Skip to content

unext

U-NeXt

Overview

U-NeXt is a modification of U-Net that utilizes techniques from ResNeXt and EfficientNetV2. During the encoding phase, mbconv blocks are used to efficiently process the input.

Classes:

Functions:

Additions

The U-NeXt architecture has been modified to allow the following:

  • MBConv blocks used in the encoding phase.
  • Squeeze and excitation (SE) blocks added within blocks.

Classes

UNextBlockParams

UNext block parameters

Attributes:

  • filters (int) –

    Number of filters

  • depth (int) –

    Layer depth

  • ddepth (int | None) –

    Layer decoder depth

  • kernel (int | tuple[int, int]) –

    Kernel size

  • pool (int | tuple[int, int]) –

    Pool size

  • strides (int | tuple[int, int]) –

    Stride size

  • skip (bool) –

    Add skip connection

  • expand_ratio (float) –

    Expansion ratio

  • se_ratio (float) –

    Squeeze and excite ratio

  • dropout (float | None) –

    Dropout rate

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

    Normalization type

UNextParams

UNext parameters

Attributes:

UNextModel

Helper class to generate model from parameters

Functions

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

Create layer from parameters

Source code in neuralspot_edge/models/unext.py 
428
429
430
431
432
433
@staticmethod
def layer_from_params(inputs: keras.Input, params: UNextParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = UNextParams(**params)
    return unext_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: UNextParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/unext.py 
435
436
437
438
439
@staticmethod
def model_from_params(inputs: keras.Input, params: UNextParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = UNextModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

se_block 

se_block(ratio: int = 8, name: str | None = None)

Squeeze and excite block

Source code in neuralspot_edge/models/unext.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
def se_block(ratio: int = 8, name: str | None = None):
    """Squeeze and excite block"""

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        num_chan = x.shape[-1]
        # Squeeze
        y = keras.layers.GlobalAveragePooling2D(name=f"{name}.pool" if name else None, keepdims=True)(x)

        y = keras.layers.Conv2D(
            num_chan // ratio,
            kernel_size=1,
            use_bias=True,
            name=f"{name}.sq" if name else None,
        )(y)

        y = keras.layers.Activation("relu6", name=f"{name}.relu" if name else None)(y)

        # Excite
        y = keras.layers.Conv2D(num_chan, kernel_size=1, use_bias=True, name=f"{name}.ex" if name else None)(y)
        y = keras.layers.Activation(keras.activations.hard_sigmoid, name=f"{name}.sigg" if name else None)(y)
        y = keras.layers.Multiply(name=f"{name}.mul" if name else None)([x, y])
        return y

    return layer

norm_layer 

norm_layer(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/unext.py 
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
def norm_layer(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":
            ln_axis = 2 if x.shape[1] == 1 else 1 if x.shape[2] == 1 else (1, 2)
            return keras.layers.LayerNormalization(axis=ln_axis, name=f"{name}.LN")(x)
        return x

    return layer

unext_block 

unext_block(output_filters: int, expand_ratio: float = 1, kernel_size: int | tuple[int, int] = 3, strides: int | tuple[int, int] = 1, se_ratio: float = 4, dropout: float | None = 0, norm: Literal['batch', 'layer'] | None = 'batch', name: str | None = None) -> keras.Layer

Create UNext block

Source code in neuralspot_edge/models/unext.py 
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
def unext_block(
    output_filters: int,
    expand_ratio: float = 1,
    kernel_size: int | tuple[int, int] = 3,
    strides: int | tuple[int, int] = 1,
    se_ratio: float = 4,
    dropout: float | None = 0,
    norm: Literal["batch", "layer"] | None = "batch",
    name: str | None = None,
) -> keras.Layer:
    """Create UNext block"""

    def layer(x: keras.KerasTensor) -> keras.KerasTensor:
        input_filters: int = x.shape[-1]
        strides_len = strides if isinstance(strides, int) else sum(strides) // len(strides)
        add_residual = input_filters == output_filters and strides_len == 1
        ln_axis = 2 if x.shape[1] == 1 else 1 if x.shape[2] == 1 else (1, 2)

        # Depthwise conv
        y = keras.layers.Conv2D(
            input_filters,
            kernel_size=kernel_size,
            groups=input_filters,
            strides=1,
            padding="same",
            use_bias=norm is None,
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name=f"{name}.dwconv" if name else None,
        )(x)
        if norm == "batch":
            y = keras.layers.BatchNormalization(
                name=f"{name}.norm",
            )(y)
        elif norm == "layer":
            y = keras.layers.LayerNormalization(
                axis=ln_axis,
                name=f"{name}.norm" if name else None,
            )(y)
        # END IF

        # Inverted expansion block
        if expand_ratio != 1:
            y = keras.layers.Conv2D(
                filters=int(expand_ratio * input_filters),
                kernel_size=1,
                strides=1,
                padding="same",
                use_bias=norm is None,
                groups=input_filters,
                kernel_initializer="he_normal",
                kernel_regularizer=keras.regularizers.L2(1e-3),
                name=f"{name}.expand" if name else None,
            )(y)

            y = keras.layers.Activation(
                "relu6",
                name=f"{name}.relu" if name else None,
            )(y)

        # Squeeze and excite
        if se_ratio > 1:
            name_se = f"{name}.se" if name else None
            y = se_block(ratio=se_ratio, name=name_se)(y)

        y = keras.layers.Conv2D(
            filters=output_filters,
            kernel_size=1,
            strides=1,
            padding="same",
            use_bias=norm is None,
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name=f"{name}.project" if name else None,
        )(y)

        if add_residual:
            if dropout and dropout > 0:
                y = keras.layers.Dropout(
                    dropout,
                    noise_shape=(y.shape),
                    name=f"{name}.drop" if name else None,
                )(y)
            y = keras.layers.Add(name=f"{name}.res" if name else None)([x, y])
        return y

    # END DEF
    return layer

unext_core 

unext_core(x: keras.KerasTensor, params: UNextParams) -> keras.KerasTensor

Create UNext TF functional core

Parameters:

  • x (KerasTensor) –

    Input tensor

  • params (UNextParams) –

    Model parameters.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/unext.py 
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
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
def unext_core(
    x: keras.KerasTensor,
    params: UNextParams,
) -> keras.KerasTensor:
    """Create UNext TF functional core

    Args:
        x (keras.KerasTensor): Input tensor
        params (UNextParams): Model parameters.

