def build(dim): #dim is the square dimensions of the input image
model = models.Sequential()
#convolution layer with 4 3x3 kernels, passed the input image
model.add(
layers.Conv2D(4, (3, 3),
activation='relu',
input_shape=(dim, dim, 1)))
model.add(layers.AvgPool2D(
(2, 2))) #perform average pooling 2x2 with a stride of 2 (default)
#second convolution layer with 4 3x3 kernels
model.add(layers.Conv2D(4, (3, 3), activation='relu'))
model.add(layers.AvgPool2D((2, 2)))
#Fully Connected or Densely Connected Classifier Network
model.add(layers.Flatten()) #restructure data for FC network
model.add(layers.Dropout(0.5)) #apply dropout regularization
model.add(layers.Dense(
8, activation='relu')) #Dense layer with 8 fully connected neurons
#Output layer with softmax classifier to for multiclass-single label problem
model.add(layers.Dense(4, activation='softmax'))
model.summary() #print the network parameters
num_layers = 4 #parameter used for plotting activation maps in Visualization
return model, num_layers
def unet(t_x):
input_img = layers.Input(t_x.shape[1:], name='RGB_Input')
pp_in_layer = input_img
if NET_SCALING is not None:
pp_in_layer = layers.AvgPool2D(NET_SCALING)(pp_in_layer)
#pp_in_layer = layers.GaussianNoise(GAUSSIAN_NOISE)(pp_in_layer)
pp_in_layer = layers.BatchNormalization()(pp_in_layer)
c1 = layers.Conv2D(8, (3, 3), activation='relu',
padding='same')(pp_in_layer)
c1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same')(c1)
p1 = layers.MaxPooling2D((2, 2))(c1)
c2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(p1)
c2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(c2)
p2 = layers.MaxPooling2D((2, 2))(c2)
c3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(p2)
c3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(c3)
p3 = layers.MaxPooling2D((2, 2))(c3)
c4 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(p3)
c4 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(c4)
p4 = layers.MaxPooling2D(pool_size=(2, 2))(c4)
c5 = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(p4)
c5 = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(c5)
u6 = upsample(64, (2, 2), strides=(2, 2), padding='same')(c5)
u6 = layers.concatenate([u6, c4])
c6 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(u6)
c6 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(c6)
u7 = upsample(32, (2, 2), strides=(2, 2), padding='same')(c6)
u7 = layers.concatenate([u7, c3])
c7 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(u7)
c7 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(c7)
u8 = upsample(16, (2, 2), strides=(2, 2), padding='same')(c7)
u8 = layers.concatenate([u8, c2])
c8 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(u8)
c8 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(c8)
u9 = upsample(8, (2, 2), strides=(2, 2), padding='same')(c8)
u9 = layers.concatenate([u9, c1])
c9 = layers.Conv2D(8, (3, 3), activation='relu', padding='same')(u9)
c9 = layers.Conv2D(8, (3, 3), activation='relu', padding='same')(c9)
d = layers.Conv2D(1, (1, 1), activation='sigmoid')(c9)
#d = layers.Cropping2D((EDGE_CROP, EDGE_CROP))(d)
#d = layers.ZeroPadding2D((EDGE_CROP, EDGE_CROP))(d)
if NET_SCALING is not None:
d = layers.UpSamplling2D(NET_SCALING)(d)
segmentation_model = models.Model(input=[input_img], outputs=[d])
segmentation_model.summary()
return segmentation_model
def darknet(channels,
odd_pointwise,
avg_pool_size,
cls_activ,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
DarkNet model from 'Darknet: Open source neural networks in c,' https://github.com/pjreddie/darknet.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
odd_pointwise : bool
Whether pointwise convolution layer is used for each odd unit.
avg_pool_size : int
Window size of the final average pooling.
cls_activ : bool
Whether classification convolution layer uses an activation.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = input
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
x = dark_convYxY(x=x,
in_channels=in_channels,
out_channels=out_channels,
pointwise=(len(channels_per_stage) > 1)
and not (((j + 1) % 2 == 1) ^ odd_pointwise),
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
if i != len(channels) - 1:
x = nn.MaxPool2D(pool_size=2,
strides=2,
name="features/pool{}".format(i + 1))(x)
x = nn.Conv2D(filters=classes, kernel_size=1, name="output/final_conv")(x)
if cls_activ:
x = nn.LeakyReLU(alpha=0.1, name="output/final_activ")(x)
x = nn.AvgPool2D(pool_size=avg_pool_size,
strides=1,
name="output/final_pool")(x)
x = nn.Flatten()(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def menet(channels,
init_block_channels,
side_channels,
groups,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
ShuffleNet model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,'
https://arxiv.org/abs/1707.01083.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
side_channels : int
Number of side channels in a ME-unit.
