Skip to content

tcn

Temporal Convolutional Network (TCN)

Overview

Temporal convolutional network (TCN) is a type of convolutional neural network (CNN) that is commonly used for sequence modeling tasks such as speech recognition, text generation, and video classification. TCN is a fully convolutional network that consists of a series of dilated causal convolutional layers. The dilated convolutional layers allow TCN to have a large receptive field while maintaining a small number of parameters. TCN is also fully parallelizable, which allows for faster training and inference times.

For more info, refer to the original paper Temporal Convolutional Networks: A Unified Approach to Action Segmentation.

Classes:

Functions:

Additions

The TCN architecture has been modified to allow the following:

  • Convolutional pairs can be factorized into depthwise separable convolutions.
  • Squeeze and excitation (SE) blocks can be added between convolutional pairs.
  • Normalization can be set between batch normalization and layer normalization.
  • ReLU is replaced with the approximated ReLU6.

Usage

The following example demonstrates how to create a TCN model using the Tcn class. The model is defined using a set of parameters defined in the TcnParams and TcnBlockParams classes.

import keras
import neuralspot_edge as nse

inputs = keras.Input(shape=(800, 1), name="inputs")
num_classes = 5

params = nse.models.TcnParams(
    input_kernel=(1, 3),
    input_norm="batch",
    blocks=[
        nse.models.TcnBlockParams(filters=8, kernel=(1, 3), dilation=(1, 1), dropout=0.1, ex_ratio=1, se_ratio=0, norm="batch"),
        nse.models.TcnBlockParams(filters=16, kernel=(1, 3), dilation=(1, 2), dropout=0.1, ex_ratio=1, se_ratio=0, norm="batch"),
        nse.models.TcnBlockParams(filters=24, kernel=(1, 3), dilation=(1, 4), dropout=0.1, ex_ratio=1, se_ratio=4, norm="batch"),
        nse.models.TcnBlockParams(filters=32, kernel=(1, 3), dilation=(1, 8), dropout=0.1, ex_ratio=1, se_ratio=4, norm="batch"),
    ],
    output_kernel=(1, 3),
    include_top=True,
    use_logits=True,
    model_name="tcn",
)

model = nse.models.TcnModel.model_from_params(
    inputs=inputs,
    params=params,
    num_classes=num_classes,
)

Classes

TcnBlockParams

TCN block parameters

Attributes:

  • depth (int) –

    Layer depth

  • branch (int) –

    Number of branches

  • filters (int) –

    Number of filters

  • kernel (int | tuple[int, int]) –

    Kernel size

  • dilation (int | tuple[int, int]) –

    Dilation rate

  • ex_ratio (float) –

    Expansion ratio

  • se_ratio (float) –

    Squeeze and excite ratio

  • dropout (float | None) –

    Dropout rate

  • norm (Literal['batch', 'layer'] | None) –

    Normalization type

  • activation (str) –

    Activation function

TcnParams

TCN parameters

Attributes:

  • input_kernel (int | tuple[int, int] | None) –

    Input kernel size

  • input_norm (Literal['batch', 'layer'] | None) –

    Input normalization type

  • block_type (Literal['lg', 'mb', 'sm']) –

    Block type

  • blocks (list[TcnBlockParams]) –

    TCN blocks

  • output_kernel (int | tuple[int, int]) –

    Output kernel size

  • include_top (bool) –

    Include top

  • use_logits (bool) –

    Use logits

  • output_activation (str | None) –

    Output activation

  • name (str) –

    Model name

TcnModel

Helper class to generate model from parameters

Functions

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

Create layer from parameters

Source code in neuralspot_edge/models/tcn.py 
502
503
504
505
506
507
@staticmethod
def layer_from_params(inputs: keras.Input, params: TcnParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = TcnParams(**params)
    return tcn_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: TcnParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/tcn.py 
509
510
511
512
513
@staticmethod
def model_from_params(inputs: keras.Input, params: TcnParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = TcnModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

normalization 

normalization(norm: str, name: str) -> keras.Layer

Normalization layer

Parameters:

