def build(self, input_shape):
if self._expand_ratio != 1:
# First 1x1 conv for channel expansion.
self._conv0 = tf.keras.layers.Conv2D(
filters=self._in_filters * self._expand_ratio,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
# Depthwise conv.
self._conv1 = tf.keras.layers.DepthwiseConv2D(
kernel_size=(self._kernel_size, self._kernel_size),
strides=self._strides,
padding='same',
use_bias=False,
depthwise_initializer=self._kernel_initializer,
depthwise_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
# Squeeze and excitation.
if self._se_ratio is not None and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._in_filters,
se_ratio=self._se_ratio,
expand_ratio=self._expand_ratio,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
# Last 1x1 conv.
self._conv2 = tf.keras.layers.Conv2D(
filters=self._out_filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
super(InvertedBottleneckBlock, self).build(input_shape)
def build(self, input_shape):
if self._use_projection:
self._shortcut = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=1,
strides=self._strides,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv1 = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=3,
strides=self._strides,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv2 = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=3,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters,
out_filters=self._filters,
se_ratio=self._se_ratio,
use_3d_input=True,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
super(ResidualBlock3DVolume, self).build(input_shape)
def build(self, input_shape):
expand_filters = self._in_filters
if self._expand_ratio > 1:
# First 1x1 conv for channel expansion.
expand_filters = nn_layers.make_divisible(
self._in_filters * self._expand_ratio, self._divisible_by)
expand_kernel = 1 if self._use_depthwise else self._kernel_size
expand_stride = 1 if self._use_depthwise else self._strides
self._conv0 = tf.keras.layers.Conv2D(
filters=expand_filters,
kernel_size=expand_kernel,
strides=expand_stride,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._use_depthwise:
# Depthwise conv.
self._conv1 = tf.keras.layers.DepthwiseConv2D(
kernel_size=(self._kernel_size, self._kernel_size),
strides=self._strides,
padding='same',
depth_multiplier=1,
dilation_rate=self._dilation_rate,
use_bias=False,
depthwise_initializer=self._kernel_initializer,
depthwise_regularizer=self._depthsize_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
# Squeeze and excitation.
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
logging.info('Use Squeeze and excitation.')
in_filters = self._in_filters
if self._expand_se_in_filters:
in_filters = expand_filters
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=in_filters,
out_filters=expand_filters,
se_ratio=self._se_ratio,
divisible_by=self._divisible_by,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer,
activation=self._se_inner_activation,
gating_activation=self._se_gating_activation)
else:
self._squeeze_excitation = None
# Last 1x1 conv.
self._conv2 = tf.keras.layers.Conv2D(
filters=self._out_filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
super(InvertedBottleneckBlock, self).build(input_shape)
def build(self, input_shape):
if self._use_projection:
if self._resnetd_shortcut:
self._shortcut0 = tf.keras.layers.AveragePooling2D(
pool_size=2, strides=self._strides, padding='same')
self._shortcut1 = tf.keras.layers.Conv2D(
filters=self._filters * 4,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._shortcut = tf.keras.layers.Conv2D(
filters=self._filters * 4,
kernel_size=1,
strides=self._strides,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._conv1 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._activation1 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
self._conv2 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=3,
strides=self._strides,
dilation_rate=self._dilation_rate,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._activation2 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
self._conv3 = tf.keras.layers.Conv2D(
filters=self._filters * 4,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm3 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._activation3 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters * 4,
out_filters=self._filters * 4,
se_ratio=self._se_ratio,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
self._add = tf.keras.layers.Add()
super(BottleneckBlock, self).build(input_shape)
def build(self, input_shape):
if self._use_projection:
self._shortcut = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=1,
strides=self._strides,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
conv1_padding = 'same'
# explicit padding here is added for centernet
if self._use_explicit_padding:
self._pad = tf.keras.layers.ZeroPadding2D(padding=(1, 1))
conv1_padding = 'valid'
self._conv1 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=3,
strides=self._strides,
padding=conv1_padding,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._conv2 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=3,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters,
out_filters=self._filters,
se_ratio=self._se_ratio,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
super(ResidualBlock, self).build(input_shape)
def build(self, input_shape):
self._shortcut_maxpool = tf.keras.layers.MaxPool3D(
pool_size=[1, 1, 1],
strides=[
self._temporal_strides, self._spatial_strides,
self._spatial_strides
])
self._shortcut_conv = tf.keras.layers.Conv3D(
filters=4 * self._filters,
kernel_size=1,
strides=[
self._temporal_strides, self._spatial_strides,
self._spatial_strides
],
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._temporal_conv = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=[self._temporal_kernel_size, 1, 1],
strides=[self._temporal_strides, 1, 1],
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._spatial_conv = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=[1, 3, 3],
strides=[1, self._spatial_strides, self._spatial_strides],
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._expand_conv = tf.keras.layers.Conv3D(
filters=4 * self._filters,
kernel_size=[1, 1, 1],
strides=[1, 1, 1],
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm3 = self._norm(axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters * 4,
out_filters=self._filters * 4,
se_ratio=self._se_ratio,
use_3d_input=True,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
if self._use_self_gating:
self._self_gating = SelfGating(filters=4 * self._filters)
else:
self._self_gating = None
super(BottleneckBlock3D, self).build(input_shape)