groups : int
Number of groups in convolution layers.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = me_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
downsample = (j == 0)
ignore_group = (i == 0) and (j == 0)
x = me_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
side_channels=side_channels,
groups=groups,
downsample=downsample,
ignore_group=ignore_group,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def senet(channels,
init_block_channels,
cardinality,
bottleneck_width,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
SENet model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
cardinality: int
Number of groups.
bottleneck_width: int
Width of bottleneck block.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, in_size[0], in_size[1]) if is_channels_first() else\
(in_size[0], in_size[1], in_channels)
input = nn.Input(shape=input_shape)
x = senet_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
identity_conv3x3 = (i != 0)
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = senet_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
cardinality=cardinality,
bottleneck_width=bottleneck_width,
identity_conv3x3=identity_conv3x3,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
# x = nn.Flatten()(x)
x = flatten(x)
x = nn.Dropout(rate=0.2, name="output/dropout")(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output/fc")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def transition_block(x, in_channels, out_channels, name="transition_block"):
"""
DenseNet's auxiliary block, which can be treated as the initial part of the DenseNet unit, triggered only in the
first unit of each stage.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
name : str, default 'transition_block'
Unit name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor.
"""
x = pre_conv1x1_block(x=x,
in_channels=in_channels,
out_channels=out_channels,
name=name + "/conv")
x = nn.AvgPool2D(pool_size=2,
strides=2,
padding="valid",
name=name + "/pool")(x)
return x
def preresnet(channels,
init_block_channels,
bottleneck,
conv1_stride,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
PreResNet model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
bottleneck : bool
Whether to use a bottleneck or simple block in units.
conv1_stride : bool
Whether to use stride in the first or the second convolution layer in units.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == "channels_first" else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = preres_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = preres_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
bottleneck=bottleneck,
conv1_stride=conv1_stride,
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
x = preres_activation(x=x, name="features/post_activ")
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def densenet(channels,
init_block_channels,
dropout_rate=0.0,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
DenseNet model from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
dropout_rate : float, default 0.0
Parameter of Dropout layer. Faction of the input units to drop.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, in_size[0], in_size[1]) if is_channels_first() else\
(in_size[0], in_size[1], in_channels)
input = nn.Input(shape=input_shape)
x = preres_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
if i != 0:
x = transition_block(x=x,
in_channels=in_channels,
out_channels=(in_channels // 2),
name="features/stage{}/trans{}".format(
i + 1, i + 1))
in_channels = in_channels // 2
for j, out_channels in enumerate(channels_per_stage):
x = dense_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
dropout_rate=dropout_rate,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = preres_activation(x=x, name="features/post_activ")
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
# x = nn.Flatten()(x)
x = flatten(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def mobilenet(channels,
first_stage_stride,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
MobileNet model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
https://arxiv.org/abs/1704.04861. Also this class implements FD-MobileNet from 'FD-MobileNet: Improved MobileNet
with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
first_stage_stride : bool
Whether stride is used at the first stage.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224, 224) if K.image_data_format() == 'channels_first' else (224, 224, in_channels)
input = nn.Input(shape=input_shape)
init_block_channels = channels[0][0]
x = conv3x3_block(
x=input,
in_channels=in_channels,
out_channels=init_block_channels,
strides=2,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels[1:]):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and ((i != 0) or first_stage_stride) else 1
x = dws_conv_block(
x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(
pool_size=7,
strides=1,
name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(
units=classes,
input_dim=in_channels,
name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def aux_clf(in_tensor):
avg_pool = layers.AvgPool2D(5, 3)(in_tensor)
conv = conv1x1(128)(avg_pool)
flattened = layers.Flatten()(conv)
dense = layers.Dense(1024, activation='relu')(flattened)
dropout = layers.Dropout(0.7)(dense)
out = layers.Dense(1000, activation='softmax')(dropout)
return out
def se_block(x,
channels,
reduction=16,
approx_sigmoid=False,
round_mid=False,
activation="relu",
name="se_block"):
"""
Squeeze-and-Excitation block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
channels : int
Number of channels.
