rvc/lib/algorithm/attentions.py

import math
import torch

from rvc.lib.algorithm.commons import convert_pad_shape


class MultiHeadAttention(torch.nn.Module):
    """
    Multi-head attention module with optional relative positional encoding and proximal bias.

    Args:
        channels (int): Number of input channels.
        out_channels (int): Number of output channels.
        n_heads (int): Number of attention heads.
        p_dropout (float, optional): Dropout probability. Defaults to 0.0.
        window_size (int, optional): Window size for relative positional encoding. Defaults to None.
        heads_share (bool, optional): Whether to share relative positional embeddings across heads. Defaults to True.
        block_length (int, optional): Block length for local attention. Defaults to None.
        proximal_bias (bool, optional): Whether to use proximal bias in self-attention. Defaults to False.
        proximal_init (bool, optional): Whether to initialize the key projection weights the same as query projection weights. Defaults to False.
    """

    def __init__(
        self,
        channels,
        out_channels,
        n_heads,
        p_dropout=0.0,
        window_size=None,
        heads_share=True,
        block_length=None,
        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.p_dropout = p_dropout
        self.window_size = window_size
        self.heads_share = heads_share
        self.block_length = block_length
        self.proximal_bias = proximal_bias
        self.proximal_init = proximal_init
        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)
        self.conv_o = torch.nn.Conv1d(channels, out_channels, 1)
        self.drop = torch.nn.Dropout(p_dropout)

        if window_size is not None:
            n_heads_rel = 1 if heads_share else n_heads
            rel_stddev = self.k_channels**-0.5
            self.emb_rel_k = torch.nn.Parameter(
                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
                * rel_stddev
            )
            self.emb_rel_v = torch.nn.Parameter(
                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
                * rel_stddev
            )

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

    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):
        # reshape [b, d, t] -> [b, n_h, t, d_k]
        b, d, t_s, t_t = (*key.size(), query.size(2))
        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)

        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
        if self.window_size is not None:
            assert (
                t_s == t_t
            ), "Relative attention is only available for self-attention."
            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
            rel_logits = self._matmul_with_relative_keys(
                query / math.sqrt(self.k_channels), key_relative_embeddings
            )
            scores_local = self._relative_position_to_absolute_position(rel_logits)
            scores = scores + scores_local
        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)
            if self.block_length is not None:
                assert (
                    t_s == t_t
                ), "Local attention is only available for self-attention."
                block_mask = (
                    torch.ones_like(scores)
                    .triu(-self.block_length)
                    .tril(self.block_length)
                )
                scores = scores.masked_fill(block_mask == 0, -1e4)
        p_attn = torch.nn.functional.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
        p_attn = self.drop(p_attn)
        output = torch.matmul(p_attn, value)
        if self.window_size is not None:
            relative_weights = self._absolute_position_to_relative_position(p_attn)
            value_relative_embeddings = self._get_relative_embeddings(
                self.emb_rel_v, t_s
            )
            output = output + self._matmul_with_relative_values(
                relative_weights, value_relative_embeddings
            )
        output = (
            output.transpose(2, 3).contiguous().view(b, d, t_t)
        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
        return output, p_attn

    def _matmul_with_relative_values(self, x, y):
        """
        x: [b, h, l, m]
        y: [h or 1, m, d]
        ret: [b, h, l, d]
        """
        ret = torch.matmul(x, y.unsqueeze(0))
        return ret

    def _matmul_with_relative_keys(self, x, y):
        """
        x: [b, h, l, d]
        y: [h or 1, m, d]
        ret: [b, h, l, m]
        """
        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
        return ret

    def _get_relative_embeddings(self, relative_embeddings, length):
        # Pad first before slice to avoid using cond ops.
        pad_length = max(length - (self.window_size + 1), 0)
        slice_start_position = max((self.window_size + 1) - length, 0)
        slice_end_position = slice_start_position + 2 * length - 1
        if pad_length > 0:
            padded_relative_embeddings = torch.nn.functional.pad(
                relative_embeddings,
                convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
            )
        else:
            padded_relative_embeddings = relative_embeddings
        used_relative_embeddings = padded_relative_embeddings[
            :, slice_start_position:slice_end_position
        ]
        return used_relative_embeddings

    def _relative_position_to_absolute_position(self, x):
        """
        x: [b, h, l, 2*l-1]
        ret: [b, h, l, l]
        """
        batch, heads, length, _ = x.size()

        # Concat columns of pad to shift from relative to absolute indexing.
        x = torch.nn.functional.pad(
            x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])
        )

        # Concat extra elements so to add up to shape (len+1, 2*len-1).
        x_flat = x.view([batch, heads, length * 2 * length])
        x_flat = torch.nn.functional.pad(
            x_flat, convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])
        )

        # Reshape and slice out the padded elements.
        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
            :, :, :length, length - 1 :
        ]
        return x_final

    def _absolute_position_to_relative_position(self, x):
        """
        x: [b, h, l, l]
        ret: [b, h, l, 2*l-1]
        """
        batch, heads, length, _ = x.size()
        # padd along column
        x = torch.nn.functional.pad(
            x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
        )
        x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
        # add 0's in the beginning that will skew the elements after reshape
        x_flat = torch.nn.functional.pad(
            x_flat, convert_pad_shape([[0, 0], [0, 0], [length, 0]])
        )
        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
        return x_final

    def _attention_bias_proximal(self, length):
        """Bias for self-attention to encourage attention to close positions.
        Args:
            length: an integer scalar.
        """
        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(torch.nn.Module):
    """
    Feed-forward network module.

    Args:
        in_channels (int): Number of input channels.
        out_channels (int): Number of output channels.
        filter_channels (int): Number of filter channels in the convolution layers.
        kernel_size (int): Kernel size of the convolution layers.
        p_dropout (float, optional): Dropout probability. Defaults to 0.0.
        activation (str, optional): Activation function to use. Defaults to None.
        causal (bool, optional): Whether to use causal padding in the convolution layers. Defaults to False.
    """

    def __init__(
        self,
        in_channels,
        out_channels,
        filter_channels,
        kernel_size,
        p_dropout=0.0,
        activation=None,
        causal=False,
    ):
        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.activation = activation
        self.causal = causal

        if causal:
            self.padding = self._causal_padding
        else:
            self.padding = self._same_padding

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

    def forward(self, x, x_mask):
        x = self.conv_1(self.padding(x * x_mask))
        if self.activation == "gelu":
            x = x * torch.sigmoid(1.702 * x)
        else:
            x = torch.relu(x)
        x = self.drop(x)
        x = self.conv_2(self.padding(x * x_mask))
        return x * x_mask

    def _causal_padding(self, x):
        if self.kernel_size == 1:
            return x
        pad_l = self.kernel_size - 1
        pad_r = 0
        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
        x = torch.nn.functional.pad(x, convert_pad_shape(padding))
        return x

    def _same_padding(self, x):
        if self.kernel_size == 1:
            return x
        pad_l = (self.kernel_size - 1) // 2
        pad_r = self.kernel_size // 2
        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
        x = torch.nn.functional.pad(x, convert_pad_shape(padding))
        return x