Skip to content

conformer

Conformer Model

Conformer model implementation in Keras.

Classes:

Functions:

Classes

SubsampleBlockParams

Subsample block parameters

Attributes:

  • depth (int) –

    Depth

  • kernel_size (int) –

    Kernel size

  • strides (int) –

    Stride size

ConformerBlockParams

Conformer block parameters

Attributes:

  • depth (int) –

    Depth

  • fc_ex_factor (float) –

    FC expansion factor

  • fc_res_factor (float) –

    FC residual factor

  • embedding (str) –

    Embedding type

  • num_heads (int) –

    Number of heads

  • kernel_size (int) –

    Kernel size

  • dropout (float) –

    Dropout rate

  • use_bias (bool) –

    Use bias

ConformerParams

Conformer parameters

Attributes:

ConformerModel

Helper class to generate model from parameters

Functions

layer_from_params staticmethod 
layer_from_params(inputs: keras.Input, params: ConformerParams | dict, num_classes: int | None = None)

Create layer from parameters

Source code in neuralspot_edge/models/conformer.py 
463
464
465
466
467
468
@staticmethod
def layer_from_params(inputs: keras.Input, params: ConformerParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = ConformerParams(**params)
    return conformer_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: ConformerParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/conformer.py 
470
471
472
473
474
@staticmethod
def model_from_params(inputs: keras.Input, params: ConformerParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = ConformerModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

subsampler 

subsampler(blocks: SubsampleBlockParams, kernel_initializer: str = 'glorot_uniform', bias_initializer: str = 'zeros', kernel_regularizer=None, bias_regularizer=None, name: str | None = None) -> keras.Layer

Subsampler block

Parameters:

  • blocks (SubsampleBlockParams) –

    Subsample block parameters

  • kernel_initializer (str, default: 'glorot_uniform' ) –

    Kernel initializer. Defaults to "glorot_uniform".

  • bias_initializer (str, default: 'zeros' ) –

    Bias initializer. Defaults to "zeros".

  • kernel_regularizer ([type], default: None ) –

    Kernel regularizer. Defaults to None.

  • bias_regularizer ([type], default: None ) –

    Bias regularizer. Defaults to None.

  • name (str, default: None ) –

    Name. Defaults to None.

Returns:

  • Layer –

    keras.Layer: Subsampler layer

Source code in neuralspot_edge/models/conformer.py 
 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
def subsampler(
    blocks: SubsampleBlockParams,
    kernel_initializer: str = "glorot_uniform",
    bias_initializer: str = "zeros",
    kernel_regularizer=None,
    bias_regularizer=None,
    name: str | None = None,
) -> keras.Layer:
    """Subsampler block

    Args:
        blocks (SubsampleBlockParams): Subsample block parameters
        kernel_initializer (str, optional): Kernel initializer. Defaults to "glorot_uniform".
        bias_initializer (str, optional): Bias initializer. Defaults to "zeros".
        kernel_regularizer ([type], optional): Kernel regularizer. Defaults to None.
        bias_regularizer ([type], optional): Bias regularizer. Defaults to None.
        name (str, optional): Name. Defaults to None.

    Returns:
        keras.Layer: Subsampler layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        y = x
        # Apply subsampling blocks
        for i, block in enumerate(blocks):
            y = keras.layers.SeparableConv2D(
                filters=block.depth,
                kernel_size=block.kernel_size,
                strides=block.strides,
                depthwise_initializer=kernel_initializer,
                pointwise_initializer=kernel_initializer,
                depthwise_regularizer=kernel_regularizer,
                pointwise_regularizer=kernel_regularizer,
                bias_initializer=bias_initializer,
                bias_regularizer=bias_regularizer,
                padding="same",
                name=f"{name}_conv{i+1}",
            )(y)
            y = relu(name=f"{name}_relu{i+1}")(y)
        # END FOR

        # Swap from (b,f,t,c) to (b,t,c,f) and merge (c,f) to get (b,t,c*f)
        y = keras.ops.transpose(y, axes=[0, 2, 3, 1])
        y = keras.layers.Reshape((y.shape[1], -1))(y)

        # Can serve as learnable embedding
        y = keras.layers.Dense(
            units=blocks[-1].depth,
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            name=f"{name}_dense",
        )(y)  # (b,t,c*f) -> (b,t,d)
        return y

    # END DEF
    return layer

fc_block 

fc_block(depth: int, ex_factor: int = 4, residual_factor: float = 0.5, dropout: float = 0, use_bias: bool = True, kernel_initializer: str = 'glorot_uniform', bias_initializer: str = 'zeros', kernel_regularizer=None, bias_regularizer=None, name: str = 'fc_block') -> keras.Layer

Fully connected block

Parameters:

  • depth (int) –

    Depth

  • ex_factor (int, default: 4 ) –

    Expansion factor. Defaults to 4.