reduction : int, default 16
Squeeze reduction value.
approx_sigmoid : bool, default False
Whether to use approximated sigmoid function.
round_mid : bool, default False
Whether to round middle channel number (make divisible by 8).
activation : function or str, default 'relu'
Activation function or name of activation function.
name : str, default 'se_block'
Block name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
assert(len(x._keras_shape) == 4)
mid_channels = channels // reduction if not round_mid else round_channels(float(channels) / reduction)
pool_size = x._keras_shape[2:4] if is_channels_first() else x._keras_shape[1:3]
w = nn.AvgPool2D(
pool_size=pool_size,
name=name + "/pool")(x)
w = conv1x1(
x=w,
in_channels=channels,
out_channels=mid_channels,
use_bias=True,
name=name + "/conv1")
w = get_activation_layer(
x=w,
activation=activation,
name=name + "/activ")
w = conv1x1(
x=w,
in_channels=mid_channels,
out_channels=channels,
use_bias=True,
name=name + "/conv2")
w = HSigmoid(name=name + "/hsigmoid")(w) if approx_sigmoid else nn.Activation("sigmoid", name=name + "/sigmoid")(w)
x = nn.multiply([x, w], name=name + "/mul")
return x
def resnext(channels,
init_block_channels,
cardinality,
bottleneck_width,
use_se,
in_channels=3,
classes=1000):
"""
ResNeXt model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431.
Also this class implements SE-ResNeXt from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
cardinality: int
Number of groups.
bottleneck_width: int
Width of bottleneck block.
use_se : bool
Whether to use SE block.
in_channels : int, default 3
Number of input channels.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = resnext_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = resnext_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
cardinality=cardinality,
bottleneck_width=bottleneck_width,
use_se=use_se,
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
return model
def resnext(channels,
init_block_channels,
cardinality,
bottleneck_width,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
ResNeXt model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
cardinality: int
Number of groups.
bottleneck_width: int
Width of bottleneck block.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, in_size[0], in_size[1]) if is_channels_first() else\
(in_size[0], in_size[1], in_channels)
input = nn.Input(shape=input_shape)
x = res_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = resnext_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
cardinality=cardinality,
bottleneck_width=bottleneck_width,
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
# x = nn.Flatten()(x)
x = flatten(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def squeezenext(channels,
init_block_channels,
final_block_channels,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
SqueezeNext model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = sqnxt_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = sqnxt_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = conv1x1_block(x=x,
in_channels=in_channels,
out_channels=final_block_channels,
use_bias=True,
name="features/final_block")
in_channels = final_block_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def seresnet(channels,
init_block_channels,
bottleneck,
conv1_stride,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
SE-ResNet model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
bottleneck : bool
Whether to use a bottleneck or simple block in units.
conv1_stride : bool
Whether to use stride in the first or the second convolution layer in units.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, in_size[0], in_size[1]) if is_channels_first() else\
(in_size[0], in_size[1], in_channels)
input = nn.Input(shape=input_shape)
x = res_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = seres_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
bottleneck=bottleneck,
conv1_stride=conv1_stride,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
# x = nn.Flatten()(x)
x = flatten(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def resnet(channels,
init_block_channels,
bottleneck,
conv1_stride,
use_se,
in_channels=3,
classes=1000):
"""
ResNet model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Also this class
implements SE-ResNet from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
bottleneck : bool
Whether to use a bottleneck or simple block in units.
conv1_stride : bool
Whether to use stride in the first or the second convolution layer in units.
use_se : bool
Whether to use SE block.
