pflow/models/components/speech_prompt_encoder.py

""" from https://github.com/jaywalnut310/glow-tts """

import math

import torch
import torch.nn as nn
from einops import rearrange

import pflow.utils as utils
from pflow.utils.model import sequence_mask
from pflow.models.components import commons
from pflow.models.components.vits_posterior import PosteriorEncoder
from pflow.models.components.transformer import BasicTransformerBlock

log = utils.get_pylogger(__name__)

class LayerNorm(nn.Module):
    def __init__(self, channels, eps=1e-4):
        super().__init__()
        self.channels = channels
        self.eps = eps

        self.gamma = torch.nn.Parameter(torch.ones(channels))
        self.beta = torch.nn.Parameter(torch.zeros(channels))

    def forward(self, x):
        n_dims = len(x.shape)
        mean = torch.mean(x, 1, keepdim=True)
        variance = torch.mean((x - mean) ** 2, 1, keepdim=True)

        x = (x - mean) * torch.rsqrt(variance + self.eps)

        shape = [1, -1] + [1] * (n_dims - 2)
        x = x * self.gamma.view(*shape) + self.beta.view(*shape)
        return x


class ConvReluNorm(nn.Module):
    def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
        super().__init__()
        self.in_channels = in_channels
        self.hidden_channels = hidden_channels
        self.out_channels = out_channels
        self.kernel_size = kernel_size
        self.n_layers = n_layers
        self.p_dropout = p_dropout

        self.conv_layers = torch.nn.ModuleList()
        self.norm_layers = torch.nn.ModuleList()
        self.conv_layers.append(torch.nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size // 2))
        self.norm_layers.append(LayerNorm(hidden_channels))
        self.relu_drop = torch.nn.Sequential(torch.nn.ReLU(), torch.nn.Dropout(p_dropout))
        for _ in range(n_layers - 1):
            self.conv_layers.append(
                torch.nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size // 2)
            )
            self.norm_layers.append(LayerNorm(hidden_channels))
        self.proj = torch.nn.Conv1d(hidden_channels, out_channels, 1)
        self.proj.weight.data.zero_()
        self.proj.bias.data.zero_()

    def forward(self, x, x_mask):
        x_org = x
        for i in range(self.n_layers):
            x = self.conv_layers[i](x * x_mask)
            x = self.norm_layers[i](x)
            x = self.relu_drop(x)
        x = x_org + self.proj(x)
        return x * x_mask


class DurationPredictor(nn.Module):
    def __init__(self, in_channels, filter_channels, kernel_size, p_dropout):
        super().__init__()
        self.in_channels = in_channels
        self.filter_channels = filter_channels
        self.p_dropout = p_dropout

        self.drop = torch.nn.Dropout(p_dropout)
        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
        self.norm_1 = LayerNorm(filter_channels)
        self.conv_2 = torch.nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2)
        self.norm_2 = LayerNorm(filter_channels)
        self.proj = torch.nn.Conv1d(filter_channels, 1, 1)

    def forward(self, x, x_mask):
        x = self.conv_1(x * x_mask)
        x = torch.relu(x)
        x = self.norm_1(x)
        x = self.drop(x)
        x = self.conv_2(x * x_mask)
        x = torch.relu(x)
        x = self.norm_2(x)
        x = self.drop(x)
        x = self.proj(x * x_mask)
        # x = torch.relu(x)
        return x * x_mask
    
class DurationPredictorNS2(nn.Module):
    def __init__(
        self, in_channels, filter_channels, kernel_size, p_dropout=0.5
    ):
        super().__init__()

        self.in_channels = in_channels
        self.filter_channels = filter_channels
        self.kernel_size = kernel_size
        self.p_dropout = p_dropout

        self.drop = nn.Dropout(p_dropout)
        self.conv_1 = nn.Conv1d(
            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
        )
        self.norm_1 = LayerNorm(filter_channels)
        
        self.module_list = nn.ModuleList()
        self.module_list.append(self.conv_1)
        self.module_list.append(nn.ReLU())
        self.module_list.append(self.norm_1)
        self.module_list.append(self.drop)
        
        for i in range(12):
            self.module_list.append(nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2))
            self.module_list.append(nn.ReLU())
            self.module_list.append(LayerNorm(filter_channels))
            self.module_list.append(nn.Dropout(p_dropout))
            
