Skip to content

unet

U-Net

Overview

U-Net is a type of convolutional neural network (CNN) that is commonly used for segmentation tasks. U-Net is a fully convolutional network that consists of a series of convolutional layers and pooling layers. The pooling layers are used to downsample the input while the convolutional layers are used to upsample the input. The skip connections between the pooling layers and convolutional layers allow U-Net to preserve spatial/temporal information while also allowing for faster training and inference times.

For more info, refer to the original paper U-Net: Convolutional Networks for Biomedical Image Segmentation.

Classes:

Functions:

Additions

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

  • Enable 1D and 2D variants.
  • Convolutional pairs can factorized into depthwise separable convolutions.
  • Specifiy the number of convolutional layers per block both downstream and upstream.
  • Normalization can be set between batch normalization and layer normalization.
  • ReLU is replaced with the approximated ReLU6.

Usage

Instantiate from UNetParams:

import keras
from neuralspot_edge.models import UNet, UNetParams, UNetBlockParams

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

model = UNet(
    x=inputs,
    params=UNetParams(
        blocks=[
            UNetBlockParams(filters=12, depth=2, ddepth=1, kernel=(1, 5), pool=(1, 3), strides=(1, 2), skip=True, seperable=True),
            UNetBlockParams(filters=24, depth=2, ddepth=1, kernel=(1, 5), pool=(1, 3), strides=(1, 2), skip=True, seperable=True),
            UNetBlockParams(filters=32, depth=2, ddepth=1, kernel=(1, 5), pool=(1, 3), strides=(1, 2), skip=True, seperable=True),
            UNetBlockParams(filters=48, depth=2, ddepth=1, kernel=(1, 5), pool=(1, 3), strides=(1, 2), skip=True, seperable=True)
        ],
        output_kernel_size=(1, 5),
        include_top=True,
        use_logits=True,
        model_name="unet"
    ),
    num_classes=num_classes,
)

Instantiate from object:

params = {
    "name": "unet",
    "params": {
        "blocks": [
            {"filters": 12, "depth": 2, "ddepth": 1, "kernel": [1, 5], "pool": [1, 3], "strides": [1, 2], "skip": true, "seperable": true},
            {"filters": 24, "depth": 2, "ddepth": 1, "kernel": [1, 5], "pool": [1, 3], "strides": [1, 2], "skip": true, "seperable": true},
            {"filters": 32, "depth": 2, "ddepth": 1, "kernel": [1, 5], "pool": [1, 3], "strides": [1, 2], "skip": true, "seperable": true},
            {"filters": 48, "depth": 2, "ddepth": 1, "kernel": [1, 5], "pool": [1, 3], "strides": [1, 2], "skip": true, "seperable": true}
        ],
        "output_kernel_size": [1, 5],
        "include_top": true,
        "use_logits": true,
        "model_name": "efficientnetv2"
    }
}

model = unet_from_object(inputs, params, num_classes)

Classes

UNetBlockParams

UNet block parameters

Attributes:

  • filters (int) –

    Number of filters

  • depth (int) –

    Layer depth

  • ddepth (int | None) –

    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

  • seperable (bool) –

    Use seperable convs

  • dropout (float | None) –

    Dropout rate

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

    Normalization type

  • activation (Literal['relu', 'relu6', 'leaky_relu', 'elu', 'selu']) –

    Activation

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

    Dilation factor

UNetParams

UNet parameters

Attributes:

UNetModel

Helper class to generate model from parameters

Functions

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

Create layer from parameters

Source code in neuralspot_edge/models/unet.py 
363
364
365
366
367
368
@staticmethod
def layer_from_params(inputs: keras.Input, params: UNetParams | dict, num_classes: int | None = None):
    """Create layer from parameters"""
    if isinstance(params, dict):
        params = UNetParams(**params)
    return unet_layer(x=inputs, params=params, num_classes=num_classes)
model_from_params staticmethod 
model_from_params(inputs: keras.Input, params: UNetParams | dict, num_classes: int | None = None)