in_channels : int, default 3
Number of input channels.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = res_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
x = res_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
bottleneck=bottleneck,
conv1_stride=conv1_stride,
use_se=use_se,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
return model
def build(width=400, height=400, depth=3, classes=3):
#create a sequential model
model = models.Sequential()
#define the input dimensions
input_format = Input(shape=(400, 400, 3), name='image_input')
#use the pretrained model until the last conv-pool block
vgg_model = VGG16(weights='imagenet',
include_top=False,
input_tensor=input_format)
#freeze the layers for training
for layer in vgg_model.layers:
layer.trainable = False
#add the VGG convolutional base model
model.add(vgg_model)
#add new 3 layers of convolution and then avg_pooling and then softmax
#filters = 4096 kernel size = (6,6), stride = 6
model.add(
layers.Conv2D(filters=4096,
kernel_size=(6, 6),
activation='relu',
strides=(6, 6)))
#filters = 4096 kernel size = (1,1), stride = 1
model.add(
layers.Conv2D(filters=4096,
kernel_size=(1, 1),
activation='relu',
strides=(1, 1)))
#filters = 3 kernel size = (1,1), stride = 1
model.add(
layers.Conv2D(filters=3,
kernel_size=(1, 1),
activation='relu',
strides=(1, 1)))
#add avg_pooling layer pool_size=(2,2)
model.add(layers.AvgPool2D(pool_size=(2, 2)))
#flatten the output
model.add(layers.Flatten())
#add softmax
model.add(layers.Dense(units=3, activation="softmax"))
return model
def se_block(x,
channels,
reduction=16,
name="se_block"):
"""
Squeeze-and-Excitation block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
channels : int
Number of channels.
reduction : int, default 16
Squeeze reduction value.
name : str, default 'se_block'
Block name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
assert(len(x._keras_shape) == 4)
mid_cannels = channels // reduction
pool_size = x._keras_shape[2:4] if is_channels_first() else x._keras_shape[1:3]
w = nn.AvgPool2D(
pool_size=pool_size,
name=name + "/pool")(x)
w = conv1x1(
x=w,
in_channels=channels,
out_channels=mid_cannels,
use_bias=True,
name=name + "/conv1")
w = nn.Activation("relu", name=name + "/relu")(w)
w = conv1x1(
x=w,
in_channels=mid_cannels,
out_channels=channels,
use_bias=True,
name=name + "/conv2")
w = nn.Activation("sigmoid", name=name + "/sigmoid")(w)
x = nn.multiply([x, w], name=name + "/mul")
return x
def inception_v1(in_shape=(224, 224, 3), n_classes=1000, opt='sgd'):
in_layer = layers.Input(in_shape)
conv1 = layers.Conv2D(64, 7, strides=2, activation='relu',
padding='same')(in_layer)
#pool1 = layers.MaxPool2D(3, 2, padding='same')(conv1)
pad1 = layers.ZeroPadding2D()(conv1)
pool1 = layers.MaxPool2D(3, 2)(pad1)
conv2 = layers.Conv2D(192, 3, strides=1, activation='relu',
padding='same')(pool1)
pad2 = layers.ZeroPadding2D()(conv2)
pool2 = layers.MaxPool2D(3, 2)(pad2)
inception_3a = inception_module(pool2, 64, 96, 128, 16, 32, 32)
inception_3b = inception_module(inception_3a, 128, 128, 192, 32, 96, 64)
pad3 = layers.ZeroPadding2D()(inception_3b)
pool3 = layers.MaxPool2D(3, 2)(pad3)
inception_4a = inception_module(pool3, 192, 96, 208, 16, 48, 64)
inception_4b = inception_module(inception_4a, 160, 112, 224, 24, 64, 64)
inception_4c = inception_module(inception_4b, 128, 128, 256, 24, 64, 64)
inception_4d = inception_module(inception_4c, 112, 144, 288, 32, 64, 64)
inception_4e = inception_module(inception_4d, 256, 160, 320, 32, 128, 128)
pad4 = layers.ZeroPadding2D()(inception_4e)
pool4 = layers.MaxPool2D(3, 2)(pad4)
aux0 = aux_clf(inception_4a, n_classes)
aux1 = aux_clf(inception_4d, n_classes)
inception_5a = inception_module(pool4, 256, 160, 320, 32, 128, 128)
inception_5b = inception_module(inception_5a, 384, 192, 384, 48, 128, 128)
#pad5 = layers.ZeroPadding2D()(inception_5b)
pool5 = layers.AvgPool2D(7, 1)(inception_5b)
drop = layers.Dropout(0.4)(pool5)
flat = layers.Flatten()(drop)
preds = layers.Dense(n_classes, activation='softmax')(flat)
model = Model(in_layer, [preds, aux0, aux1])
model.compile(loss="categorical_crossentropy",
optimizer=opt,
metrics=["accuracy"])
return model
def se_block(x, channels, reduction=16, name="se_block"):
"""
Squeeze-and-Excitation block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
channels : int
Number of channels.
