def __init__(self,
_in_channel,
output_dim,
stride=1,
downsample=None,
**kwargs):
super(Bottleneck, self).__init__(**kwargs)
self.downsample = downsample
self.conv1 = Conv2D(filters=output_dim,
kernel_size=1,
padding="same",
use_bias=False)
self.bn1 = GroupNormalization(groups=GN_GROUPS)
self.conv2 = Conv2D(filters=output_dim,
kernel_size=3,
strides=stride,
padding="same",
use_bias=False)
self.bn2 = GroupNormalization(groups=GN_GROUPS)
self.conv3 = Conv2D(filters=output_dim * self.expansion,
kernel_size=1,
padding="same",
use_bias=False)
self.bn3 = GroupNormalization(groups=GN_GROUPS)
self.relu = ReLU()
def single_hourglass(output_channels,
n_levels=4,
channels=32,
channels_growth=2,
spatial_dims=2,
spacing=1,
norm_groups=4):
'''Combines an hourglass block with input/output blocks to
increase/decrease the number of channels.
Notes:
Expects the input to be already padded)
'''
Conv = get_nd_conv(spatial_dims)
hglass = _stack_layers([
Conv(channels, kernel_size=1, padding='same'),
bottleneck_conv_block(channels, spatial_dims, norm_groups),
hourglass_block(n_levels, channels, channels_growth, spatial_dims,
spacing, norm_groups),
bottleneck_conv_block(channels, spatial_dims, norm_groups),
Activation(),
GroupNormalization(groups=norm_groups, axis=-1),
Conv(channels, kernel_size=3, padding='same'),
Activation(),
GroupNormalization(groups=norm_groups, axis=-1),
Conv(output_channels, kernel_size=1, padding='same'),
])
return hglass
def conv_block_3d(x, growth_rate, name, GN=True):
"""A building block for a dense block.
Arguments:
x: input tensor.
growth_rate: float, growth rate at dense layers.
name: string, block label.
Returns:
Output tensor for the block.
"""
if GN:
x1 = GroupNormalization(groups=2, axis=-1, name=name + '_0_gn')(x)
else:
x1 = layers.BatchNormalization(name=name + '_0_bn')(x)
# x1 = layers.BatchNormalization(axis=-1, epsilon=1.001e-5, name=name + '_0_bn')(x)
x1 = layers.Activation('selu', name=name + '_0_selu')(x1)
x1 = layers.Conv3D(2 * growth_rate,
1,
use_bias=False,
name=name + '_1_conv',
padding='same')(x1)
# x1 = layers.BatchNormalization(axis=-1, epsilon=1.001e-5, name=name + '_1_bn')(x1)
if GN:
x1 = GroupNormalization(groups=2, axis=-1, name=name + '_1_gn')(x1)
else:
x1 = layers.BatchNormalization(name=name + '_1_bn')(x1)
x1 = layers.Activation('selu', name=name + '_1_selu')(x1)
x1 = layers.Conv3D(growth_rate,
3,
padding='same',
use_bias=False,
name=name + '_2_conv')(x1)
x = layers.Concatenate(axis=-1, name=name + '_concat')([x, x1])
return x
def transition_layer1(x, out_filters_list=[32, 64], GROUPS = 18):
x0 = Conv2D(out_filters_list[0], 3, padding='same', use_bias=False, kernel_initializer='he_normal')(x)
x0 = GroupNormalization(axis=3, groups=GROUPS)(x0)
x0 = Activation('relu')(x0)
x1 = Conv2D(out_filters_list[1], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x)
x1 = GroupNormalization(axis=3, groups=GROUPS)(x1)
x1 = Activation('relu')(x1)
return [x0, x1]
def _make_fuse_layers(self):
if self.num_branches == 1:
return None
num_branches = self.num_branches
num_inchannels = self.num_inchannels
fuse_layers = []
for i in range(num_branches if self.multi_scale_output else 1):
fuse_layer = []
for j in range(num_branches):
if j > i:
fuse_layer.append(Sequential([
Conv2D(
filters=num_inchannels[i],
kernel_size=(1, 1),
strides=(1, 1),
padding="same",
use_bias=False
),
GroupNormalization(groups=self.GN_GROUPS),