        
        # attention layer every 3 layers
        self.attn_list = nn.ModuleList()
        for i in range(4):
            self.attn_list.append(
                Encoder(
                    filter_channels,
                    filter_channels,
                    8,
                    10,
                    3,
                    p_dropout=p_dropout,
                )
            )

        for i in range(30):
            if i+1 % 3 == 0:
                self.module_list.append(self.attn_list[i//3])
        
        self.proj = nn.Conv1d(filter_channels, 1, 1)

    def forward(self, x, x_mask):
        x = torch.detach(x)
        for layer in self.module_list:
            x = layer(x * x_mask)
        x = self.proj(x * x_mask)
        # x = torch.relu(x)
        return x * x_mask
    
class RotaryPositionalEmbeddings(nn.Module):
    """
    ## RoPE module

    Rotary encoding transforms pairs of features by rotating in the 2D plane.
    That is, it organizes the $d$ features as $\frac{d}{2}$ pairs.
    Each pair can be considered a coordinate in a 2D plane, and the encoding will rotate it
    by an angle depending on the position of the token.
    """

    def __init__(self, d: int, base: int = 10_000):
        r"""
        * `d` is the number of features $d$
        * `base` is the constant used for calculating $\Theta$
        """
        super().__init__()

        self.base = base
        self.d = int(d)
        self.cos_cached = None
        self.sin_cached = None

    def _build_cache(self, x: torch.Tensor):
        r"""
        Cache $\cos$ and $\sin$ values
        """
        # Return if cache is already built
        if self.cos_cached is not None and x.shape[0] <= self.cos_cached.shape[0]:
            return

        # Get sequence length
        seq_len = x.shape[0]

        # $\Theta = {\theta_i = 10000^{-\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
        theta = 1.0 / (self.base ** (torch.arange(0, self.d, 2).float() / self.d)).to(x.device)

        # Create position indexes `[0, 1, ..., seq_len - 1]`
        seq_idx = torch.arange(seq_len, device=x.device).float().to(x.device)

        # Calculate the product of position index and $\theta_i$
        idx_theta = torch.einsum("n,d->nd", seq_idx, theta)

        # Concatenate so that for row $m$ we have
        # $[m \theta_0, m \theta_1, ..., m \theta_{\frac{d}{2}}, m \theta_0, m \theta_1, ..., m \theta_{\frac{d}{2}}]$
        idx_theta2 = torch.cat([idx_theta, idx_theta], dim=1)

        # Cache them
        self.cos_cached = idx_theta2.cos()[:, None, None, :]
        self.sin_cached = idx_theta2.sin()[:, None, None, :]

    def _neg_half(self, x: torch.Tensor):
        # $\frac{d}{2}$
        d_2 = self.d // 2

        # Calculate $[-x^{(\frac{d}{2} + 1)}, -x^{(\frac{d}{2} + 2)}, ..., -x^{(d)}, x^{(1)}, x^{(2)}, ..., x^{(\frac{d}{2})}]$
        return torch.cat([-x[:, :, :, d_2:], x[:, :, :, :d_2]], dim=-1)

    def forward(self, x: torch.Tensor):
        """
        * `x` is the Tensor at the head of a key or a query with shape `[seq_len, batch_size, n_heads, d]`
        """
        # Cache $\cos$ and $\sin$ values
        x = rearrange(x, "b h t d -> t b h d")

        self._build_cache(x)

        # Split the features, we can choose to apply rotary embeddings only to a partial set of features.
        x_rope, x_pass = x[..., : self.d], x[..., self.d :]

        # Calculate
        # $[-x^{(\frac{d}{2} + 1)}, -x^{(\frac{d}{2} + 2)}, ..., -x^{(d)}, x^{(1)}, x^{(2)}, ..., x^{(\frac{d}{2})}]$
        neg_half_x = self._neg_half(x_rope)

        x_rope = (x_rope * self.cos_cached[: x.shape[0]]) + (neg_half_x * self.sin_cached[: x.shape[0]])

        return rearrange(torch.cat((x_rope, x_pass), dim=-1), "t b h d -> b h t d")


class MultiHeadAttention(nn.Module):
    def __init__(
        self,
        channels,
        out_channels,
        n_heads,
        heads_share=True,
        p_dropout=0.0,
        proximal_bias=False,
        proximal_init=False,
    ):
        super().__init__()
        assert channels % n_heads == 0

        self.channels = channels
        self.out_channels = out_channels
        self.n_heads = n_heads
        self.heads_share = heads_share
        self.proximal_bias = proximal_bias
        self.p_dropout = p_dropout
        self.attn = None

        self.k_channels = channels // n_heads
        self.conv_q = torch.nn.Conv1d(channels, channels, 1)
        self.conv_k = torch.nn.Conv1d(channels, channels, 1)
        self.conv_v = torch.nn.Conv1d(channels, channels, 1)