reduction : int, default 16
Squeeze reduction value.
name : str, default 'se_block'
Block name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
mid_cannels = channels // reduction
conv1 = conv1x1(out_channels=mid_cannels,
use_bias=True,
name=name + "/conv1")
relu = nn.Activation('relu', name=name + "/relu")
conv2 = conv1x1(out_channels=channels, use_bias=True, name=name + "/conv2")
sigmoid = nn.Activation('sigmoid', name=name + "/sigmoid")
assert (len(x.shape) == 4)
pool_size = x.shape[2:4] if K.image_data_format(
) == 'channels_first' else x.shape[1:3]
w = nn.AvgPool2D(pool_size=pool_size, name=name + "/pool")(x)
w = conv1(w)
w = relu(w)
w = conv2(w)
w = sigmoid(w)
x = nn.multiply([x, w], name=name + "/mul")
return x
def shufflenetv2(channels,
init_block_channels,
final_block_channels,
use_se=False,
use_residual=False,
in_channels=3,
classes=1000):
"""
ShuffleNetV2 model from 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,'
https://arxiv.org/abs/1807.11164.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
use_se : bool, default False
Whether to use SE block.
use_residual : bool, default False
Whether to use residual connections.
in_channels : int, default 3
Number of input channels.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = shuffle_init_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
downsample = (j == 0)
x = shuffle_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
downsample=downsample,
use_se=use_se,
use_residual=use_residual,
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
x = shuffle_conv1x1(x=x,
in_channels=in_channels,
out_channels=final_block_channels,
name="features/final_block")
in_channels = final_block_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
return model
def avgpool2d(x, pool_size, strides, padding=0, ceil_mode=False, name=None):
"""
Average pooling operation for two dimensional (spatial) data.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
pool_size : int or tuple/list of 2 int
Size of the max pooling windows.
strides : int or tuple/list of 2 int
Strides of the pooling.
padding : int or tuple/list of 2 int, default 0
Padding value for convolution layer.
ceil_mode : bool, default False
When `True`, will use ceil instead of floor to compute the output shape.
name : str, default None
Layer name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
if isinstance(pool_size, int):
pool_size = (pool_size, pool_size)
if isinstance(strides, int):
strides = (strides, strides)
if isinstance(padding, int):
padding = (padding, padding)
assert (padding[0] == 0) or (padding[0] == (pool_size[0] - 1) // 2)
assert (padding[1] == 0) or (padding[1] == (pool_size[1] - 1) // 2)
padding_ke = "valid" if padding[0] == 0 else "same"
if K.backend() == "tensorflow":
if ceil_mode:
height = int(x.shape[2])
out_height = float(height + 2 * padding[0] -
pool_size[0]) / strides[0] + 1.0
if math.ceil(out_height) > math.floor(out_height):
padding = (padding[0] + 1, padding[1])
width = int(x.shape[3])
out_width = float(width + 2 * padding[1] -
pool_size[1]) / strides[1] + 1.0
if math.ceil(out_width) > math.floor(out_width):
padding = (padding[0], padding[1] + 1)
if (padding[0] > 0) or (padding[1] > 0):
import tensorflow as tf
x = nn.Lambda((lambda z: tf.pad(z, [[0, 0], [0, 0],
list(padding),
list(padding)],
mode="REFLECT")
) if is_channels_first() else (
lambda z: tf.pad(z, [[0, 0],
list(padding),
list(padding), [0, 0]],
mode="REFLECT")))(x)
x = nn.AvgPool2D(pool_size=pool_size,
strides=1,
padding="valid",
name=name + "/pool")(x)
if (strides[0] > 1) or (strides[1] > 1):
x = nn.AvgPool2D(pool_size=1,
strides=strides,
padding="valid",
name=name + "/stride")(x)
return x
x = nn.AvgPool2D(pool_size=pool_size,
strides=strides,
padding=padding_ke,
name=name + "/pool")(x)
return x
def shuffle_unit(x,
in_channels,
out_channels,
groups,
downsample,
ignore_group,
name="shuffle_unit"):
"""
ShuffleNet unit.
Parameters:
----------
x : keras.backend tensor/variable/symbol
Input tensor/variable/symbol.
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
groups : int
Number of groups in convolution layers.
downsample : bool
Whether do downsample.
ignore_group : bool
Whether ignore group value in the first convolution layer.
name : str, default 'shuffle_unit'
Unit name.
