def __init__(self,
num_anchors,
kernel_initializer,
kernel_regularizer,
name='classification_subnet'):
super().__init__(name=name)
self.num_anchors = num_anchors
self.conv_pre = Sequential([
Sequential([
L.Conv2D(256,
3,
1,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization(),
L.Activation(tf.nn.relu),
]) for _ in range(4)
])
self.conv_out = L.Conv2D(num_anchors * 4,
3,
1,
padding='same',
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
def __init__(self,
kernel_initializer,
kernel_regularizer,
name='upsample_merge'):
super().__init__(name=name)
self.conv_lateral = Sequential([
L.Conv2D(256,
1,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization()
])
self.conv_merge = Sequential([
L.Conv2D(256,
3,
1,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization()
])
def __init__(self,
filters,
dropout_rate,
kernel_initializer,
kernel_regularizer,
name='bottleneck_composite_function'):
layers = [
L.BatchNormalization(),
L.Activation(tf.nn.relu),
L.Conv2D(
filters * 4,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Dropout(dropout_rate),
L.BatchNormalization(),
L.Activation(tf.nn.relu),
L.Conv2D(
filters,
3,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Dropout(dropout_rate),
]
super().__init__(layers, name=name)
def __init__(self,
kernel_initializer=None,
kernel_regularizer=None,
name='feature_pyramid_network'):
if kernel_initializer is None:
kernel_initializer = tf.random_normal_initializer(mean=0.0,
stddev=0.01)
if kernel_regularizer is None:
kernel_regularizer = tf.contrib.layers.l2_regularizer(scale=1e-4)
super().__init__(name=name)
self.p6_from_c5 = Sequential([
L.Conv2D(256,
3,
2,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization()
])
self.p7_from_p6 = Sequential([
L.Activation(tf.nn.relu),
L.Conv2D(256,
3,
2,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization()
])
self.p5_from_c5 = Sequential([
L.Conv2D(256,
1,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization()
])
self.p4_from_c4p5 = UpsampleMerge(
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='upsample_merge_c4p5')
self.p3_from_c3p4 = UpsampleMerge(
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='upsample_merge_c3p4')
def __init__(self,
filters,
strides,
expansion_factor,
dropout_rate,
kernel_initializer,
kernel_regularizer,
name='bottleneck'):
super().__init__(name=name)
self.expand_conv = Sequential([
L.Conv2D(
filters *
expansion_factor, # FIXME: should be `input_shape[3].value * expansion_factor`
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization(),
L.Activation(tf.nn.relu6),
L.Dropout(dropout_rate)
])
self.depthwise_conv = Sequential([
L.DepthwiseConv2D(3,
strides=strides,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization(),
L.Activation(tf.nn.relu6),
L.Dropout(dropout_rate)
])
self.linear_conv = Sequential([
L.Conv2D(filters,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization(),
L.Dropout(dropout_rate)
])
def __init__(self,
kernel_initializer,
kernel_regularizer,
name='resnext_conv1'):
super().__init__(name=name)
self.conv = L.Conv2D(64,
7,
2,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bn = L.BatchNormalization()
def __init__(self,
num_anchors,
num_classes,
kernel_initializer,
kernel_regularizer,
name='classification_subnet'):
super().__init__(name=name)
self.num_anchors = num_anchors
self.num_classes = num_classes
self.conv_pre = Sequential([
Sequential([
L.Conv2D(256,
3,
1,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Normalization(),
L.Activation(tf.nn.relu),
]) for _ in range(4)
])
pi = 0.01
bias_prior_initializer = tf.constant_initializer(-math.log((1 - pi) /
pi))
self.conv_out = L.Conv2D(num_anchors * num_classes,
3,
1,
padding='same',
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_initializer=bias_prior_initializer)
def __init__(self,
input_filters,
compression_factor,
dropout_rate,
kernel_initializer,
kernel_regularizer,
name='transition_layer'):
self.input_filters = input_filters
filters = int(input_filters * compression_factor)
layers = [
L.BatchNormalization(),
L.Conv2D(
filters,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.Dropout(dropout_rate),
L.AveragePooling2D(2, 2, padding='same')
]
super().__init__(layers, name=name)
def __init__(self,
dropout_rate,
kernel_initializer=None,
kernel_regularizer=None,
name='mobilenet_v2'):
if kernel_initializer is None:
kernel_initializer = tf.contrib.layers.variance_scaling_initializer(
factor=2.0, mode='FAN_IN', uniform=False)
if kernel_regularizer is None:
kernel_regularizer = tf.contrib.layers.l2_regularizer(scale=4e-5)
super().__init__(name=name)
self.input_conv = Sequential([
L.Conv2D(32,
3,
strides=2,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization(),