        # from https://nn.labml.ai/transformers/rope/index.html
        self.query_rotary_pe = RotaryPositionalEmbeddings(self.k_channels * 0.5)
        self.key_rotary_pe = RotaryPositionalEmbeddings(self.k_channels * 0.5)

        self.conv_o = torch.nn.Conv1d(channels, out_channels, 1)
        self.drop = torch.nn.Dropout(p_dropout)

        torch.nn.init.xavier_uniform_(self.conv_q.weight)
        torch.nn.init.xavier_uniform_(self.conv_k.weight)
        if proximal_init:
            self.conv_k.weight.data.copy_(self.conv_q.weight.data)
            self.conv_k.bias.data.copy_(self.conv_q.bias.data)
        torch.nn.init.xavier_uniform_(self.conv_v.weight)

    def forward(self, x, c, attn_mask=None):
        q = self.conv_q(x)
        k = self.conv_k(c)
        v = self.conv_v(c)

        x, self.attn = self.attention(q, k, v, mask=attn_mask)

        x = self.conv_o(x)
        return x

    def attention(self, query, key, value, mask=None):
        b, d, t_s, t_t = (*key.size(), query.size(2))
        query = rearrange(query, "b (h c) t-> b h t c", h=self.n_heads)
        key = rearrange(key, "b (h c) t-> b h t c", h=self.n_heads)
        value = rearrange(value, "b (h c) t-> b h t c", h=self.n_heads)

        query = self.query_rotary_pe(query)
        key = self.key_rotary_pe(key)

        scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(self.k_channels)

        if self.proximal_bias:
            assert t_s == t_t, "Proximal bias is only available for self-attention."
            scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
        if mask is not None:
            scores = scores.masked_fill(mask == 0, -1e4)
        p_attn = torch.nn.functional.softmax(scores, dim=-1)
        p_attn = self.drop(p_attn)
        output = torch.matmul(p_attn, value)
        output = output.transpose(2, 3).contiguous().view(b, d, t_t)
        return output, p_attn

    @staticmethod
    def _attention_bias_proximal(length):
        r = torch.arange(length, dtype=torch.float32)
        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)


class FFN(nn.Module):
    def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0):
        super().__init__()
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.filter_channels = filter_channels
        self.kernel_size = kernel_size
        self.p_dropout = p_dropout

        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
        self.conv_2 = torch.nn.Conv1d(filter_channels, out_channels, kernel_size, padding=kernel_size // 2)
        self.drop = torch.nn.Dropout(p_dropout)

    def forward(self, x, x_mask):
        x = self.conv_1(x * x_mask)
        x = torch.relu(x)
        x = self.drop(x)
        x = self.conv_2(x * x_mask)
        return x * x_mask


class Encoder(nn.Module):
    def __init__(
        self,
        hidden_channels,
        filter_channels,
        n_heads,
        n_layers,
        kernel_size=1,
        p_dropout=0.0,
        **kwargs,
    ):
        super().__init__()
        self.hidden_channels = hidden_channels
        self.filter_channels = filter_channels
        self.n_heads = n_heads
        self.n_layers = n_layers
        self.kernel_size = kernel_size
        self.p_dropout = p_dropout

        self.drop = torch.nn.Dropout(p_dropout)
        self.attn_layers = torch.nn.ModuleList()
        self.norm_layers_1 = torch.nn.ModuleList()
        self.ffn_layers = torch.nn.ModuleList()
        self.norm_layers_2 = torch.nn.ModuleList()
        for _ in range(self.n_layers):
            self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
            self.norm_layers_1.append(LayerNorm(hidden_channels))
            self.ffn_layers.append(
                FFN(
                    hidden_channels,
                    hidden_channels,
                    filter_channels,
                    kernel_size,
                    p_dropout=p_dropout,
                )
            )
            self.norm_layers_2.append(LayerNorm(hidden_channels))

    def forward(self, x, x_mask):
        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
        for i in range(self.n_layers):
            x = x * x_mask
            y = self.attn_layers[i](x, x, attn_mask)
            y = self.drop(y)
            x = self.norm_layers_1[i](x + y)
            y = self.ffn_layers[i](x, x_mask)
            y = self.drop(y)
            x = self.norm_layers_2[i](x + y)
        x = x * x_mask
        return x