Returns
-------
keras.backend tensor/variable/symbol
Resulted tensor/variable/symbol.
"""
mid_channels = out_channels // 4
if downsample:
out_channels -= in_channels
identity = x
x = conv1x1(x=x,
in_channels=in_channels,
out_channels=mid_channels,
groups=(1 if ignore_group else groups),
name=name + "/compress_conv1")
x = GluonBatchNormalization(name=name + "/compress_bn1")(x)
x = nn.Activation("relu", name=name + "/activ")(x)
x = channel_shuffle_lambda(channels=mid_channels,
groups=groups,
name=name + "/c_shuffle")(x)
x = depthwise_conv3x3(x=x,
channels=mid_channels,
strides=(2 if downsample else 1),
name=name + "/dw_conv2")
x = GluonBatchNormalization(name=name + "/dw_bn2")(x)
x = conv1x1(x=x,
in_channels=mid_channels,
out_channels=out_channels,
groups=groups,
name=name + "/expand_conv3")
x = GluonBatchNormalization(name=name + "/expand_bn3")(x)
x._keras_shape = tuple([d if d != 0 else None for d in x.shape])
if downsample:
identity = nn.AvgPool2D(pool_size=3,
strides=2,
padding="same",
name=name + "/avgpool")(identity)
channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
x = nn.concatenate([x, identity],
axis=channel_axis,
name=name + "/concat")
else:
x = nn.add([x, identity], name=name + "/add")
x = nn.Activation("relu", name=name + "/final_activ")(x)
return x
def mobilenetv2(channels,
init_block_channels,
final_block_channels,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
MobileNetV2 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if is_channels_first() else (224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = conv3x3_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
strides=2,
activation="relu6",
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
expansion = (i != 0) or (j != 0)
x = linear_bottleneck(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
expansion=expansion,
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
x = conv1x1_block(x=x,
in_channels=in_channels,
out_channels=final_block_channels,
activation="relu6",
name="features/final_block")
in_channels = final_block_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = conv1x1(x=x,
in_channels=in_channels,
out_channels=classes,
use_bias=False,
name="output")
# x = nn.Flatten()(x)
x = flatten(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def mnasnet_model(channels,
init_block_channels,
final_block_channels,
kernels3,
exp_factors,
se_factors,
init_block_use_skip,
final_block_use_skip,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
MnasNet model from 'MnasNet: Platform-Aware Neural Architecture Search for Mobile,'
https://arxiv.org/abs/1807.11626.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : list of 2 int
Number of output channels for the initial unit.
final_block_channels : list of 2 int
Number of output channels for the final block of the feature extractor.
kernels3 : list of list of int/bool
Using 3x3 (instead of 5x5) kernel for each unit.
exp_factors : list of list of int
Expansion factor for each unit.
se_factors : list of list of int
SE reduction factor for each unit.
init_block_use_skip : bool
Whether to use skip connection in the initial unit.
final_block_use_skip : bool
Whether to use skip connection in the final block of the feature extractor.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, in_size[0], in_size[1]) if is_channels_first() else\
(in_size[0], in_size[1], in_channels)
input = nn.Input(shape=input_shape)
x = mnas_init_block(
x=input,
in_channels=in_channels,
out_channels=init_block_channels[1],
mid_channels=init_block_channels[0],
use_skip=init_block_use_skip,
name="features/init_block")
in_channels = init_block_channels[1]
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) else 1
use_kernel3 = kernels3[i][j] == 1
exp_factor = exp_factors[i][j]
se_factor = se_factors[i][j]
x = dws_exp_se_res_unit(
x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
use_kernel3=use_kernel3,
exp_factor=exp_factor,
se_factor=se_factor,
name="features/stage{}/unit{}".format(i + 1, j + 1))
in_channels = out_channels
x = mnas_final_block(
x=x,
in_channels=in_channels,
out_channels=final_block_channels[1],
mid_channels=final_block_channels[0],
use_skip=final_block_use_skip,
name="features/final_block")
in_channels = final_block_channels[1]
x = nn.AvgPool2D(
pool_size=7,
strides=1,
name="features/final_pool")(x)
# x = nn.Flatten()(x)
x = flatten(x)
x = nn.Dense(
units=classes,
input_dim=in_channels,
