Skip to content

conv1d

Author: Sashi Novitasari Affiliation: NAIST (-2022) Date: 2022.08

Author: Heli Qi Affiliation: NAIST Date: 2022.09

Conv1dEv

Bases: Module

A 1D convolutional layer with support for different padding modes.

Attributes:

Name Type Description
cutoff bool

Indicates whether the output should be cut off for the 'same' padding mode.
causal_padding int

Additional padding required for the 'causal' padding mode.
dilation int

The dilation rate of the convolutional layer.
conv_lyr Conv1d

The 1D convolutional layer.
Source code in speechain/module/prenet/conv1d.py 
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 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
 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
class Conv1dEv(torch.nn.Module):
    """A 1D convolutional layer with support for different padding modes.

    Attributes:
        cutoff (bool):
            Indicates whether the output should be cut off for the 'same' padding mode.
        causal_padding (int):
            Additional padding required for the 'causal' padding mode.
        dilation (int):
            The dilation rate of the convolutional layer.
        conv_lyr (torch.nn.Conv1d):
            The 1D convolutional layer.
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        stride: int = 1,
        dilation: int = 1,
        padding_mode: str = "same",
        bias: bool = True,
        use_weight_norm: bool = False,
        groups: int = 1,
    ):
        """Initializes the Conv1dEv module with the specified parameters.

        Args:
            in_channels (int):
                Number of channels in the input feature.
            out_channels (int):
                Number of channels produced by the convolution.
            kernel_size (int):
                Size of the convolutional kernel.
            stride (int, optional):
                Stride of the convolution. Defaults to 1.
            dilation (int, optional):
                The dilation rate of the kernel. Defaults to 1.
            padding_mode (str, optional):
                Padding mode. Supported values are 'valid', 'full', 'same' and 'causal'. Defaults to 'same'.
            bias (bool, optional):
                If True, adds a learnable bias to the output. Defaults to True.

        Raises:
            ValueError: If an unsupported padding mode is specified.
        """
        super().__init__()

        self.cutoff = False
        self.causal_padding = 0
        self.dilation = dilation

        # no padding is used
        if padding_mode == "valid":
            padding = 0
        # full padding
        elif padding_mode == "full":
            padding = dilation * (kernel_size - 1)
        # same padding, the output is the same in dimension with input
        elif padding_mode == "same":
            assert stride == 1, "Stride should be 1 for 'same' padding mode"
            if kernel_size % 2 == 0:
                padding = dilation * kernel_size // 2
                self.cutoff = True
            else:
                padding = dilation * (kernel_size - 1) // 2
        # causal padding
        elif padding_mode == "causal":
            padding = 0
            self.causal_padding = dilation * (kernel_size - 1)
        else:
            raise ValueError(
                "Unsupported padding mode. Supported modes are 'valid', 'full', 'same' and 'causal'."
            )

        self.conv_lyr = torch.nn.Conv1d(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding,
            bias=bias,
            groups=groups,
        )
        if use_weight_norm:
            self.conv_lyr = weight_norm(self.conv_lyr)

    def forward(self, feat: torch.Tensor):
        """Performs a forward pass through the convolutional layer.

        Args:
            feat (torch.Tensor):
                The input feature tensor. Shape: (batch, feat_dim, feat_maxlen).

        Returns:
            torch.Tensor:
                The output tensor. Shape: (batch, out_channels, output_len).
        """
        # attach additional paddings at the end for the 'causal' padding mode
        if self.causal_padding > 0:
            feat = F.pad(feat, (self.causal_padding, 0))
        output = self.conv_lyr(feat)
        # cut off the redundant tails for the 'same' padding mode
        if self.cutoff:
            output = output[:, :, : -self.dilation]
        return output

__init__(in_channels, out_channels, kernel_size, stride=1, dilation=1, padding_mode='same', bias=True, use_weight_norm=False, groups=1)

Initializes the Conv1dEv module with the specified parameters.

Parameters:

Name Type Description Default
in_channels int

Number of channels in the input feature.

 required
out_channels int

Number of channels produced by the convolution.

 required
kernel_size int

Size of the convolutional kernel.

 required
stride int

Stride of the convolution. Defaults to 1.

 1
dilation int

The dilation rate of the kernel. Defaults to 1.

 1
padding_mode str

Padding mode. Supported values are 'valid', 'full', 'same' and 'causal'. Defaults to 'same'.

 'same'
bias bool

If True, adds a learnable bias to the output. Defaults to True.