Create model from parameters

Source code in neuralspot_edge/models/unet.py 
370
371
372
373
374
@staticmethod
def model_from_params(inputs: keras.Input, params: UNetParams | dict, num_classes: int | None = None):
    """Create model from parameters"""
    outputs = UNetModel.layer_from_params(inputs=inputs, params=params, num_classes=num_classes)
    return keras.Model(inputs=inputs, outputs=outputs)

Functions

unet_layer 

unet_layer(x: keras.KerasTensor, params: UNetParams, num_classes: int) -> keras.KerasTensor

Create UNet TF functional model

Parameters:

  • x (KerasTensor) –

    Input tensor

  • params (ResNetParams) –

    Model parameters.

  • num_classes (int) –

    Number of classes.

Returns:

  • KerasTensor –

    keras.KerasTensor: Output tensor

Source code in neuralspot_edge/models/unet.py 
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
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
def unet_layer(
    x: keras.KerasTensor,
    params: UNetParams,
    num_classes: int,
) -> keras.KerasTensor:
    """Create UNet TF functional model

    Args:
        x (keras.KerasTensor): Input tensor
        params (ResNetParams): Model parameters.
        num_classes (int, optional): 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
    # END IF

    #### ENCODER ####
    skip_layers: list[keras.layers.Layer | None] = []
    for i, block in enumerate(params.blocks):
        name = f"ENC{i+1}"
        ym = y
        for d in range(block.depth):
            dname = f"{name}.D{d+1}"
            if block.dilation is None:
                dilation_rate = (1, 1)
            elif isinstance(block.dilation, int):
                dilation_rate = (block.dilation**d, block.dilation**d)
            else:
                dilation_rate = (block.dilation[0] ** d, block.dilation[1] ** d)
            if block.seperable:
                ym = keras.layers.SeparableConv2D(
                    block.filters,
                    kernel_size=block.kernel,
                    strides=(1, 1),
                    padding="same",
                    dilation_rate=dilation_rate,
                    depthwise_initializer="he_normal",
                    pointwise_initializer="he_normal",
                    depthwise_regularizer=keras.regularizers.L2(1e-3),
                    pointwise_regularizer=keras.regularizers.L2(1e-3),
                    use_bias=block.norm is None,
                    name=f"{dname}.conv",
                )(ym)
            else:
                ym = keras.layers.Conv2D(
                    block.filters,
                    kernel_size=block.kernel,
                    strides=(1, 1),
                    padding="same",
                    dilation_rate=dilation_rate,
                    kernel_initializer="he_normal",
                    kernel_regularizer=keras.regularizers.L2(1e-3),
                    use_bias=block.norm is None,
                    name=f"{dname}.conv",
                )(ym)
            if block.norm == "layer":
                ym = layer_normalization(name=dname, axis=[1, 2])(ym)
            elif block.norm == "batch":
                ym = batch_normalization(name=dname, momentum=0.99)(ym)
            ym = keras.layers.Activation(block.activation, name=f"{dname}.act")(ym)
        # END FOR

        # Project residual
        yr = keras.layers.Conv2D(
            block.filters,
            kernel_size=(1, 1),
            strides=(1, 1),
            padding="same",
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name=f"{name}.skip",
        )(y)

        if block.dropout is not None:
            ym = keras.layers.Dropout(block.dropout, noise_shape=ym.shape)(ym)
        y = keras.layers.add([ym, yr], name=f"{name}.add")

        skip_layers.append(y if block.skip else None)