  • residual_factor (float, default: 0.5 ) –

    Residual factor. Defaults to 0.5.

  • dropout (float, default: 0 ) –

    Dropout rate. Defaults to 0.

  • use_bias (bool, default: True ) –

    Use bias. Defaults to True.

  • kernel_initializer (str, default: 'glorot_uniform' ) –

    Kernel initializer. Defaults to "glorot_uniform".

  • bias_initializer (str, default: 'zeros' ) –

    Bias initializer. Defaults to "zeros".

  • kernel_regularizer ([type], default: None ) –

    Kernel regularizer. Defaults to None.

  • bias_regularizer ([type], default: None ) –

    Bias regularizer. Defaults to None.

  • name (str, default: 'fc_block' ) –

    Name. Defaults to "fc_block".

Returns:

  • Layer –

    keras.Layer: Functional layer

Source code in neuralspot_edge/models/conformer.py 
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
def fc_block(
    depth: int,
    ex_factor: int = 4,
    residual_factor: float = 0.5,
    dropout: float = 0,
    use_bias: bool = True,
    kernel_initializer: str = "glorot_uniform",
    bias_initializer: str = "zeros",
    kernel_regularizer=None,
    bias_regularizer=None,
    name: str = "fc_block",
) -> keras.Layer:
    """Fully connected block

    Args:
        depth (int): Depth
        ex_factor (int, optional): Expansion factor. Defaults to 4.
        residual_factor (float, optional): Residual factor. Defaults to 0.5.
        dropout (float, optional): Dropout rate. Defaults to 0.
        use_bias (bool, optional): Use bias. Defaults to True.
        kernel_initializer (str, optional): Kernel initializer. Defaults to "glorot_uniform".
        bias_initializer (str, optional): Bias initializer. Defaults to "zeros".
        kernel_regularizer ([type], optional): Kernel regularizer. Defaults to None.
        bias_regularizer ([type], optional): Bias regularizer. Defaults to None.
        name (str, optional): Name. Defaults to "fc_block".

    Returns:
        keras.Layer: Functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        y = x

        y = layer_normalization(name=f"{name}.ln")(y)

        y = keras.layers.Dense(
            int(ex_factor * depth),
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            kernel_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            use_bias=use_bias,
            name=f"{name}.fc1",
        )(y)
        y = swish(name=f"{name}.swish")(y)
        if 0 < dropout < 1:
            y = keras.layers.Dropout(dropout, name=f"{name}.dropout1")(y)
        # END IF
        y = keras.layers.Dense(
            depth,
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            kernel_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            use_bias=use_bias,
            name=f"{name}.fc2",
        )(y)
        if 0 < dropout < 1:
            y = keras.layers.Dropout(dropout, name=f"{name}.dropout2")(y)
        # END IF
        return x + residual_factor * y

    # END DEF
    return layer

conv_block 

conv_block(depth: int, kernel_size: int = 9, dropout: float = 0.0, padding: str = 'same', scale_factor: int = 2, kernel_initializer: str = 'glorot_uniform', bias_initializer: str = 'zeros', kernel_regularizer=None, bias_regularizer=None, name: str = 'conv_module') -> keras.Layer

Convolutional block

Parameters:

  • depth (int) –

    Depth

  • kernel_size (int, default: 9 ) –

    Kernel size. Defaults to 9.

  • dropout (float, default: 0.0 ) –

    Dropout rate. Defaults to 0.0.

  • padding (str, default: 'same' ) –

    Padding. Defaults to "same".

  • scale_factor (int, default: 2 ) –

    Scale factor. Defaults to 2.

  • kernel_initializer (str, default: 'glorot_uniform' ) –

    Kernel initializer. Defaults to "glorot_uniform".

  • bias_initializer (str, default: 'zeros' ) –

    Bias initializer. Defaults to "zeros".

  • kernel_regularizer ([type], default: None ) –

    Kernel regularizer. Defaults to None.

  • bias_regularizer ([type], default: None ) –

    Bias regularizer. Defaults to None.

  • name (str, default: 'conv_module' ) –

    Name. Defaults to "conv_module".