name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def Unet(GAUSSIAN_NOISE=0.1,
UPSAMPLE_MODE='SIMPLE',
NET_SCALING=(1, 1),
EDGE_CROP=16):
def upsample_conv(filters, kernel_size, strides, padding):
return layers.Conv2DTranspose(filters,
kernel_size,
strides=strides,
padding=padding)
def upsample_simple(filters, kernel_size, strides, padding):
return layers.UpSampling2D(strides)
if UPSAMPLE_MODE == 'DECONV':
upsample = upsample_conv
else:
upsample = upsample_simple
input_img = layers.Input((768, 768, 3), name='RGB_Input')
pp_in_layer = input_img
if NET_SCALING is not None:
pp_in_layer = layers.AvgPool2D(NET_SCALING)(pp_in_layer)
pp_in_layer = layers.GaussianNoise(GAUSSIAN_NOISE)(pp_in_layer)
pp_in_layer = layers.BatchNormalization()(pp_in_layer)
c1 = layers.Conv2D(16, (3, 3), activation='relu',
padding='same')(pp_in_layer)
c1 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(c1)
p1 = layers.MaxPooling2D((2, 2))(c1)
c2 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(p1)
c2 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(c2)
p2 = layers.MaxPooling2D((2, 2))(c2)
c3 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(p2)
c3 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(c3)
p3 = layers.MaxPooling2D((2, 2))(c3)
c4 = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(p3)
c4 = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(c4)
p4 = layers.MaxPooling2D(pool_size=(2, 2))(c4)
c5 = layers.Conv2D(256, (3, 3), activation='relu', padding='same')(p4)
c5 = layers.Conv2D(256, (3, 3), activation='relu', padding='same')(c5)
u6 = upsample(128, (2, 2), strides=(2, 2), padding='same')(c5)
u6 = layers.concatenate([u6, c4])
c6 = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(u6)
c6 = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(c6)
u7 = upsample(64, (2, 2), strides=(2, 2), padding='same')(c6)
u7 = layers.concatenate([u7, c3])
c7 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(u7)
c7 = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(c7)
u8 = upsample(32, (2, 2), strides=(2, 2), padding='same')(c7)
u8 = layers.concatenate([u8, c2])
c8 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(u8)
c8 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(c8)
u9 = upsample(16, (2, 2), strides=(2, 2), padding='same')(c8)
u9 = layers.concatenate([u9, c1], axis=3)
c9 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(u9)
c9 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(c9)
d = layers.Conv2D(1, (1, 1), activation='sigmoid')(c9)
# d = layers.Cropping2D((EDGE_CROP, EDGE_CROP))(d)
# d = layers.ZeroPadding2D((EDGE_CROP, EDGE_CROP))(d)
if NET_SCALING is not None:
d = layers.UpSampling2D(NET_SCALING)(d)
seg_model = models.Model(inputs=[input_img], outputs=[d])
seg_model.summary()
return seg_model
# Build U-Net model
def upsample_conv(filters, kernel_size, strides, padding):
return layers.Conv2DTranspose(filters, kernel_size, strides=strides, padding=padding)
def upsample_simple(filters, kernel_size, strides, padding):
return layers.UpSampling2D(strides)
if UPSAMPLE_MODE=='DECONV':
upsample=upsample_conv
else:
upsample=upsample_simple
input_img = layers.Input(t_x.shape[1:], name = 'RGB_Input')
pp_in_layer = input_img
if NET_SCALING is not None:
pp_in_layer = layers.AvgPool2D(NET_SCALING)(pp_in_layer)
pp_in_layer = layers.GaussianNoise(GAUSSIAN_NOISE)(pp_in_layer)
pp_in_layer = layers.BatchNormalization()(pp_in_layer)
c1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same') (pp_in_layer)
c1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same') (c1)
p1 = layers.MaxPooling2D((2, 2)) (c1)
c2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same') (p1)
c2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same') (c2)
p2 = layers.MaxPooling2D((2, 2)) (c2)
c3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same') (p2)
c3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same') (c3)
p3 = layers.MaxPooling2D((2, 2)) (c3)
def igcv3(channels,
init_block_channels,
final_block_channels,
in_channels=3,
in_size=(224, 224),
classes=1000):
"""
IGCV3 model from 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,'
https://arxiv.org/abs/1806.00178.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
in_channels : int, default 3
Number of input channels.
in_size : tuple of two ints, default (224, 224)
Spatial size of the expected input image.
classes : int, default 1000
Number of classification classes.
