Skip to content

var_pred

Conv1dVarPredictor

Bases: Module

The Conv1d variance predictor for FastSpeech2. This module is made up of: 1. (mandatory) The Conv1d part contains two or more Conv1d blocks which are composed of the components below 1. (mandatory) a Conv1d layer 2. (mandatory) a ReLU function 3. (mandatory) a LayerNorm layer 4. (mandatory) a Dropout layer. 2. (mandatory) The Linear part contains one Linear block which is composed of the component below 1. (mandatory) a Linear layer

Reference

Fastspeech 2: Fast and high-quality end-to-end text to speech https://arxiv.org/pdf/2006.04558

Source code in speechain/module/prenet/var_pred.py 
 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
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 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
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
class Conv1dVarPredictor(Module):
    """
    The Conv1d variance predictor for FastSpeech2.
    This module is made up of:
        1. (mandatory) The Conv1d part contains two or more Conv1d blocks which are composed of the components below
            1. (mandatory) a Conv1d layer
            2. (mandatory) a ReLU function
            3. (mandatory) a LayerNorm layer
            4. (mandatory) a Dropout layer.
        2. (mandatory) The Linear part contains one Linear block which is composed of the component below
            1. (mandatory) a Linear layer

    Reference:
        Fastspeech 2: Fast and high-quality end-to-end text to speech
        https://arxiv.org/pdf/2006.04558
    """

    def module_init(
        self,
        feat_dim: int = None,
        conv_dims: int or List[int] = [256, 256],
        conv_kernel: int = 3,
        conv_stride: int = 1,
        use_gate: bool = False,
        conv_dropout: float or List[float] = 0.5,
        use_conv_emb: bool = True,
        conv_emb_kernel: int = 1,
        conv_emb_dropout: float = 0.0,
    ):
        """

        Args:
            feat_dim: int
                The dimension of input acoustic feature tensors.
                Used for calculating the in_features of the first Linear layer.
            conv_dims: List[int] or int
                The values of out_channels of each Conv1d layer.
                If a list of integers is given, multiple Conv1d layers will be initialized.
                If an integer is given, there will be only one Conv1d layer
            conv_kernel: int
                The value of kernel_size of all Conv1d layers.
            conv_stride: int
                The value of stride of all Conv1d layers.
            conv_dropout: float or List[float]
                The values of p rate of the Dropout layer after each Linear layer.
            use_conv_emb: bool
                Whether to embed the predicted scalar back to an embedding vector.
                This argument needs to be False for duration predictor.
            conv_emb_kernel: int
                The value of kernel_size for the conv1d embedding layer.
                Only effective when use_conv_emb is True.
            conv_emb_dropout: float
                The value of p reate of the Dropout layer after the conv1d embedding layer.
                Only effective when use_conv_emb is True.
        """
        # --- 0. Argument Checking --- #
        # Convolution arguments checking
        assert isinstance(
            conv_dims, (List, int)
        ), "The dimensions of convolutional layers must be given as a list of integers or an integer!"
        assert isinstance(
            conv_kernel, int
        ), "The sizes of convolutional kernels must be given as an integer!"
        assert isinstance(
            conv_stride, int
        ), "The lengths of convolutional strides must be given as an integer!"
        if conv_dropout is not None:
            assert isinstance(
                conv_dropout, (List, float)
            ), "The dropout rates of convolutional layers must be given as a list of integers or an integer!"

        # input_size initialization
        if self.input_size is not None:
            feat_dim = self.input_size
        else:
            assert feat_dim is not None
            self.input_size = feat_dim

        # --- 1. Convolutional Part Initialization --- #
        # register convolution arguments
        self.conv_dims = conv_dims if isinstance(conv_dims, List) else [conv_dims]
        self.conv_kernel = conv_kernel
        self.conv_stride = conv_stride
        self.conv_dropout = conv_dropout

        # Conv1d blocks construction
        _prev_dim = feat_dim
        _tmp_conv = []
        for i in range(len(self.conv_dims)):
            # 0 means go back to the input feat_dim
            if self.conv_dims[i] == 0:
                self.conv_dims[i] = feat_dim
            # -1 means equal to the previous layer
            elif self.conv_dims[i] == -1:
                self.conv_dims[i] = self.conv_dims[i - 1]
            # Conv1d layer
            _tmp_conv.append(
                Conv1dEv(
                    in_channels=_prev_dim,
                    out_channels=self.conv_dims[i],
                    kernel_size=self.conv_kernel,
                    stride=self.conv_stride,
                    padding_mode="same",
                )
            )
            # ReLU function
            _tmp_conv.append(torch.nn.ReLU())
            # LayerNorm layer
            _tmp_conv.append(LayerNorm(self.conv_dims[i], dim=1))
            # Dropout layer
            if conv_dropout is not None:
                _tmp_conv.append(
                    torch.nn.Dropout(
                        p=(
                            self.conv_dropout
                            if not isinstance(self.conv_dropout, List)
                            else self.conv_dropout[i]
                        )
                    )
                )
            _prev_dim = conv_dims[i]
        self.conv = torch.nn.Sequential(*_tmp_conv)