    Returns:
        keras.KerasTensor: Output tensor
    """

    y = x

    #### ENCODER ####
    skip_layers: list[keras.layers.Layer | None] = []
    for i, block in enumerate(params.blocks):
        name = f"ENC{i+1}"
        for d in range(block.depth):
            y = unext_block(
                output_filters=block.filters,
                expand_ratio=block.expand_ratio,
                kernel_size=block.kernel,
                strides=1,
                se_ratio=block.se_ratio,
                dropout=block.dropout,
                norm=block.norm,
                name=f"{name}.D{d+1}",
            )(y)
        # END FOR
        skip_layers.append(y if block.skip else None)

        # Downsample using strided conv
        y = keras.layers.Conv2D(
            filters=block.filters,
            kernel_size=block.pool,
            strides=block.strides,
            padding="same",
            use_bias=block.norm is None,
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name=f"{name}.pool",
        )(y)
        if block.norm == "batch":
            y = keras.layers.BatchNormalization(
                name=f"{name}.norm",
            )(y)
        elif block.norm == "layer":
            ln_axis = 2 if y.shape[1] == 1 else 1 if y.shape[2] == 1 else (1, 2)
            y = keras.layers.LayerNormalization(
                axis=ln_axis,
                name=f"{name}.norm",
            )(y)
        # END IF
    # END FOR

    #### DECODER ####
    for i, block in enumerate(reversed(params.blocks)):
        name = f"DEC{i+1}"
        for d in range(block.ddepth or block.depth):
            y = unext_block(
                output_filters=block.filters,
                expand_ratio=block.expand_ratio,
                kernel_size=block.kernel,
                strides=1,
                se_ratio=block.se_ratio,
                dropout=block.dropout,
                norm=block.norm,
                name=f"{name}.D{d+1}",
            )(y)
        # END FOR

        # Upsample using transposed conv
        # y = keras.layers.Conv1DTranspose(
        #     filters=block.filters,
        #     kernel_size=block.pool,
        #     strides=block.strides,
        #     padding="same",
        #     kernel_initializer="he_normal",
        #     kernel_regularizer=keras.regularizers.L2(1e-3),
        #     name=f"{name}.unpool",
        # )(y)

        y = keras.layers.Conv2D(
            filters=block.filters,
            kernel_size=block.pool,
            strides=1,
            padding="same",
            use_bias=block.norm is None,
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name=f"{name}.conv",
        )(y)
        y = keras.layers.UpSampling2D(size=block.strides, name=f"{name}.unpool")(y)

        # Skip connection
        skip_layer = skip_layers.pop()
        if skip_layer is not None:
            # y = keras.layers.Concatenate(name=f"{name}.S1.cat")([y, skip_layer])
            y = keras.layers.Add(name=f"{name}.S1.cat")([y, skip_layer])

            # Use conv to reduce filters
            y = keras.layers.Conv2D(
                block.filters,
                kernel_size=1,  # block.kernel,
                padding="same",
                kernel_initializer="he_normal",
                kernel_regularizer=keras.regularizers.L2(1e-3),
                use_bias=block.norm is None,
                name=f"{name}.S1.conv",
            )(y)

            if block.norm == "batch":
                y = keras.layers.BatchNormalization(
                    name=f"{name}.S1.norm",
                )(y)
            elif block.norm == "layer":
                ln_axis = 2 if y.shape[1] == 1 else 1 if y.shape[2] == 1 else (1, 2)
                y = keras.layers.LayerNormalization(
                    axis=ln_axis,
                    name=f"{name}.S1.norm",
                )(y)
            # END IF

            y = keras.layers.Activation(
                "relu6",
                name=f"{name}.S1.relu" if name else None,
            )(y)
        # END IF

        y = unext_block(
            output_filters=block.filters,
            expand_ratio=block.expand_ratio,
            kernel_size=block.kernel,
            strides=1,
            se_ratio=block.se_ratio,
            dropout=block.dropout,
            norm=block.norm,
            name=f"{name}.D{block.depth+1}",
        )(y)

    # END FOR
    return y

unext_layer 

unext_layer(inputs: keras.KerasTensor, params: UNextParams, num_classes: int | None = None) -> keras.KerasTensor

Create UNext TF functional model

Parameters:

  • inputs (KerasTensor) –

    Input tensor

  • params (UNextParams) –

    Model parameters.

  • num_classes (int, default: None ) –

    Number of classes.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/unext.py 
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
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
def unext_layer(
    inputs: keras.KerasTensor,
    params: UNextParams,
    num_classes: int | None = None,
) -> keras.KerasTensor:
    """Create UNext TF functional model

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

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

    y = unext_core(y, params)

    if params.include_top:
        # Add a per-point classification layer
        y = keras.layers.Conv2D(
            num_classes,
            kernel_size=params.output_kernel_size,
            padding="same",
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name="NECK.conv",
            use_bias=True,
        )(y)
        if 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