  • norm (str) –

    Normalization type

  • name (str) –

    Name

Returns:

  • Layer –

    keras.Layer: Layer

Source code in neuralspot_edge/models/tcn.py 
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
def normalization(norm: str, name: str) -> keras.Layer:
    """Normalization layer

    Args:
        norm (str): Normalization type
        name (str): Name

    Returns:
        keras.Layer: Layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        """Functional normalization layer

        Args:
            x (keras.KerasTensor): Input tensor

        Returns:
            keras.KerasTensor: Output tensor
        """
        if norm == "batch":
            return keras.layers.BatchNormalization(axis=-1, name=f"{name}.BN")(x)
        if norm == "layer":
            return keras.layers.LayerNormalization(axis=(1, 2), name=f"{name}.LN")(x)
        return x

    return layer

tcn_block_lg 

tcn_block_lg(params: TcnBlockParams, name: str) -> keras.Layer

TCN large block

Parameters:

Returns:

  • Layer –

    keras.Layer: Layer

Source code in neuralspot_edge/models/tcn.py 
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
def tcn_block_lg(params: TcnBlockParams, name: str) -> keras.Layer:
    """TCN large block

    Args:
        params (TcnBlockParams): Parameters
        name (str): Name

    Returns:
        keras.Layer: Layer
    """

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

        for d in range(params.depth):
            lcl_name = f"{name}.D{d+1}"
            y_skip = y

            y = keras.layers.Conv2D(
                filters=params.filters,
                kernel_size=params.kernel,
                strides=(1, 1),
                padding="same",
                use_bias=False,
                dilation_rate=params.dilation,
                kernel_initializer="he_normal",
                kernel_regularizer=keras.regularizers.L2(1e-3),
                name=f"{lcl_name}.CN1",
            )(y)
            y = normalization(params.norm, f"{lcl_name}.CN1")(y)

            y = keras.layers.Conv2D(
                filters=params.filters,
                kernel_size=params.kernel,
                strides=(1, 1),
                padding="same",
                use_bias=params.norm is None,
                dilation_rate=params.dilation,
                kernel_initializer="he_normal",
                kernel_regularizer=keras.regularizers.L2(1e-3),
                name=f"{lcl_name}.CN2",
            )(y)
            y = normalization(params.norm, f"{lcl_name}.CN2")(y)

            if y_skip.shape[-1] == y.shape[-1]:
                y = keras.layers.Add(name=f"{lcl_name}.ADD")([y, y_skip])

            y = keras.layers.Activation(params.activation, name=f"{lcl_name}.ACT")(y)

            # Squeeze and excite
            if params.se_ratio > 0:
                y = se_layer(ratio=params.se_ratio, name=f"{lcl_name}.SE")(y)
            # END IF

            if params.dropout and params.dropout > 0:
                y = keras.layers.SpatialDropout2D(rate=params.dropout, name=f"{lcl_name}.DROP")(y)
            # END IF

        # END FOR
        return y

    return layer

tcn_block_mb 

tcn_block_mb(params: TcnBlockParams, name: str) -> keras.Layer

TCN mbconv block

Parameters:

Returns:

  • Layer –

    keras.Layer: Layer

Source code in neuralspot_edge/models/tcn.py 
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
def tcn_block_mb(params: TcnBlockParams, name: str) -> keras.Layer:
    """TCN mbconv block

    Args:
        params (TcnBlockParams): Parameters
        name (str): Name

    Returns:
        keras.Layer: Layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        """TCN block layer"""
        y = x
        y_skip = y
        for d in range(params.depth):
            lcl_name = f"{name}.D{d+1}"

            if params.ex_ratio != 1:
                y = keras.layers.Conv2D(
                    filters=int(params.filters * params.ex_ratio),
                    kernel_size=(1, 1),
                    strides=(1, 1),
                    padding="same",
                    use_bias=params.norm is None,
                    kernel_initializer="he_normal",
                    kernel_regularizer=keras.regularizers.L2(1e-3),
                    name=f"{lcl_name}.EX.CN",
                )(y)
                y = normalization(params.norm, f"{lcl_name}.EX")(y)
                y = keras.layers.Activation(params.activation, name=f"{lcl_name}.EX.ACT")(y)
            # END IF