        # --- 2. Linear Part Initialization --- #
        self.linear = torch.nn.Linear(_prev_dim, 1)
        if use_gate:
            self.gate_linear = torch.nn.Linear(_prev_dim, 1)

        # --- 3. Scalar Embedding Part Initialization --- #
        if use_conv_emb:
            _tmp_conv_emb = [
                Conv1dEv(
                    in_channels=1,
                    out_channels=self.input_size,
                    kernel_size=conv_emb_kernel,
                    padding_mode="same",
                )
            ]
            if conv_emb_dropout > 0:
                _tmp_conv_emb.append(torch.nn.Dropout(p=conv_emb_dropout))
            self.conv_emb = torch.nn.Sequential(*_tmp_conv_emb)
        self.output_size = self.input_size

    def forward(self, feat: torch.Tensor, feat_len: torch.Tensor):
        """

        Args:
            feat: (batch, feat_maxlen, feat_dim)
                The input feature tensors.
            feat_len: (batch,)
                The length of each feature tensor.

        Returns: feat, feat_len
            The embedded feature vectors with their lengths.

        """
        # forward the convolutional layers
        # (batch, feat_maxlen, feat_dim) -> (batch, feat_dim, feat_maxlen)
        feat = feat.transpose(1, 2)
        # (batch, feat_dim, feat_maxlen) -> (batch, conv_dim, feat_maxlen)
        feat = self.conv(feat)
        # (batch, conv_dim, feat_maxlen) -> (batch, feat_maxlen, conv_dim)
        feat = feat.transpose(1, 2)

        # forward the linear layer
        # (batch, feat_maxlen, conv_dim) -> (batch, feat_maxlen, 1) -> (batch, feat_maxlen)
        feat_pred = self.linear(feat).squeeze(-1)

        # return feat_len for the compatibility with other prenets
        if not hasattr(self, "gate_linear"):
            return feat_pred, feat_len
        else:
            feat_gate = self.gate_linear(feat).squeeze(-1)
            return feat_pred, feat_len, feat_gate

    def emb_pred_scalar(self, pred_scalar: torch.Tensor):
        """

        Args:
            pred_scalar: (batch, feat_maxlen, 1) or (batch, feat_maxlen)
                The predicted scalar vectors calculated in the forward().

        """
        assert hasattr(
            self, "conv_emb"
        ), "Please set the argument 'use_conv_emb' to True if you want to embed the predicted scalar!"
        if len(pred_scalar.shape) == 2:
            pred_scalar = pred_scalar.unsqueeze(-1)
        else:
            assert len(pred_scalar.shape) == 3 and pred_scalar.size(-1) == 1

        return self.conv_emb(pred_scalar.transpose(1, 2)).transpose(1, 2)

emb_pred_scalar(pred_scalar)

Parameters:

Name Type Description Default
pred_scalar Tensor

(batch, feat_maxlen, 1) or (batch, feat_maxlen) The predicted scalar vectors calculated in the forward().

 required
Source code in speechain/module/prenet/var_pred.py 
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
def emb_pred_scalar(self, pred_scalar: torch.Tensor):
    """

    Args:
        pred_scalar: (batch, feat_maxlen, 1) or (batch, feat_maxlen)
            The predicted scalar vectors calculated in the forward().

    """
    assert hasattr(
        self, "conv_emb"
    ), "Please set the argument 'use_conv_emb' to True if you want to embed the predicted scalar!"
    if len(pred_scalar.shape) == 2:
        pred_scalar = pred_scalar.unsqueeze(-1)
    else:
        assert len(pred_scalar.shape) == 3 and pred_scalar.size(-1) == 1

    return self.conv_emb(pred_scalar.transpose(1, 2)).transpose(1, 2)

forward(feat, feat_len)

Parameters:

Name Type Description Default
feat Tensor

(batch, feat_maxlen, feat_dim) The input feature tensors.

 required
feat_len Tensor

(batch,) The length of each feature tensor.

 required

feat, feat_len

Type Description

The embedded feature vectors with their lengths.
Source code in speechain/module/prenet/var_pred.py 
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
def forward(self, feat: torch.Tensor, feat_len: torch.Tensor):
    """

    Args:
        feat: (batch, feat_maxlen, feat_dim)
            The input feature tensors.
        feat_len: (batch,)
            The length of each feature tensor.