Returns:

  • Layer –

    keras.Layer: Functional layer

Source code in neuralspot_edge/models/conformer.py 
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
def conv_block(
    depth: int,
    kernel_size: int = 9,
    dropout: float = 0.0,
    padding: str = "same",
    scale_factor: int = 2,
    kernel_initializer: str = "glorot_uniform",
    bias_initializer: str = "zeros",
    kernel_regularizer=None,
    bias_regularizer=None,
    name: str = "conv_module",
) -> keras.Layer:
    """Convolutional block

    Args:
        depth (int): Depth
        kernel_size (int, optional): Kernel size. Defaults to 9.
        dropout (float, optional): Dropout rate. Defaults to 0.0.
        padding (str, optional): Padding. Defaults to "same".
        scale_factor (int, optional): Scale factor. Defaults to 2.
        kernel_initializer (str, optional): Kernel initializer. Defaults to "glorot_uniform".
        bias_initializer (str, optional): Bias initializer. Defaults to "zeros".
        kernel_regularizer ([type], optional): Kernel regularizer. Defaults to None.
        bias_regularizer ([type], optional): Bias regularizer. Defaults to None.
        name (str, optional): Name. Defaults to "conv_module".

    Returns:
        keras.Layer: Functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        y = x

        y = layer_normalization(name=f"{name}.ln")(y)

        y = keras.layers.Conv1D(
            filters=scale_factor * depth,
            kernel_size=1,
            strides=1,
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            kernel_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            name=f"{name}.pw_conv1",
        )(y)

        y = glu(name=f"{name}.glu")(y)

        y = keras.layers.DepthwiseConv1D(
            kernel_size=kernel_size,
            strides=1,
            padding=padding,
            depthwise_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            depthwise_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            name=f"{name}.dw_conv",
        )(y)
        y = batch_normalization(name=f"{name}.bn")(y)

        y = swish(name=f"{name}.swish")(y)

        y = keras.layers.Conv1D(
            filters=depth,
            kernel_size=1,
            strides=1,
            kernel_initializer=kernel_initializer,
            bias_initializer=bias_initializer,
            kernel_regularizer=kernel_regularizer,
            bias_regularizer=bias_regularizer,
            name=f"{name}.pw_conv2",
        )(y)

        if 0 < dropout < 1:
            y = keras.layers.Dropout(rate=dropout, name=f"{name}.dropout")(y)
        # END IF

        # Residual connection
        return x + y

    # END DEF
    return layer

att_block 

att_block(depth: int, embedding: str = 'rel', num_heads: int = 4, dropout: float = 0.1, name: str = 'att_block') -> keras.Layer

Attention block

Parameters:

  • depth (int) –

    Depth

  • embedding (str, default: 'rel' ) –

    Embedding type. Defaults to "rel".

  • num_heads (int, default: 4 ) –

    Number of heads. Defaults to 4.

  • dropout (float, default: 0.1 ) –

    Dropout rate. Defaults to 0.1.

  • name (str, default: 'att_block' ) –

    Name. Defaults to "att_block".

Returns:

  • Layer –

    keras.Layer: Functional layer

Source code in neuralspot_edge/models/conformer.py 
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
325
326
327
328
329
330
331
332
333
334
335
def att_block(
    depth: int,
    embedding: str = "rel",
    num_heads: int = 4,
    dropout: float = 0.1,
    name: str = "att_block",
) -> keras.Layer:
    """Attention block

    Args:
        depth (int): Depth
        embedding (str, optional): Embedding type. Defaults to "rel".
        num_heads (int, optional): Number of heads. Defaults to 4.
        dropout (float, optional): Dropout rate. Defaults to 0.1.
        name (str, optional): Name. Defaults to "att_block".

    Returns:
        keras.Layer: Functional layer

    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        y = x
        y = layer_normalization(name=f"{name}.ln")(y)

        # Att type: rel, abs, none
        y = keras.layers.MultiHeadAttention(
            num_heads=num_heads,
            key_dim=depth // num_heads,
            value_dim=depth // num_heads,
            dropout=dropout,
            name=f"{name}.mha",
        )(y, y)

        if 0 < dropout < 1:
            y = keras.layers.Dropout(dropout, name=f"{name}.dropout")(y)

        return y

    # END DEF
    return layer

conformer_block 

conformer_block(depth: int, fc_ex_factor: int = 4, fc_res_factor: int = 0.5, embedding: str = 'relative', num_heads: int = 4, kernel_size: int = 9, dropout: float = 0.1, use_bias: bool = True, name: str = 'cf_block') -> keras.Layer

Conformer block

Parameters:

  • depth (int) –

    Depth

  • fc_ex_factor (int, default: 4 ) –

    FC expansion factor. Defaults to 4.