UpSampling2D(size=(2 ** (j - i), 2 ** (j - i)), interpolation='bilinear')
]))
elif j == i:
fuse_layer.append(None)
else:
conv3x3s = []
for k in range(i - j):
if k == i - j - 1:
num_outchannels_conv3x3 = num_inchannels[i]
conv3x3s.append(Sequential([
Conv2D(
filters=num_outchannels_conv3x3,
kernel_size=(3,3),
strides=(2,2),
padding="same",
use_bias=False),
GroupNormalization(groups=self.GN_GROUPS)
]))
else:
num_outchannels_conv3x3 = num_inchannels[j]
conv3x3s.append(Sequential([
Conv2D(
filters=num_outchannels_conv3x3,
kernel_size=(3,3),
strides=(2,2),
padding="same",
use_bias=False),
GroupNormalization(groups=self.GN_GROUPS),
ReLU()
]))
fuse_layer.append(Sequential(conv3x3s))
fuse_layers.append(fuse_layer)
return fuse_layers
def __init__(self, _in_channel, output_dim, stride=1, downsample=None, GN_GROUPS=32, after_stem=False, **kwargs):
super(BasicBlock, self).__init__(**kwargs)
self.downsample = downsample
self.conv1 = conv3x3(output_dim, stride)
if after_stem:
self.gn1 = GroupNormalization(groups=output_dim)
else:
self.gn1 = GroupNormalization(groups=GN_GROUPS)
self.relu = ReLU()
self.conv2 = conv3x3(output_dim)
self.gn2 = GroupNormalization(groups=GN_GROUPS)
def fuse_layer1(x, out_filters_list=[32, 64], GROUPS = 18):
x0_0 = x[0]
x0_1 = Conv2D(out_filters_list[0], 1, use_bias=False, kernel_initializer='he_normal')(x[1])
x0_1 = GroupNormalization(axis=3, groups=GROUPS)(x0_1)
x0_1 = UpSampling2D(size=(2, 2), interpolation="bilinear")(x0_1)
x0 = add([x0_0, x0_1])
x1_0 = Conv2D(out_filters_list[1], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x[0])
x1_0 = GroupNormalization(axis=3, groups=GROUPS)(x1_0)
x1_1 = x[1]
x1 = add([x1_0, x1_1])
return [x0, x1]
def __init__(self, filters, gp_num=48, **kwargs):
super(BasicShuffleUnitV2, self).__init__()
filters //= 2
self.model = tf.keras.models.Sequential([
GroupNormalization(groups=gp_num, axis=-1),
Conv2D(filters, 1, use_bias=False),
ReLU(),
GroupNormalization(groups=gp_num, axis=-1),
DepthwiseConv2D(3, padding='same', use_bias=False),
GroupNormalization(groups=gp_num, axis=-1),
Conv2D(filters, 1, use_bias=False),
ReLU(),
])
def __init__(self,
_in_channel,
output_dim,
stride=1,
downsample=None,
**kwargs):
super(BasicBlock, self).__init__(**kwargs)
self.downsample = downsample
self.conv1 = conv3x3(output_dim, stride)
self.bn1 = GroupNormalization(groups=GN_GROUPS)
self.relu = ReLU()
self.conv2 = conv3x3(output_dim)
self.bn2 = GroupNormalization(groups=GN_GROUPS)
def _make_head(self, pre_stage_channels, GN_GROUPS):
head_block = Bottleneck
head_channels = [32, 64, 128, 256]
# Increasing the #channels on each resolution
# from C, 2C, 4C, 8C to 128, 256, 512, 1024
incre_modules = []
for i, channels in enumerate(pre_stage_channels):
incre_module = self._make_layer(
block=head_block,
planes=head_channels[i],
blocks=1,
stride=1,
head_layer=True
)
incre_modules.append(incre_module)
# downsampling modules
downsamp_modules = []
for i in range(len(pre_stage_channels)-1):
out_channels = head_channels[i+1] * head_block.expansion
downsamp_module = Sequential([
Conv2D(
filters=out_channels,
kernel_size=(3,3),
strides=(2,2),
padding='same'
),
GroupNormalization(groups=GN_GROUPS),
ReLU()
])
downsamp_modules.append(downsamp_module)
final_layer = Sequential([
Conv2D(
filters=2048,
kernel_size=(1,1),
strides=(1,1),
padding='valid'
),
GroupNormalization(groups=GN_GROUPS),
ReLU(),