class Decoder(nn.Module):
    def __init__(
        self,
        hidden_channels,
        filter_channels,
        n_heads,
        n_layers,
        kernel_size=1,
        p_dropout=0.0,
        proximal_bias=False,
        proximal_init=True,
        **kwargs
    ):
        super().__init__()
        self.hidden_channels = hidden_channels
        self.filter_channels = filter_channels
        self.n_heads = n_heads
        self.n_layers = n_layers
        self.kernel_size = kernel_size
        self.p_dropout = p_dropout
        self.proximal_bias = proximal_bias
        self.proximal_init = proximal_init

        self.drop = nn.Dropout(p_dropout)
        self.self_attn_layers = nn.ModuleList()
        self.norm_layers_0 = nn.ModuleList()
        self.encdec_attn_layers = nn.ModuleList()
        self.norm_layers_1 = nn.ModuleList()
        self.ffn_layers = nn.ModuleList()
        self.norm_layers_2 = nn.ModuleList()
        for i in range(self.n_layers):
            self.self_attn_layers.append(
                MultiHeadAttention(
                    hidden_channels,
                    hidden_channels, 
                    n_heads, 
                    p_dropout=p_dropout
                    )
                )
            self.norm_layers_0.append(LayerNorm(hidden_channels))
            self.encdec_attn_layers.append(
                MultiHeadAttention(
                    hidden_channels,
                    hidden_channels, 
                    n_heads, 
                    p_dropout=p_dropout
                    )
                )
            self.norm_layers_1.append(LayerNorm(hidden_channels))
            self.ffn_layers.append(
                FFN(
                    hidden_channels,
                    hidden_channels,
                    filter_channels,
                    kernel_size,
                    p_dropout=p_dropout,
                )
            )
            self.norm_layers_2.append(LayerNorm(hidden_channels))

    def forward(self, x, x_mask, h, h_mask):
        """
        x: decoder input
        h: encoder output
        """
        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
            device=x.device, dtype=x.dtype
        )
        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
        x = x * x_mask
        for i in range(self.n_layers):
            y = self.self_attn_layers[i](x, x, self_attn_mask)
            y = self.drop(y)
            x = self.norm_layers_0[i](x + y)

            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
            y = self.drop(y)
            x = self.norm_layers_1[i](x + y)

            y = self.ffn_layers[i](x, x_mask)
            y = self.drop(y)
            x = self.norm_layers_2[i](x + y)
        x = x * x_mask
        return x
    
class TextEncoder(nn.Module):
    def __init__(
        self,
        encoder_type,
        encoder_params,
        duration_predictor_params,
        n_vocab,
        speech_in_channels,
    ):
        super().__init__()
        self.encoder_type = encoder_type
        self.n_vocab = n_vocab
        self.n_feats = encoder_params.n_feats
        self.n_channels = encoder_params.n_channels

        self.emb = torch.nn.Embedding(n_vocab, self.n_channels)
        torch.nn.init.normal_(self.emb.weight, 0.0, self.n_channels**-0.5)

        self.speech_in_channels = speech_in_channels
        self.speech_out_channels = self.n_channels
        self.speech_prompt_proj = torch.nn.Conv1d(self.speech_in_channels, self.speech_out_channels, 1)
        # self.speech_prompt_proj = PosteriorEncoder(
        #     self.speech_in_channels,
        #     self.speech_out_channels,
        #     self.speech_out_channels,
        #     1,
        #     1,
        #     1,
        #     gin_channels=0,
        # )

        self.prenet = ConvReluNorm(
            self.n_channels,
            self.n_channels,
            self.n_channels,
            kernel_size=5,
            n_layers=3,
            p_dropout=0,
        )

        self.speech_prompt_encoder = Encoder(
            encoder_params.n_channels,
            encoder_params.filter_channels,
            encoder_params.n_heads,
            encoder_params.n_layers,
            encoder_params.kernel_size,
            encoder_params.p_dropout,
        )

        self.text_base_encoder = Encoder(
            encoder_params.n_channels,
            encoder_params.filter_channels,
            encoder_params.n_heads,
            encoder_params.n_layers,
            encoder_params.kernel_size,
            encoder_params.p_dropout,
        )

        self.decoder = Decoder(
            encoder_params.n_channels,
            encoder_params.filter_channels,
            encoder_params.n_heads,
            encoder_params.n_layers,
            encoder_params.kernel_size,
            encoder_params.p_dropout,
        )

