Skip to content

Embeddings

FixedEmbedding

Bases: Module

Fixed Embedding class that creates a fixed embedding of a given maximum length and features.

Source code in models/tts/styledtts2/diffusion/embeddings.py 
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
class FixedEmbedding(nn.Module):
    r"""Fixed Embedding class that creates a fixed embedding of a given maximum length and features."""

    def __init__(self, max_length: int, features: int):
        r"""Initialize the FixedEmbedding with a maximum length and features.

        Args:
            max_length (int): The maximum length of the embedding.
            features (int): The number of features of the embedding.
        """
        super().__init__()
        self.max_length = max_length
        self.embedding = nn.Embedding(max_length, features)

    def forward(self, x: Tensor) -> Tensor:
        r"""Forward pass of the FixedEmbedding.

        Args:
            x (Tensor): The input tensor.

        Returns:
            Tensor: The output tensor.
        """
        batch_size, length, device = *x.shape[0:2], x.device

        assert length <= self.max_length, "Input sequence length must be <= max_length"

        position = torch.arange(length, device=device)

        fixed_embedding = self.embedding(position)
        fixed_embedding = repeat(fixed_embedding, "n d -> b n d", b=batch_size)

        return fixed_embedding

__init__(max_length, features)

Initialize the FixedEmbedding with a maximum length and features.

Parameters:

Name Type Description Default
max_length int

The maximum length of the embedding.

 required
features int

The number of features of the embedding.

 required
Source code in models/tts/styledtts2/diffusion/embeddings.py 
90
91
92
93
94
95
96
97
98
99
def __init__(self, max_length: int, features: int):
    r"""Initialize the FixedEmbedding with a maximum length and features.

    Args:
        max_length (int): The maximum length of the embedding.
        features (int): The number of features of the embedding.
    """
    super().__init__()
    self.max_length = max_length
    self.embedding = nn.Embedding(max_length, features)

forward(x)

Forward pass of the FixedEmbedding.

Parameters:

Name Type Description Default
x Tensor

The input tensor.

 required

Returns:

Name Type Description
Tensor Tensor

The output tensor.
Source code in models/tts/styledtts2/diffusion/embeddings.py 
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
def forward(self, x: Tensor) -> Tensor:
    r"""Forward pass of the FixedEmbedding.

    Args:
        x (Tensor): The input tensor.

    Returns:
        Tensor: The output tensor.
    """
    batch_size, length, device = *x.shape[0:2], x.device

    assert length <= self.max_length, "Input sequence length must be <= max_length"

    position = torch.arange(length, device=device)

    fixed_embedding = self.embedding(position)
    fixed_embedding = repeat(fixed_embedding, "n d -> b n d", b=batch_size)

    return fixed_embedding

LearnedPositionalEmbedding

Bases: Module

Learned Positional Embedding class that creates a learned positional embedding of a given dimension. Used for continuous time.

Source code in models/tts/styledtts2/diffusion/embeddings.py 
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
class LearnedPositionalEmbedding(nn.Module):
    r"""Learned Positional Embedding class that creates a learned positional embedding of a given dimension. Used for continuous time."""

    def __init__(self, dim: int):
        r"""Initialize the LearnedPositionalEmbedding with a dimension.

        Args:
            dim (int): The dimension of the embedding.
        """
        super().__init__()
        assert (dim % 2) == 0
        half_dim = dim // 2
        self.weights = nn.Parameter(torch.randn(half_dim))

    def forward(self, x: Tensor) -> Tensor:
        r"""Forward pass of the LearnedPositionalEmbedding.

        Args:
            x (Tensor): The input tensor.

        Returns:
            Tensor: The output tensor.
        """
        x = rearrange(x, "b -> b 1")

        freqs = x * rearrange(self.weights, "d -> 1 d") * 2 * pi

        fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1)
        fouriered = torch.cat((x, fouriered), dim=-1)

        return fouriered

__init__(dim)

Initialize the LearnedPositionalEmbedding with a dimension.

Parameters:

Name Type Description Default
dim int

The dimension of the embedding.

 required
Source code in models/tts/styledtts2/diffusion/embeddings.py 
41
42
43
44
45
46
47
48
49
50
def __init__(self, dim: int):
    r"""Initialize the LearnedPositionalEmbedding with a dimension.