GlobalAveragePooling2D(),
Dense(self.NUM_CLASSES, dtype='float32')
])
return incre_modules, downsamp_modules, final_layer
def _normalization(inputs, norm='bn', name=None):
if norm == 'bn':
return BatchNormalization(name=name)(inputs)
elif norm == 'syncbn':
return SyncBatchNormalization(name=name)(inputs)
elif norm == 'in':
return InstanceNormalization(name=name)(inputs)
elif norm == 'ln':
return LayerNormalization(name=name)(inputs)
elif 'gn' in norm:
if len(norm) == 2:
return GroupNormalization(name=name)(inputs)
else:
return GroupNormalization(groups=int(norm[2:]), name=name)(inputs)
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = Sequential([
Conv2D(planes * block.expansion,
kernel_size=(1, 1),
strides=(stride, stride),
padding="same",
use_bias=False),
GroupNormalization(groups=self.GN_GROUPS)
])
layers = [
block(self.inplanes,
planes,
stride,
downsample,
GN_GROUPS=self.GN_GROUPS)
]
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(
block(self.inplanes, planes, GN_GROUPS=self.GN_GROUPS))
return Sequential(layers)
def _make_layer(self, block, planes, blocks, stride=1, after_stem=False, head_layer=False):
downsample = None
if stride != 1 or self.inplanes != planes*block.expansion:
# TODO: make it so this if statement = True when C = 32
if head_layer:
GN_GROUPS = 32
else:
GN_GROUPS = self.GN_GROUPS
downsample = Sequential([
Conv2D(
planes * block.expansion,
kernel_size=(1,1),
strides=(stride,stride),
padding="same",
use_bias=False),
GroupNormalization(groups=GN_GROUPS)
])
if after_stem:
layers = [block(self.inplanes, planes, stride, downsample, GN_GROUPS=self.GN_GROUPS, after_stem=True)]
elif head_layer:
layers = [block(self.inplanes, planes, stride, downsample, GN_GROUPS=32, after_stem=False)]
else:
layers = [block(self.inplanes, planes, stride, downsample)]
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
if after_stem:
layers.append(block(self.inplanes, planes, GN_GROUPS=self.GN_GROUPS, after_stem=True))
elif head_layer:
layers.append(block(self.inplanes, planes, GN_GROUPS=32, after_stem=False))
else:
layers.append(block(self.inplanes, planes))
return Sequential(layers)
def test_group_normalization(self, use_cpu_only, backend, rank, groups,
axis, epsilon, center, scale):
tensorflow_addons = pytest.importorskip("tensorflow_addons")
from tensorflow_addons.layers import GroupNormalization
shape = np.random.randint(low=2, high=4, size=rank)
shape[-1] = shape[-1] * groups # groups must be a multiple of channels
model = tf.keras.Sequential([
GroupNormalization(
batch_input_shape=shape,
groups=groups,
axis=axis,
epsilon=epsilon,
center=center,
scale=scale,
)
])
run_compare_tf_keras(
model,
[random_gen(shape, rand_min=-1, rand_max=1)],
use_cpu_only=use_cpu_only,
backend=backend,
atol=1e-3,
rtol=1e-4,
)
def _make_one_branch(self,
branch_index,
block,
num_blocks,
num_channels,
stride=1):
downsample = None
if stride != 1 or self.num_inchannels[
branch_index] != num_channels[branch_index] * block.expansion:
downsample = Sequential(
Conv2D(self.num_inchannels[branch_index],
num_channels[branch_index] * block.expansion,
kernel_size=1,
strides=stride,
bias=False),
GroupNormalization(groups=GN_GROUPS),
)
layers = [
block(self.num_inchannels[branch_index],
num_channels[branch_index], stride, downsample)
]
self.num_inchannels[
branch_index] = num_channels[branch_index] * block.expansion
for i in range(1, num_blocks[branch_index]):