 True

Raises:

Type Description
ValueError

If an unsupported padding mode is specified.
Source code in speechain/module/prenet/conv1d.py 
 35
 36
 37
 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
 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
def __init__(
    self,
    in_channels: int,
    out_channels: int,
    kernel_size: int,
    stride: int = 1,
    dilation: int = 1,
    padding_mode: str = "same",
    bias: bool = True,
    use_weight_norm: bool = False,
    groups: int = 1,
):
    """Initializes the Conv1dEv module with the specified parameters.

    Args:
        in_channels (int):
            Number of channels in the input feature.
        out_channels (int):
            Number of channels produced by the convolution.
        kernel_size (int):
            Size of the convolutional kernel.
        stride (int, optional):
            Stride of the convolution. Defaults to 1.
        dilation (int, optional):
            The dilation rate of the kernel. Defaults to 1.
        padding_mode (str, optional):
            Padding mode. Supported values are 'valid', 'full', 'same' and 'causal'. Defaults to 'same'.
        bias (bool, optional):
            If True, adds a learnable bias to the output. Defaults to True.

    Raises:
        ValueError: If an unsupported padding mode is specified.
    """
    super().__init__()

    self.cutoff = False
    self.causal_padding = 0
    self.dilation = dilation

    # no padding is used
    if padding_mode == "valid":
        padding = 0
    # full padding
    elif padding_mode == "full":
        padding = dilation * (kernel_size - 1)
    # same padding, the output is the same in dimension with input
    elif padding_mode == "same":
        assert stride == 1, "Stride should be 1 for 'same' padding mode"
        if kernel_size % 2 == 0:
            padding = dilation * kernel_size // 2
            self.cutoff = True
        else:
            padding = dilation * (kernel_size - 1) // 2
    # causal padding
    elif padding_mode == "causal":
        padding = 0
        self.causal_padding = dilation * (kernel_size - 1)
    else:
        raise ValueError(
            "Unsupported padding mode. Supported modes are 'valid', 'full', 'same' and 'causal'."
        )

    self.conv_lyr = torch.nn.Conv1d(
        in_channels=in_channels,
        out_channels=out_channels,
        kernel_size=kernel_size,
        stride=stride,
        dilation=dilation,
        padding=padding,
        bias=bias,
        groups=groups,
    )
    if use_weight_norm:
        self.conv_lyr = weight_norm(self.conv_lyr)

forward(feat)

Performs a forward pass through the convolutional layer.

Parameters:

Name Type Description Default
feat Tensor

The input feature tensor. Shape: (batch, feat_dim, feat_maxlen).

 required

Returns:

Type Description

torch.Tensor: The output tensor. Shape: (batch, out_channels, output_len).
Source code in speechain/module/prenet/conv1d.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
def forward(self, feat: torch.Tensor):
    """Performs a forward pass through the convolutional layer.

    Args:
        feat (torch.Tensor):
            The input feature tensor. Shape: (batch, feat_dim, feat_maxlen).

    Returns:
        torch.Tensor:
            The output tensor. Shape: (batch, out_channels, output_len).
    """
    # attach additional paddings at the end for the 'causal' padding mode
    if self.causal_padding > 0:
        feat = F.pad(feat, (self.causal_padding, 0))
    output = self.conv_lyr(feat)
    # cut off the redundant tails for the 'same' padding mode
    if self.cutoff:
        output = output[:, :, : -self.dilation]
    return output

Conv1dPrenet

Bases: Module

The Conv1d prenet. Usually used before the TTS encoder. This prenet is made up of two parts: 1. (mandatory) The Conv1d part contains one or more Conv1d blocks which are composed of the components below 1. (mandatory) a Conv1d layer 2. (optional) a BatchNorm1d layer 3. (optional) an activation function 4. (optional) a Dropout layer. 2. (optional) The Linear part contains one or more Linear blocks which are composed of the components below 1. (mandatory) a Linear layer 2. (optional) an activation function 3. (optional) a Dropout layer.

Reference

Neural Speech Synthesis with Transformer Network https://ojs.aaai.org/index.php/AAAI/article/view/4642/4520

Source code in speechain/module/prenet/conv1d.py 
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
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
class Conv1dPrenet(Module):
    """
    The Conv1d prenet. Usually used before the TTS encoder.
    This prenet is made up of two parts:
        1. (mandatory) The Conv1d part contains one or more Conv1d blocks which are composed of the components below
            1. (mandatory) a Conv1d layer
            2. (optional) a BatchNorm1d layer
            3. (optional) an activation function
            4. (optional) a Dropout layer.
        2. (optional) The Linear part contains one or more Linear blocks which are composed of the components below
            1. (mandatory) a Linear layer
            2. (optional) an activation function
            3. (optional) a Dropout layer.