        self.transformerblock = BasicTransformerBlock(
            encoder_params.n_channels,
            encoder_params.n_heads,
            encoder_params.n_channels // encoder_params.n_heads,
            encoder_params.p_dropout,
            encoder_params.n_channels,
            activation_fn="gelu",
            attention_bias=False,
            only_cross_attention=False,
            double_self_attention=False,
            upcast_attention=False,
            norm_elementwise_affine=True,
            norm_type="layer_norm",
            final_dropout=False,
        )
        self.proj_m = torch.nn.Conv1d(self.n_channels, self.n_feats, 1)

        self.proj_w = DurationPredictor(
            self.n_channels,
            duration_predictor_params.filter_channels_dp,
            duration_predictor_params.kernel_size,
            duration_predictor_params.p_dropout,
        )
        # self.proj_w = DurationPredictorNS2(
        #     self.n_channels,
        #     duration_predictor_params.filter_channels_dp,
        #     duration_predictor_params.kernel_size,
        #     duration_predictor_params.p_dropout,
        # )

        # self.speech_prompt_pos_emb = RotaryPositionalEmbeddings(self.n_channels * 0.5)
        # self.text_pos_emb = RotaryPositionalEmbeddings(self.n_channels * 0.5)
    
    def forward(
            self, 
            x_input, 
            x_lengths, 
            speech_prompt,
            ):
        """Run forward pass to the transformer based encoder and duration predictor

        Args:
            x (torch.Tensor): text input
                shape: (batch_size, max_text_length)
            x_lengths (torch.Tensor): text input lengths
                shape: (batch_size,)
            speech_prompt (torch.Tensor): speech prompt input

        Returns:
            mu (torch.Tensor): average output of the encoder
                shape: (batch_size, n_feats, max_text_length)
            logw (torch.Tensor): log duration predicted by the duration predictor
                shape: (batch_size, 1, max_text_length)
            x_mask (torch.Tensor): mask for the text input
                shape: (batch_size, 1, max_text_length)
        """
 
        x_emb = self.emb(x_input) * math.sqrt(self.n_channels)
        x_emb = torch.transpose(x_emb, 1, -1)
        x_emb_mask = torch.unsqueeze(sequence_mask(x_lengths, x_emb.size(2)), 1).to(x_emb.dtype)
        x_emb = self.text_base_encoder(x_emb, x_emb_mask)

        x_speech_lengths = x_lengths + speech_prompt.size(2)
        speech_lengths = x_speech_lengths - x_lengths
        speech_mask = torch.unsqueeze(sequence_mask(speech_lengths, speech_prompt.size(2)), 1).to(x_emb.dtype)
         
        speech_prompt_proj = self.speech_prompt_proj(speech_prompt)
        # speech_prompt_proj, speech_mask = self.speech_prompt_proj(speech_prompt, speech_lengths)
        # speech_prompt_proj = self.speech_prompt_encoder(speech_prompt_proj, speech_mask)

        x_speech_cat = torch.cat([speech_prompt_proj, x_emb], dim=2)
        x_speech_mask = torch.unsqueeze(sequence_mask(x_speech_lengths, x_speech_cat.size(2)), 1).to(x_speech_cat.dtype)      
        
        x_prenet = self.prenet(x_speech_cat, x_speech_mask)
        # split speech prompt and text input
        speech_prompt_proj = x_prenet[:, :, :speech_prompt_proj.size(2)]
        x_split = x_prenet[:, :, speech_prompt_proj.size(2):]
        
        # add positional encoding to speech prompt and x_split
        # x_split = self.text_pos_emb(x_split.unsqueeze(1).transpose(-2,-1)).squeeze(1).transpose(-2,-1)
        x_split_mask = torch.unsqueeze(sequence_mask(x_lengths, x_split.size(2)), 1).to(x_split.dtype)      
               
        # speech_prompt = self.speech_prompt_pos_emb(speech_prompt_proj.unsqueeze(1).transpose(-2,-1)).squeeze(1).transpose(-2,-1)
        # x_split = self.decoder(x_split, x_split_mask, speech_prompt, speech_mask)

        x_split = self.transformerblock(x_split.transpose(1,2), x_split_mask, speech_prompt_proj.transpose(1,2), speech_mask)
        x_split = x_split.transpose(1,2)
        
        # x_split_mask = torch.unsqueeze(sequence_mask(x_lengths, x_split.size(2)), 1).to(x.dtype)
        
        # x_split = x_split + x_emb

        mu = self.proj_m(x_split) * x_split_mask

        x_dp = torch.detach(x_split)
        logw = self.proj_w(x_dp, x_split_mask)

        return mu, logw, x_split_mask