layers.append(
block(self.num_inchannels[branch_index],
num_channels[branch_index]))
return Sequential(layers)
def __init__(self, c, name=None, drop_rate=0.0, groups=32):
super().__init__(name=name)
self.c = c
self.drop_rate = drop_rate
self.conv1 = Conv2D(c, 3, padding='same')
self.conv2 = Conv2D(c, 3, padding='same')
self.norm1 = GroupNormalization(groups=groups)
self.norm2 = GroupNormalization(groups=groups)
self.temb_proj = Dense(c)
if drop_rate > 0.01:
self.drop_fn = Dropout(drop_rate)
else:
self.drop_fn = tf.identity
def fuse_layer3(x, out_filters_list=[32, 64, 128, 256], GROUPS = 18):
x0_0 = x[0]
x0_1 = Conv2D(out_filters_list[1], 1, use_bias=False, kernel_initializer='he_normal')(x[1])
x0_1 = GroupNormalization(axis=3, groups=GROUPS)(x0_1)
x0_1 = UpSampling2D(size=(2, 2), interpolation="bilinear")(x0_1)
x0_2 = Conv2D(out_filters_list[2], 1, use_bias=False, kernel_initializer='he_normal')(x[2])
x0_2 = GroupNormalization(axis=3, groups=GROUPS)(x0_2)
x0_2 = UpSampling2D(size=(4, 4), interpolation="bilinear")(x0_2)
x0_3 = Conv2D(out_filters_list[3], 1, use_bias=False, kernel_initializer='he_normal')(x[3])
x0_3 = GroupNormalization(axis=3, groups=GROUPS)(x0_3)
x0_3 = UpSampling2D(size=(8, 8), interpolation="bilinear")(x0_3)
x0 = concatenate([x0_0, x0_1, x0_2, x0_3], axis=-1)
return x0
def mydensenet(blocks_in_dense=2, dense_conv_blocks=2, dense_layers=1, num_dense_connections=256, filters=16,
growth_rate=16, reduction=0.5, **kwargs):
"""
:param blocks_in_dense: how many convolution blocks are in a single size layer
:param dense_conv_blocks: how many dense blocks before a max pooling to occur
:param dense_layers: number of dense layers
:param num_dense_connections:
:param filters:
:param growth_rate:
:param kwargs:
:return:
"""
blocks_in_dense = int(blocks_in_dense)
dense_conv_blocks = int(dense_conv_blocks)
dense_layers = int(dense_layers)
num_dense_connections = int(num_dense_connections)
filters = int(filters)
growth_rate = int(growth_rate)
reduction = float(reduction)
input_shape = (32, 64, 64, 2)
img_input = layers.Input(shape=input_shape)
x = img_input
inputs = (img_input,)
x = layers.Conv3D(filters, (3, 7, 7), strides=2, use_bias=False, name='conv1/conv', padding='Same')(x)
# x = layers.BatchNormalization(axis=-1, epsilon=1.001e-5, name='conv1/bn')(x)
x = GroupNormalization(groups=2, axis=-1, name='conv1/gn')(x)
x = layers.Activation('selu', name='conv1/selu')(x)
for i in range(dense_conv_blocks):
x = dense_block3d(x=x, growth_rate=growth_rate, blocks=blocks_in_dense, name='conv{}'.format(i))
x = transition_block(x=x, reduction=reduction, name='pool{}'.format(i))
# x = layers.BatchNormalization(axis=-1, epsilon=1.001e-5, name='bn')(x)
x = GroupNormalization(groups=2, axis=-1, name='gn')(x)
x = layers.Activation('selu', name='selu')(x)
x = layers.AveragePooling3D(pool_size=(2, 2, 2), name='final_average_pooling')(x)
x = layers.Flatten()(x)
for i in range(dense_layers):
x = layers.Dense(num_dense_connections, activation='selu', kernel_regularizer=regularizers.l2(0.001))(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(2, activation='softmax', name='prediction', dtype='float32')(x)
model = Model(inputs=inputs, outputs=(x,), name='my_3d_densenet')
return model
def fuse_layer2(x, out_filters_list=[32, 64, 128], GROUPS = 18):
# add( identity (x0) | upsample x 2 (x1) | upsample x 4 (x2) ) --> x0
x0_0 = x[0]