    Reference:
        Neural Speech Synthesis with Transformer Network
        https://ojs.aaai.org/index.php/AAAI/article/view/4642/4520
    """

    def module_init(
        self,
        feat_dim: int = None,
        conv_dims: int or List[int] = [512, 512, 512],
        conv_kernel: int = 5,
        conv_stride: int = 1,
        conv_batchnorm: bool = True,
        conv_activation: str = "ReLU",
        conv_dropout: float or List[float] = None,
        lnr_dims: int or List[int] = -1,
        lnr_activation: str = None,
        lnr_dropout: int or List[int] = None,
        zero_centered: bool = False,
    ):
        """

        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_batchnorm: bool
                Whether a BatchNorm1d layer is added right after a Conv1d layer
            conv_activation: str
                The type of the activation function after all Conv1d layers.
                None means no activation function is needed.
            conv_dropout: float or List[float]
                The values of p rate of the Dropout layer after each Linear layer.
            lnr_dims: int or List[int]
                The values of out_features of each Linear layer.
                The first value in the List represents the out_features of the first Linear layer.
                -1: same size as the last convolutional layer's dim
            lnr_activation: str
                The type of the activation function after all Linear layers.
                None means no activation function is needed.
            lnr_dropout: float or List[float]
                The values of p rate of the Dropout layer after each Linear layer.
            zero_centered: bool
                Whether the output of this module is centered at 0.
                If the specified activation function changes the centroid of the output distribution, e.g. ReLU and
                LeakyReLU, the activation function won't be attached to the final Linear layer if zer_centered is set
                to 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!"

        # Linear arguments checking
        if lnr_dropout is not None:
            assert isinstance(
                lnr_dropout, (List, float)
            ), "The dropout rates of linear layers must be given as a list of integers or an integer!"
        if lnr_dims is not None:
            assert isinstance(
                lnr_dims, (List, int)
            ), "The dimensions of linear 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

        # --- 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)):
            _tmp_conv.append(
                # don't include bias in the convolutional layer if it is followed by a batchnorm layer
                # reference: https://stackoverflow.com/questions/46256747/can-not-use-both-bias-and-batch-normalization-in-convolution-layers
                Conv1dEv(
                    in_channels=_prev_dim,
                    out_channels=self.conv_dims[i],
                    kernel_size=self.conv_kernel,
                    stride=self.conv_stride,
                    padding_mode="same",
                    bias=not conv_batchnorm,
                )
            )
            # BatchNorm is better to be placed before activation
            # reference: https://stackoverflow.com/questions/39691902/ordering-of-batch-normalization-and-dropout
            if conv_batchnorm:
                _tmp_conv.append(torch.nn.BatchNorm1d(self.conv_dims[i]))
            if conv_activation is not None:
                # no 'ReLU'-series activation is added for the last layer if zero_centered is specified
                if not (i == len(self.conv_dims) - 1 and lnr_dims is None) or not (
                    zero_centered and "ReLU" in conv_activation
                ):
                    _tmp_conv.append(getattr(torch.nn, conv_activation)())
            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)
        self.output_size = _prev_dim

        # --- 2. Linear Part Initialization --- #
        if lnr_dims is not None:
            lnr_dims = lnr_dims if isinstance(lnr_dims, List) else [lnr_dims]
            for i in range(len(lnr_dims)):
                _prev_dim = self.conv_dims[-1] if i == 0 else lnr_dims[i - 1]
                if lnr_dims[i] == -1:
                    lnr_dims[i] = _prev_dim

            self.linear = LinearPrenet(
                feat_dim=self.output_size,
                lnr_dims=lnr_dims,
                lnr_activation=lnr_activation,
                lnr_dropout=lnr_dropout,
                zero_centered=zero_centered,
            )
            self.output_size = self.linear.output_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 layers
        if hasattr(self, "linear"):
            # (batch, feat_maxlen, conv_dim) -> (batch, feat_maxlen, lnr_dim)
            feat, feat_len = self.linear(feat, feat_len)

        # return both feat & feat_len for the compatibility with other prenet
        return feat, feat_len

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/conv1d.py 
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
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 layers
    if hasattr(self, "linear"):
        # (batch, feat_maxlen, conv_dim) -> (batch, feat_maxlen, lnr_dim)
        feat, feat_len = self.linear(feat, feat_len)

    # return both feat & feat_len for the compatibility with other prenet
    return feat, feat_len

module_init(feat_dim=None, conv_dims=[512, 512, 512], conv_kernel=5, conv_stride=1, conv_batchnorm=True, conv_activation='ReLU', conv_dropout=None, lnr_dims=-1, lnr_activation=None, lnr_dropout=None, zero_centered=False)

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

 [512, 512, 512]
conv_kernel int

int The value of kernel_size of all Conv1d layers.