  • fc_res_factor (int, default: 0.5 ) –

    FC residual factor. Defaults to 0.5.

  • embedding (str, default: 'relative' ) –

    Embedding type. Defaults to "relative".

  • num_heads (int, default: 4 ) –

    Number of heads. Defaults to 4.

  • kernel_size (int, default: 9 ) –

    Kernel size. Defaults to 9.

  • dropout (float, default: 0.1 ) –

    Dropout rate. Defaults to 0.1.

  • use_bias (bool, default: True ) –

    Use bias. Defaults to True.

  • name (str, default: 'cf_block' ) –

    Name. Defaults to "cf_block".

Returns:

  • Layer –

    keras.Layer: Functional layer

Source code in neuralspot_edge/models/conformer.py 
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
def conformer_block(
    depth: int,
    fc_ex_factor: int = 4,
    fc_res_factor: int = 0.5,
    embedding: str = "relative",
    num_heads: int = 4,
    kernel_size: int = 9,
    dropout: float = 0.1,
    use_bias: bool = True,
    name: str = "cf_block",
) -> keras.Layer:
    """Conformer block

    Args:
        depth (int): Depth
        fc_ex_factor (int, optional): FC expansion factor. Defaults to 4.
        fc_res_factor (int, optional): FC residual factor. Defaults to 0.5.
        embedding (str, optional): Embedding type. Defaults to "relative".
        num_heads (int, optional): Number of heads. Defaults to 4.
        kernel_size (int, optional): Kernel size. Defaults to 9.
        dropout (float, optional): Dropout rate. Defaults to 0.1.
        use_bias (bool, optional): Use bias. Defaults to True.
        name (str, optional): Name. Defaults to "cf_block".

    Returns:
        keras.Layer: Functional layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        y = x

        # 1st FC Block
        y = fc_block(
            depth=depth,
            ex_factor=fc_ex_factor,
            residual_factor=fc_res_factor,
            dropout=dropout,
            use_bias=use_bias,
            name=f"{name}.fcb1",
        )(y)

        # ATT Block w/ residual
        y = att_block(
            depth=depth,
            embedding=embedding,
            num_heads=num_heads,
            dropout=dropout,
            name=f"{name}.att",
        )(y)

        # Conv Block w/ residual
        y = conv_block(depth=depth, kernel_size=kernel_size, dropout=dropout, name=f"{name}.cvb")(y)

        # 2nd FC Block w/ residual
        y = fc_block(
            depth=depth,
            ex_factor=fc_ex_factor,
            residual_factor=fc_res_factor,
            dropout=dropout,
            use_bias=use_bias,
            name=f"{name}.fcb2",
        )(y)

        # Output Layer Norm
        y = layer_normalization(name=f"{name}.out.ln")(y)

        return y

    # END DEF
    return layer

conformer_layer 

conformer_layer(x: keras.KerasTensor, params: ConformerParams, num_classes: int | None = None) -> keras.KerasTensor

Conformer functional layer

Parameters:

  • x (KerasTensor) –

    Input tensor

  • params (ConformerParams) –

    Model parameters.

  • num_classes (int, default: None ) –

    Number of classes.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/conformer.py 
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
def conformer_layer(
    x: keras.KerasTensor,
    params: ConformerParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """Conformer functional layer

    Args:
        x (keras.KerasTensor): Input tensor
        params (ConformerParams): Model parameters.
        num_classes (int, optional): Number of classes.

    Returns:
        keras.KerasTensor: Output tensor
    """
    y = x

    y = subsampler(
        blocks=params.subsamples,
        kernel_size=params.subsample.kernel_size,
        strides=params.subsample.strides,
        num_downsamples=params.subsample.downsamples,
        name="subsample",
    )(y)

    for i, block in enumerate(params.blocks):
        y = conformer_block(
            depth=block.depth,
            fc_ex_factor=block.fc_ex_factor,
            fc_res_factor=block.fc_res_factor,
            embedding=block.embedding,
            num_heads=block.num_heads,
            kernel_size=block.kernel_size,
            dropout=block.dropout,
            use_bias=block.use_bias,
            name=f"CB{i+1}",
        )(y)

    if params.include_top:
        name = "top"
        y = keras.layers.GlobalAveragePooling1D(name=f"{name}.gap")(y)
        if num_classes is not None:
            y = keras.layers.Dense(num_classes, name=name)(y)
        if params.output_activation:
            y = keras.layers.Activation(params.output_activation)(y)
    # END IF

    return y