    Args:
        dim (int): The dimension of the embedding.
    """
    super().__init__()
    assert (dim % 2) == 0
    half_dim = dim // 2
    self.weights = nn.Parameter(torch.randn(half_dim))

forward(x)

Forward pass of the LearnedPositionalEmbedding.

Parameters:

Name Type Description Default
x Tensor

The input tensor.

 required

Returns:

Name Type Description
Tensor Tensor

The output tensor.
Source code in models/tts/styledtts2/diffusion/embeddings.py 
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
def forward(self, x: Tensor) -> Tensor:
    r"""Forward pass of the LearnedPositionalEmbedding.

    Args:
        x (Tensor): The input tensor.

    Returns:
        Tensor: The output tensor.
    """
    x = rearrange(x, "b -> b 1")

    freqs = x * rearrange(self.weights, "d -> 1 d") * 2 * pi

    fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1)
    fouriered = torch.cat((x, fouriered), dim=-1)

    return fouriered

SinusoidalEmbedding

Bases: Module

Sinusoidal Embedding class that creates a sinusoidal embedding of a given dimension.

Source code in models/tts/styledtts2/diffusion/embeddings.py 
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
class SinusoidalEmbedding(nn.Module):
    r"""Sinusoidal Embedding class that creates a sinusoidal embedding of a given dimension."""

    def __init__(self, dim: int):
        r"""Initialize the SinusoidalEmbedding with a dimension.

        Args:
            dim (int): The dimension of the embedding.
        """
        super().__init__()
        self.dim = dim

    def forward(self, x: Tensor) -> Tensor:
        r"""Forward pass of the SinusoidalEmbedding.

        Args:
            x (Tensor): The input tensor.

        Returns:
            Tensor: The output tensor.
        """
        device, half_dim = x.device, self.dim // 2

        emb = torch.tensor(log(10000) / (half_dim - 1), device=device)
        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
        emb = rearrange(x, "i -> i 1") * rearrange(emb, "j -> 1 j")

        return torch.cat((emb.sin(), emb.cos()), dim=-1)

__init__(dim)

Initialize the SinusoidalEmbedding with a dimension.

Parameters:

Name Type Description Default
dim int

The dimension of the embedding.

 required
Source code in models/tts/styledtts2/diffusion/embeddings.py 
11
12
13
14
15
16
17
18
def __init__(self, dim: int):
    r"""Initialize the SinusoidalEmbedding with a dimension.

    Args:
        dim (int): The dimension of the embedding.
    """
    super().__init__()
    self.dim = dim

forward(x)

Forward pass of the SinusoidalEmbedding.

Parameters:

Name Type Description Default
x Tensor

The input tensor.

 required

Returns:

Name Type Description
Tensor Tensor

The output tensor.
Source code in models/tts/styledtts2/diffusion/embeddings.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
def forward(self, x: Tensor) -> Tensor:
    r"""Forward pass of the SinusoidalEmbedding.

    Args:
        x (Tensor): The input tensor.

    Returns:
        Tensor: The output tensor.
    """
    device, half_dim = x.device, self.dim // 2

    emb = torch.tensor(log(10000) / (half_dim - 1), device=device)
    emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
    emb = rearrange(x, "i -> i 1") * rearrange(emb, "j -> 1 j")

    return torch.cat((emb.sin(), emb.cos()), dim=-1)

TimePositionalEmbedding(dim, out_features)

Creates a time positional embedding of a given dimension and output features.

Parameters:

Name Type Description Default
dim int

The dimension of the embedding.

 required
out_features int

The number of output features.

 required

Returns:

Type Description
Module

nn.Module: The time positional embedding module.
Source code in models/tts/styledtts2/diffusion/embeddings.py 
71
72
73
74
75
76
77
78
79
80
81
82
83
84
def TimePositionalEmbedding(dim: int, out_features: int) -> nn.Module:
    r"""Creates a time positional embedding of a given dimension and output features.

    Args:
        dim (int): The dimension of the embedding.
        out_features (int): The number of output features.

    Returns:
        nn.Module: The time positional embedding module.
    """
    return nn.Sequential(
        LearnedPositionalEmbedding(dim),
        nn.Linear(in_features=dim + 1, out_features=out_features),
    )