            branches = []
            for b in range(params.branch):
                yb = y
                yb = keras.layers.DepthwiseConv2D(
                    kernel_size=params.kernel,
                    strides=(1, 1),
                    padding="same",
                    use_bias=params.norm is None,
                    dilation_rate=params.dilation,
                    depthwise_initializer="he_normal",
                    depthwise_regularizer=keras.regularizers.L2(1e-3),
                    name=f"{lcl_name}.DW.B{b+1}.CN",
                )(yb)
                yb = normalization(params.norm, f"{lcl_name}.DW.B{b+1}")(yb)
                branches.append(yb)
            # END FOR

            if params.branch > 1:
                y = keras.layers.Add(name=f"{lcl_name}.DW.ADD")(branches)
            else:
                y = branches[0]
            # END IF

            y = keras.layers.Activation(params.activation, name=f"{lcl_name}.DW.ACT")(y)

            # Squeeze and excite
            if params.se_ratio and y.shape[-1] // params.se_ratio > 0:
                y = se_layer(ratio=params.se_ratio, name=f"{lcl_name}.SE")(y)
            # END IF

            branches = []
            for b in range(params.branch):
                yb = y
                yb = keras.layers.Conv2D(
                    filters=params.filters,
                    kernel_size=(1, 1),
                    strides=(1, 1),
                    padding="same",
                    use_bias=params.norm is None,
                    kernel_initializer="he_normal",
                    kernel_regularizer=keras.regularizers.L2(1e-3),
                    name=f"{lcl_name}.PW.B{b+1}.CN",
                )(yb)
                yb = normalization(params.norm, f"{lcl_name}.PW.B{b+1}")(yb)
                branches.append(yb)
            # END FOR

            if params.branch > 1:
                y = keras.layers.Add(name=f"{lcl_name}.PW.ADD")(branches)
            else:
                y = branches[0]
            # END IF

            y = keras.layers.Activation(params.activation, name=f"{lcl_name}.PW.ACT")(y)
        # END FOR

        # Skip connection
        if y_skip.shape[-1] == y.shape[-1]:
            y = keras.layers.Add(name=f"{name}.ADD")([y, y_skip])

        if params.dropout and params.dropout > 0:
            y = keras.layers.SpatialDropout2D(rate=params.dropout, name=f"{name}.DROP")(y)
        # END IF
        return y

    return layer

tcn_block_sm 

tcn_block_sm(params: TcnBlockParams, name: str) -> keras.Layer

TCN small block

Parameters:

Returns:

  • Layer –

    keras.Layer: Layer

Source code in neuralspot_edge/models/tcn.py 
324
325
326
327
328
329
330
331
332
333
334
335
336
337
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 tcn_block_sm(params: TcnBlockParams, name: str) -> keras.Layer:
    """TCN small block

    Args:
        params (TcnBlockParams): Parameters
        name (str): Name

    Returns:
        keras.Layer: Layer
    """

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        """TCN block layer"""
        y = x
        y_skip = y
        for d in range(params.depth):
            lcl_name = f"{name}.D{d+1}"
            branches = []
            for b in range(params.branch):
                yb = y
                yb = keras.layers.DepthwiseConv2D(
                    kernel_size=params.kernel,
                    strides=(1, 1),
                    padding="same",
                    use_bias=params.norm is None,
                    dilation_rate=params.dilation,
                    depthwise_initializer="he_normal",
                    depthwise_regularizer=keras.regularizers.L2(1e-3),
                    name=f"{lcl_name}.DW.B{b+1}.CN",
                )(yb)
                yb = normalization(params.norm, f"{lcl_name}.DW.B{b+1}")(yb)
                branches.append(yb)
            # END FOR

            if params.branch > 1:
                y = keras.layers.Add(name=f"{lcl_name}.DW.ADD")(branches)
            else:
                y = branches[0]
            # END IF

            y = keras.layers.Activation(params.activation, name=f"{lcl_name}.DW.ACT")(y)