 5
conv_stride int

int The value of stride of all Conv1d layers.

 1
conv_batchnorm bool

bool Whether a BatchNorm1d layer is added right after a Conv1d layer

 True
conv_activation str

str The type of the activation function after all Conv1d layers. None means no activation function is needed.

 'ReLU'
conv_dropout float or List[float]

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

 None
lnr_dims int or List[int]

int or List[int] The values of out_features of each Linear layer. The first value in the List represents the out_features of the first Linear layer. -1: same size as the last convolutional layer's dim

 -1
lnr_activation str

str The type of the activation function after all Linear layers. None means no activation function is needed.

 None
lnr_dropout int or List[int]

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

 None
zero_centered bool

bool Whether the output of this module is centered at 0. If the specified activation function changes the centroid of the output distribution, e.g. ReLU and LeakyReLU, the activation function won't be attached to the final Linear layer if zer_centered is set to True.

 False
Source code in speechain/module/prenet/conv1d.py 
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
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
def module_init(
    self,
    feat_dim: int = None,
    conv_dims: int or List[int] = [512, 512, 512],
    conv_kernel: int = 5,
    conv_stride: int = 1,
    conv_batchnorm: bool = True,
    conv_activation: str = "ReLU",
    conv_dropout: float or List[float] = None,
    lnr_dims: int or List[int] = -1,
    lnr_activation: str = None,
    lnr_dropout: int or List[int] = None,
    zero_centered: bool = False,
):
    """

    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_batchnorm: bool
            Whether a BatchNorm1d layer is added right after a Conv1d layer
        conv_activation: str
            The type of the activation function after all Conv1d layers.
            None means no activation function is needed.
        conv_dropout: float or List[float]
            The values of p rate of the Dropout layer after each Linear layer.
        lnr_dims: int or List[int]
            The values of out_features of each Linear layer.
            The first value in the List represents the out_features of the first Linear layer.
            -1: same size as the last convolutional layer's dim
        lnr_activation: str
            The type of the activation function after all Linear layers.
            None means no activation function is needed.
        lnr_dropout: float or List[float]
            The values of p rate of the Dropout layer after each Linear layer.
        zero_centered: bool
            Whether the output of this module is centered at 0.
            If the specified activation function changes the centroid of the output distribution, e.g. ReLU and
            LeakyReLU, the activation function won't be attached to the final Linear layer if zer_centered is set
            to 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!"

    # Linear arguments checking
    if lnr_dropout is not None:
        assert isinstance(
            lnr_dropout, (List, float)
        ), "The dropout rates of linear layers must be given as a list of integers or an integer!"
    if lnr_dims is not None:
        assert isinstance(
            lnr_dims, (List, int)
        ), "The dimensions of linear 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

    # --- 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)):
        _tmp_conv.append(
            # don't include bias in the convolutional layer if it is followed by a batchnorm layer
            # reference: https://stackoverflow.com/questions/46256747/can-not-use-both-bias-and-batch-normalization-in-convolution-layers
            Conv1dEv(
                in_channels=_prev_dim,
                out_channels=self.conv_dims[i],
                kernel_size=self.conv_kernel,
                stride=self.conv_stride,
                padding_mode="same",
                bias=not conv_batchnorm,
            )
        )
        # BatchNorm is better to be placed before activation
        # reference: https://stackoverflow.com/questions/39691902/ordering-of-batch-normalization-and-dropout
        if conv_batchnorm:
            _tmp_conv.append(torch.nn.BatchNorm1d(self.conv_dims[i]))
        if conv_activation is not None:
            # no 'ReLU'-series activation is added for the last layer if zero_centered is specified
            if not (i == len(self.conv_dims) - 1 and lnr_dims is None) or not (
                zero_centered and "ReLU" in conv_activation
            ):
                _tmp_conv.append(getattr(torch.nn, conv_activation)())
        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)
    self.output_size = _prev_dim

    # --- 2. Linear Part Initialization --- #
    if lnr_dims is not None:
        lnr_dims = lnr_dims if isinstance(lnr_dims, List) else [lnr_dims]
        for i in range(len(lnr_dims)):
            _prev_dim = self.conv_dims[-1] if i == 0 else lnr_dims[i - 1]
            if lnr_dims[i] == -1:
                lnr_dims[i] = _prev_dim

        self.linear = LinearPrenet(
            feat_dim=self.output_size,
            lnr_dims=lnr_dims,
            lnr_activation=lnr_activation,
            lnr_dropout=lnr_dropout,
            zero_centered=zero_centered,
        )
        self.output_size = self.linear.output_size