        y = keras.layers.MaxPooling2D(block.pool, strides=block.strides, padding="same", name=f"{name}.pool")(y)
    # 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):
            dname = f"{name}.D{d+1}"
            if block.seperable:
                y = keras.layers.SeparableConv2D(
                    block.filters,
                    kernel_size=block.kernel,
                    strides=(1, 1),
                    padding="same",
                    dilation_rate=dilation_rate,
                    depthwise_initializer="he_normal",
                    pointwise_initializer="he_normal",
                    depthwise_regularizer=keras.regularizers.L2(1e-3),
                    pointwise_regularizer=keras.regularizers.L2(1e-3),
                    use_bias=block.norm is None,
                    name=f"{dname}.conv",
                )(y)
            else:
                y = keras.layers.Conv2D(
                    block.filters,
                    kernel_size=block.kernel,
                    strides=(1, 1),
                    padding="same",
                    dilation_rate=dilation_rate,
                    kernel_initializer="he_normal",
                    kernel_regularizer=keras.regularizers.L2(1e-3),
                    use_bias=block.norm is None,
                    name=f"{dname}.conv",
                )(y)
            if block.norm == "layer":
                y = layer_normalization(name=dname, axis=[1, 2])(y)
            elif block.norm == "batch":
                y = batch_normalization(name=dname, momentum=0.99)(y)
            y = keras.layers.Activation(block.activation, name=f"{dname}.act")(y)
        # END FOR

        y = keras.layers.UpSampling2D(size=block.strides, name=f"{dname}.unpool")(y)

        # Add skip connection
        dname = f"{name}.D{block.depth+1}"
        skip_layer = skip_layers.pop()
        if skip_layer is not None:
            y = keras.layers.concatenate([y, skip_layer], name=f"{dname}.cat")  # Can add or concatenate
            # Use 1x1 conv to reduce filters
            y = keras.layers.Conv2D(
                block.filters,
                kernel_size=(1, 1),
                padding="same",
                kernel_initializer="he_normal",
                kernel_regularizer=keras.regularizers.L2(1e-3),
                use_bias=block.norm is None,
                name=f"{dname}.conv",
            )(y)
            if block.norm == "layer":
                y = layer_normalization(name=dname, axis=[1, 2])(y)
            elif block.norm == "batch":
                y = batch_normalization(name=dname, momentum=0.99)(y)
            y = keras.layers.Activation(block.activation, name=f"{dname}.act")(y)
        # END IF

        dname = f"{name}.D{block.depth+2}"
        if block.seperable:
            ym = keras.layers.SeparableConv2D(
                block.filters,
                kernel_size=block.kernel,
                strides=(1, 1),
                padding="same",
                depthwise_initializer="he_normal",
                pointwise_initializer="he_normal",
                depthwise_regularizer=keras.regularizers.L2(1e-3),
                pointwise_regularizer=keras.regularizers.L2(1e-3),
                use_bias=block.norm is None,
                name=f"{dname}.conv",
            )(y)
        else:
            ym = keras.layers.Conv2D(
                block.filters,
                kernel_size=block.kernel,
                strides=(1, 1),
                padding="same",
                kernel_initializer="he_normal",
                kernel_regularizer=keras.regularizers.L2(1e-3),
                use_bias=block.norm is None,
                name=f"{dname}.conv",
            )(y)
        if block.norm == "layer":
            ym = layer_normalization(name=dname, axis=[1, 2])(ym)
        elif block.norm == "batch":
            ym = batch_normalization(name=dname, momentum=0.99)(ym)
        ym = keras.layers.Activation(block.activation, name=f"{dname}.act")(ym)

        # Project residual
        yr = keras.layers.Conv2D(
            block.filters,
            kernel_size=(1, 1),
            padding="same",
            kernel_initializer="he_normal",
            kernel_regularizer=keras.regularizers.L2(1e-3),
            name=f"{name}.skip",
        )(y)
        y = keras.layers.add([ym, yr], name=f"{name}.add")  # Add back residual
    # END FOR

    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

    if requires_reshape:
        y = keras.layers.Reshape(y.shape[2:])(y)
    # END IF

    return y