"""
input_shape = (in_channels, 224,
224) if K.image_data_format() == 'channels_first' else (
224, 224, in_channels)
input = nn.Input(shape=input_shape)
x = conv3x3_block(x=input,
in_channels=in_channels,
out_channels=init_block_channels,
strides=2,
activation="relu6",
name="features/init_block")
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
strides = 2 if (j == 0) and (i != 0) else 1
expansion = (i != 0) or (j != 0)
x = inv_res_unit(x=x,
in_channels=in_channels,
out_channels=out_channels,
strides=strides,
expansion=expansion,
name="features/stage{}/unit{}".format(
i + 1, j + 1))
in_channels = out_channels
x = conv1x1_block(x=x,
in_channels=in_channels,
out_channels=final_block_channels,
activation="relu6",
name="features/final_block")
in_channels = final_block_channels
x = nn.AvgPool2D(pool_size=7, strides=1, name="features/final_pool")(x)
x = nn.Flatten()(x)
x = nn.Dense(units=classes, input_dim=in_channels, name="output")(x)
model = Model(inputs=input, outputs=x)
model.in_size = in_size
model.classes = classes
return model
def upsample_simple(filters, kernel_size, strides, padding):
return layers.UpSampling2D(strides)
if UPSAMPLE_MODE == 'DECONV':
upsample = upsample_conv
else:
upsample = upsample_simple
input_img = layers.Input(t_x.shape[1:], name='RGB_Input')
pp_in_layer = input_img
if NET_SCALING is not None:
pp_in_layer = layers.AvgPool2D(NET_SCALING)(pp_in_layer)
pp_in_layer = layers.GaussianNoise(GAUSSIAN_NOISE)(pp_in_layer)
pp_in_layer = layers.BatchNormalization()(pp_in_layer)
c1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same')(pp_in_layer)
c1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same')(c1)
p1 = layers.MaxPooling2D((2, 2))(c1)
c2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(p1)
c2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same')(c2)
p2 = layers.MaxPooling2D((2, 2))(c2)
c3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(p2)
c3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(c3)
p3 = layers.MaxPooling2D((2, 2))(c3)
padding='same',
activation='relu',
**init_reg)(lyrout2)
lyrout4 = lyr.MaxPool2D(3, 2, 'same')(lyrout3)
lyrout5 = custom.inception(lyrout4, 64, 96, 128, 16, 32, 32, **init_reg)
lyrout6 = custom.inception(lyrout5, 128, 128, 192, 32, 96, 64, **init_reg)
lyrout7 = lyr.MaxPool2D(3, 2, 'same')(lyrout6)
lyrout8 = custom.inception(lyrout7, 192, 96, 208, 16, 48, 64, **init_reg)
lyrout9 = custom.inception(lyrout8, 160, 112, 224, 24, 64, 64, **init_reg)
lyrout10 = custom.inception(lyrout9, 128, 128, 256, 24, 64, 64, **init_reg)
lyrout11 = custom.inception(lyrout10, 112, 144, 288, 32, 64, 64, **init_reg)
lyrout12 = custom.inception(lyrout11, 256, 160, 320, 32, 128, 128, **init_reg)
lyrout13 = lyr.MaxPool2D(3, 2, 'same')(lyrout12)
lyrout14 = custom.inception(lyrout13, 256, 160, 320, 32, 128, 128, **init_reg)
lyrout15 = custom.inception(lyrout14, 384, 192, 384, 48, 128, 128, **init_reg)
lyrout16 = lyr.AvgPool2D(7, 1, 'valid')(lyrout15)
lyrout17 = lyr.Flatten()(lyrout16)
lyrout18 = lyr.Dropout(0.4)(lyrout17)
out = lyr.Dense(num_classes, activation='softmax', **init_reg)(lyrout18)
model = mod.Model(inputs=input, outputs=out)
"""
Optimizer, Loss, & Metrics
Using zero built in decay and relying exclusively on the heuristic
used in the original paper.
"""
optimizer = opt.Adam(0.001)
# This line below is what was used in the original ResNet paper. Adam is the recommended approach now though.
# We use this for GoogLeNet because its paper only specifies that they used momentum=0.9 but not what the initial
# learning rate was.