blocks": 17 } Args: in_channels (int): number of input channels. out_channels (int): number of output channels. hidden_channels (int): number of hidden channels including ResidualConv1dBNBlock layers. params (dict): dictionary for residual convolutional blocks. c s`tt|||fi||_t||d|_tt||||ddddt||d|_ dS)Nr kernel_size)num_conv_blocks) rrrres_conv_blockrr post_convrrrr0rrrrs   z ResidualConv1dBNDecoder.__init__NcCs(|||}|||}|||Sr)r7r8rr rrrr%s zResidualConv1dBNDecoder.forwardr&r'rrrrr3}s  r3csJeZdZdZdddgddgdddd ffd d Zdd dZZS)Decodera$Decodes the expanded phoneme encoding into spectrograms Args: out_channels (int): number of output channels. in_hidden_channels (int): input and hidden channels. Model keeps the input channels for the intermediate layers. decoder_type (str): decoder layer types. 'transformers' or 'residual_conv_bn'. Default 'residual_conv_bn'. decoder_params (dict): model parameters for specified decoder type. c_in_channels (int): number of channels for conditional input. Shapes: - input: (B, C, T) residual_conv_bn)r r5r;r r5)r4 dilationsr6num_res_blocksrcst|dkrt||||d|_dS|dkr't||||d|_dS|dkr9t|||||d|_dS|dkrHt||||_dStd|)Nrelative_position_transformer)rrr rr:wavenet)rrr r r fftransformerz[!] Unknown decoder type - ) rrlowerr-decoderr3r r1 ValueError)rrin_hidden_channels decoder_typedecoder_paramsr rrrrs4    zDecoder.__init__NcCs||||}|S)zi Args: x: [B, C, T] x_mask: [B, 1, T] g: [B, C_g, 1] )rDr rrrr%szDecoder.forwardrr'rrrrr9s*r9)rr"TTS.tts.layers.generic.res_conv_bnrrr"TTS.tts.layers.generic.transformerrTTS.tts.layers.generic.wavenetr#TTS.tts.layers.glow_tts.transformerrModuler r-r1r3r9rrrrs    3!&