    Returns: feat, feat_len
        The embedded feature vectors with their lengths.

    """
    # forward the convolutional layers
    # (batch, feat_maxlen, feat_dim) -> (batch, feat_dim, feat_maxlen)
    feat = feat.transpose(1, 2)
    # (batch, feat_dim, feat_maxlen) -> (batch, conv_dim, feat_maxlen)
    feat = self.conv(feat)
    # (batch, conv_dim, feat_maxlen) -> (batch, feat_maxlen, conv_dim)
    feat = feat.transpose(1, 2)

    # forward the linear layer
    # (batch, feat_maxlen, conv_dim) -> (batch, feat_maxlen, 1) -> (batch, feat_maxlen)
    feat_pred = self.linear(feat).squeeze(-1)

    # return feat_len for the compatibility with other prenets
    if not hasattr(self, "gate_linear"):
        return feat_pred, feat_len
    else:
        feat_gate = self.gate_linear(feat).squeeze(-1)
        return feat_pred, feat_len, feat_gate

module_init(feat_dim=None, conv_dims=[256, 256], conv_kernel=3, conv_stride=1, use_gate=False, conv_dropout=0.5, use_conv_emb=True, conv_emb_kernel=1, conv_emb_dropout=0.0)

Parameters:

Name Type Description Default
feat_dim int

int The dimension of input acoustic feature tensors. Used for calculating the in_features of the first Linear layer.

 None
conv_dims int or List[int]

List[int] or int The values of out_channels of each Conv1d layer. If a list of integers is given, multiple Conv1d layers will be initialized. If an integer is given, there will be only one Conv1d layer

 [256, 256]
conv_kernel int

int The value of kernel_size of all Conv1d layers.

 3
conv_stride int

int The value of stride of all Conv1d layers.

 1
conv_dropout float or List[float]

float or List[float] The values of p rate of the Dropout layer after each Linear layer.

 0.5
use_conv_emb bool

bool Whether to embed the predicted scalar back to an embedding vector. This argument needs to be False for duration predictor.

 True
conv_emb_kernel int

int The value of kernel_size for the conv1d embedding layer. Only effective when use_conv_emb is True.

 1
conv_emb_dropout float

float The value of p reate of the Dropout layer after the conv1d embedding layer. Only effective when use_conv_emb is True.

 0.0
Source code in speechain/module/prenet/var_pred.py 
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 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
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
def module_init(
    self,
    feat_dim: int = None,
    conv_dims: int or List[int] = [256, 256],
    conv_kernel: int = 3,
    conv_stride: int = 1,
    use_gate: bool = False,
    conv_dropout: float or List[float] = 0.5,
    use_conv_emb: bool = True,
    conv_emb_kernel: int = 1,
    conv_emb_dropout: float = 0.0,
):
    """

    Args:
        feat_dim: int
            The dimension of input acoustic feature tensors.
            Used for calculating the in_features of the first Linear layer.
        conv_dims: List[int] or int
            The values of out_channels of each Conv1d layer.
            If a list of integers is given, multiple Conv1d layers will be initialized.
            If an integer is given, there will be only one Conv1d layer
        conv_kernel: int
            The value of kernel_size of all Conv1d layers.
        conv_stride: int
            The value of stride of all Conv1d layers.
        conv_dropout: float or List[float]
            The values of p rate of the Dropout layer after each Linear layer.
        use_conv_emb: bool
            Whether to embed the predicted scalar back to an embedding vector.
            This argument needs to be False for duration predictor.
        conv_emb_kernel: int
            The value of kernel_size for the conv1d embedding layer.
            Only effective when use_conv_emb is True.
        conv_emb_dropout: float
            The value of p reate of the Dropout layer after the conv1d embedding layer.
            Only effective when use_conv_emb is True.
    """
    # --- 0. Argument Checking --- #
    # Convolution arguments checking
    assert isinstance(
        conv_dims, (List, int)
    ), "The dimensions of convolutional layers must be given as a list of integers or an integer!"
    assert isinstance(
        conv_kernel, int
    ), "The sizes of convolutional kernels must be given as an integer!"
    assert isinstance(
        conv_stride, int
    ), "The lengths of convolutional strides must be given as an integer!"
    if conv_dropout is not None:
        assert isinstance(
            conv_dropout, (List, float)
        ), "The dropout rates of convolutional layers must be given as a list of integers or an integer!"