L.Activation(tf.nn.relu6),
L.Dropout(dropout_rate)
])
self.bottleneck_1_1 = Bottleneck(16,
expansion_factor=1,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_2_1 = Bottleneck(24,
expansion_factor=6,
strides=2,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_2_2 = Bottleneck(24,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_3_1 = Bottleneck(32,
expansion_factor=6,
strides=2,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_3_2 = Bottleneck(32,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_3_3 = Bottleneck(32,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_4_1 = Bottleneck(64,
expansion_factor=6,
strides=2,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_4_2 = Bottleneck(64,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_4_3 = Bottleneck(64,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_4_4 = Bottleneck(64,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_5_1 = Bottleneck(96,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_5_2 = Bottleneck(96,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_5_3 = Bottleneck(96,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_6_1 = Bottleneck(160,
expansion_factor=6,
strides=2,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_6_2 = Bottleneck(160,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_6_3 = Bottleneck(160,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.bottleneck_7_1 = Bottleneck(320,
expansion_factor=6,
strides=1,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer)
self.output_conv = Sequential([
L.Conv2D(32,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization(),
L.Activation(tf.nn.relu6),
L.Dropout(dropout_rate)
])
def __init__(self,
filters,
projection_type,
kernel_initializer,
kernel_regularizer,
cardinality=32,
name='resnext_bottleneck'):
assert filters % cardinality == 0
assert projection_type in ProjectionType
super().__init__(name=name)
# identity
if projection_type is ProjectionType.DOWN:
self.conv_identity = Sequential([
L.Conv2D(filters * 4,
3,
2,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization()
])
elif projection_type is ProjectionType.CONV:
self.conv_identity = Sequential([
L.Conv2D(filters * 4,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization()
])
elif projection_type is ProjectionType.NONE:
self.conv_identity = None
# conv_1
self.conv_1 = Sequential([
L.Conv2D(filters * 2,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization(),
L.Activation(tf.nn.relu)
])
# conv_2
self.conv_2 = []
for _ in range(cardinality):
strides = 2 if projection_type is ProjectionType.DOWN else 1
self.conv_2.append(
L.Conv2D((filters * 2) // cardinality,
3,
strides,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer))
self.bn_2 = L.BatchNormalization()
# conv_3
self.conv_3 = Sequential([
L.Conv2D(filters * 4,
1,
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer),
L.BatchNormalization()
])
def __init__(self,
blocks,
growth_rate,
compression_factor,
bottleneck,
dropout_rate,
kernel_initializer,
kernel_regularizer,
name='densenet_bc_imagenet'):
super().__init__(name=name)
self.conv_1 = Sequential([
L.Conv2D(
2 * growth_rate,
7,
2,
padding='same',
use_bias=False,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='conv_1'),
L.BatchNormalization(),
L.Activation(tf.nn.relu)
])
self.conv_1_max_pool = L.MaxPooling2D(3, 2, padding='same')
self.dense_block_1 = DenseBlock(
growth_rate,
depth=blocks[1],
bottleneck=bottleneck,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='dense_block_1')
self.transition_layer_1 = TransitionLayer(
input_filters=blocks[1] * growth_rate + 64,
compression_factor=compression_factor,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='transition_layer_1')
self.dense_block_2 = DenseBlock(
growth_rate,
depth=blocks[2],
bottleneck=bottleneck,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='dense_block_2')
self.transition_layer_2 = TransitionLayer(
input_filters=blocks[2] * growth_rate + self.transition_layer_1.layers[1].filters, # FIXME:
compression_factor=compression_factor,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='transition_layer_2')
self.dense_block_3 = DenseBlock(
growth_rate,
depth=blocks[3],
bottleneck=bottleneck,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='dense_block_3')
self.transition_layer_3 = TransitionLayer(
input_filters=blocks[3] * growth_rate + self.transition_layer_2.layers[1].filters, # FIXME:
compression_factor=compression_factor,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='transition_layer_3')
self.dense_block_4 = DenseBlock(
growth_rate,
depth=blocks[4],
bottleneck=bottleneck,
dropout_rate=dropout_rate,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
name='dense_block_4')