            branches = []
            for b in range(params.branch):
                yb = y
                yb = keras.layers.Conv2D(
                    filters=params.filters,
                    kernel_size=(1, 1),
                    strides=(1, 1),
                    padding="same",
                    # groups=int(params.se_ratio) if params.se_ratio > 0 else 1,
                    use_bias=params.norm is None,
                    kernel_initializer="he_normal",
                    kernel_regularizer=keras.regularizers.L2(1e-3),
                    name=f"{lcl_name}.PW.B{b+1}.CN",
                )(yb)
                yb = normalization(params.norm, f"{lcl_name}.PW.B{b+1}")(yb)
                branches.append(yb)
            # END FOR

            if params.branch > 1:
                y = keras.layers.Add(name=f"{lcl_name}.PW.ADD")(branches)
            else:
                y = branches[0]
            # END IF

            y = keras.layers.Activation(params.activation, name=f"{lcl_name}.PW.ACT")(y)
        # END FOR

        # Squeeze and excite
        if y.shape[-1] // params.se_ratio > 1:
            y = se_layer(ratio=params.se_ratio, name=f"{name}.SE")(y)
        # END IF

        # Skip connection
        if y_skip.shape[-1] == y.shape[-1]:
            y = keras.layers.Add(name=f"{name}.ADD")([y, y_skip])

        if params.dropout and params.dropout > 0:
            y = keras.layers.SpatialDropout2D(rate=params.dropout, name=f"{name}.DROP")(y)
        # END IF
        return y

    return layer

tcn_core 

tcn_core(params: TcnParams) -> keras.Layer

TCN core

Parameters:

Returns:

  • Layer –

    keras.Layer: Layer

Source code in neuralspot_edge/models/tcn.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
def tcn_core(params: TcnParams) -> keras.Layer:
    """TCN core

    Args:
        params (TcnParams): Parameters

    Returns:
        keras.Layer: Layer
    """
    if params.block_type == "lg":
        tcn_block = tcn_block_lg
    elif params.block_type == "mb":
        tcn_block = tcn_block_mb
    elif params.block_type == "sm":
        tcn_block = tcn_block_sm
    else:
        raise ValueError(f"Invalid block type: {params.block_type}")

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        y = x
        for i, block in enumerate(params.blocks):
            name = f"B{i+1}"
            y = tcn_block(params=block, name=name)(y)
        # END IF
        return y

    return layer

tcn_layer 

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

TCN functional layer

Parameters:

  • x (KerasTensor) –

    Input tensor

  • params (TcnParams) –

    Parameters

  • num_classes (int, default: None ) –

    Number of classes

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/tcn.py 
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
def tcn_layer(
    x: keras.KerasTensor,
    params: TcnParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """TCN functional layer

    Args:
        x (keras.KerasTensor): Input tensor
        params (TcnParams): Parameters
        num_classes (int): Number of classes

    Returns:
        keras.KerasTensor: Output tensor
    """
    requires_reshape = len(x.shape) == 3
    if requires_reshape:
        y = keras.layers.Reshape((1,) + x.shape[1:])(x)
    else:
        y = x

    ### Input layer
    # Encode each channel separately
    if params.input_kernel:
        y = keras.layers.DepthwiseConv2D(
            kernel_size=params.input_kernel,
            use_bias=params.input_norm is None,
            name="ENC.CN",
            padding="same",
        )(y)
        y = normalization(params.input_norm, "ENC")(y)
    # END IF

    ### Core layers
    y = tcn_core(params)(y)

    ### Output layer
    if params.include_top:
        # Add a per-point classification layer
        y = keras.layers.Conv2D(
            num_classes,
            kernel_size=params.output_kernel,
            padding="same",
            name="NECK.conv",
            use_bias=True,
        )(y)
        if params.output_activation:
            y = keras.layers.Activation(params.output_activation)(y)
        elif not params.use_logits:
            y = keras.layers.Softmax()(y)
        # END IF
    # END IF

    # Always reshape back to original shape
    if requires_reshape:
        y = keras.layers.Reshape(y.shape[2:])(y)

    return y