    # input_size initialization
    if self.input_size is not None:
        feat_dim = self.input_size
    else:
        assert feat_dim is not None
        self.input_size = feat_dim

    # --- 1. Convolutional Part Initialization --- #
    # register convolution arguments
    self.conv_dims = conv_dims if isinstance(conv_dims, List) else [conv_dims]
    self.conv_kernel = conv_kernel
    self.conv_stride = conv_stride
    self.conv_dropout = conv_dropout

    # Conv1d blocks construction
    _prev_dim = feat_dim
    _tmp_conv = []
    for i in range(len(self.conv_dims)):
        # 0 means go back to the input feat_dim
        if self.conv_dims[i] == 0:
            self.conv_dims[i] = feat_dim
        # -1 means equal to the previous layer
        elif self.conv_dims[i] == -1:
            self.conv_dims[i] = self.conv_dims[i - 1]
        # Conv1d layer
        _tmp_conv.append(
            Conv1dEv(
                in_channels=_prev_dim,
                out_channels=self.conv_dims[i],
                kernel_size=self.conv_kernel,
                stride=self.conv_stride,
                padding_mode="same",
            )
        )
        # ReLU function
        _tmp_conv.append(torch.nn.ReLU())
        # LayerNorm layer
        _tmp_conv.append(LayerNorm(self.conv_dims[i], dim=1))
        # Dropout layer
        if conv_dropout is not None:
            _tmp_conv.append(
                torch.nn.Dropout(
                    p=(
                        self.conv_dropout
                        if not isinstance(self.conv_dropout, List)
                        else self.conv_dropout[i]
                    )
                )
            )
        _prev_dim = conv_dims[i]
    self.conv = torch.nn.Sequential(*_tmp_conv)

    # --- 2. Linear Part Initialization --- #
    self.linear = torch.nn.Linear(_prev_dim, 1)
    if use_gate:
        self.gate_linear = torch.nn.Linear(_prev_dim, 1)

    # --- 3. Scalar Embedding Part Initialization --- #
    if use_conv_emb:
        _tmp_conv_emb = [
            Conv1dEv(
                in_channels=1,
                out_channels=self.input_size,
                kernel_size=conv_emb_kernel,
                padding_mode="same",
            )
        ]
        if conv_emb_dropout > 0:
            _tmp_conv_emb.append(torch.nn.Dropout(p=conv_emb_dropout))
        self.conv_emb = torch.nn.Sequential(*_tmp_conv_emb)
    self.output_size = self.input_size

LayerNorm

Bases: LayerNorm

Layer normalization module. Borrowed from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/layer_norm.py

Parameters:

Name Type Description Default
nout int

Output dim size.

 required
dim int

Dimension to be normalized.

 -1
Source code in speechain/module/prenet/var_pred.py 
 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
36
37
38
39
class LayerNorm(torch.nn.LayerNorm):
    """
    Layer normalization module.
    Borrowed from
        https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/layer_norm.py

    Args:
        nout (int): Output dim size.
        dim (int): Dimension to be normalized.
    """

    def __init__(self, nout, dim=-1):
        """Construct an LayerNorm object."""
        super(LayerNorm, self).__init__(nout, eps=1e-12)
        self.dim = dim

    def forward(self, x):
        """Apply layer normalization.

        Args:
            x (torch.Tensor): Input tensor.
        Returns:
            torch.Tensor: Normalized tensor.
        """
        if self.dim == -1:
            return super(LayerNorm, self).forward(x)
        return (
            super(LayerNorm, self)
            .forward(x.transpose(self.dim, -1))
            .transpose(self.dim, -1)
        )

__init__(nout, dim=-1)

Construct an LayerNorm object.

Source code in speechain/module/prenet/var_pred.py 
20
21
22
23
def __init__(self, nout, dim=-1):
    """Construct an LayerNorm object."""
    super(LayerNorm, self).__init__(nout, eps=1e-12)
    self.dim = dim

forward(x)

Apply layer normalization.

Parameters:

Name Type Description Default
x Tensor

Input tensor.

 required

Returns: torch.Tensor: Normalized tensor.

Source code in speechain/module/prenet/var_pred.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
def forward(self, x):
    """Apply layer normalization.

    Args:
        x (torch.Tensor): Input tensor.
    Returns:
        torch.Tensor: Normalized tensor.
    """
    if self.dim == -1:
        return super(LayerNorm, self).forward(x)
    return (
        super(LayerNorm, self)
        .forward(x.transpose(self.dim, -1))
        .transpose(self.dim, -1)
    )