x0_1 = Conv2D(out_filters_list[0], 1, use_bias=False, kernel_initializer='he_normal')(x[1])
x0_1 = GroupNormalization(axis=3, groups=GROUPS)(x0_1)
x0_1 = UpSampling2D(size=(2, 2), interpolation="bilinear")(x0_1)
x0_2 = Conv2D(out_filters_list[0], 1, use_bias=False, kernel_initializer='he_normal')(x[2])
x0_2 = GroupNormalization(axis=3, groups=GROUPS)(x0_2)
x0_2 = UpSampling2D(size=(4, 4), interpolation="bilinear")(x0_2)
x0 = add([x0_0, x0_1, x0_2])
# add( downsample x 2 (x0) | identity (x1) | upsample x 2 (x2) ) --> x1
x1_0 = Conv2D(out_filters_list[1], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x[0])
x1_0 = GroupNormalization(axis=3, groups=GROUPS)(x1_0)
x1_1 = x[1]
x1_2 = Conv2D(out_filters_list[1], 1, use_bias=False, kernel_initializer='he_normal')(x[2])
x1_2 = GroupNormalization(axis=3, groups=GROUPS)(x1_2)
x1_2 = UpSampling2D(size=(2, 2), interpolation="bilinear")(x1_2)
x1 = add([x1_0, x1_1, x1_2])
# add( downsample x 4 (x0) | downsample x 2 (x1) | identity (x2) ) --> x2
x2_0 = Conv2D(out_filters_list[0], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x[0])
x2_0 = GroupNormalization(axis=3, groups=GROUPS)(x2_0)
x2_0 = Activation('relu')(x2_0)
x2_0 = Conv2D(out_filters_list[2], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x2_0)
x2_0 = GroupNormalization(axis=3, groups=GROUPS)(x2_0)
x2_1 = Conv2D(out_filters_list[2], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x[1])
x2_1 = GroupNormalization(axis=3, groups=GROUPS)(x2_1)
x2_2 = x[2]
x2 = add([x2_0, x2_1, x2_2])
return [x0, x1, x2]
def __init__(self,
filters,
kernel_size,
gp_num=3,
pad_type="constant",
**kwargs):
super(ResBlock, self).__init__()
padding = (kernel_size - 1) // 2
padding = (padding, padding)
self.model = tf.keras.models.Sequential()
self.model.add(get_padding(pad_type, padding))
self.model.add(GroupNormalization(groups=gp_num, axis=-1))
self.model.add(Conv2D(filters, kernel_size))
self.model.add(ReLU())
self.model.add(get_padding(pad_type, padding))
self.model.add(GroupNormalization(groups=gp_num, axis=-1))
self.model.add(Conv2D(filters, kernel_size))
self.add = Add()
def transition_layer3(x, out_filters_list=[32, 64, 128, 256], GROUPS = 18):
x0 = Conv2D(out_filters_list[0], 3, padding='same', use_bias=False, kernel_initializer='he_normal')(x[0])
x0 = GroupNormalization(axis=3, groups=GROUPS)(x0)
x0 = Activation('relu')(x0)
x1 = Conv2D(out_filters_list[1], 3, padding='same', use_bias=False, kernel_initializer='he_normal')(x[1])
x1 = GroupNormalization(axis=3, groups=GROUPS)(x1)
x1 = Activation('relu')(x1)
x2 = Conv2D(out_filters_list[2], 3, padding='same', use_bias=False, kernel_initializer='he_normal')(x[2])
x2 = GroupNormalization(axis=3, groups=GROUPS)(x2)
x2 = Activation('relu')(x2)
x3 = Conv2D(out_filters_list[3], 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x[2])
x3 = GroupNormalization(axis=3, groups=GROUPS)(x3)
x3 = Activation('relu')(x3)
return [x0, x1, x2, x3]
def __init__(self,
filters=64,
lrelu_alpha=0.2,
pad_type="constant",
gp_num=3,
**kwargs):
super(StridedConv, self).__init__(name="StridedConv")
self.model = tf.keras.models.Sequential()
self.model.add(get_padding(pad_type, (1, 1)))
self.model.add(GroupNormalization(groups=gp_num, axis=-1))
self.model.add(Conv2D(filters, 3, strides=(2, 2)))
self.model.add(LeakyReLU(lrelu_alpha))
self.model.add(get_padding(pad_type, (1, 1)))
self.model.add(GroupNormalization(groups=gp_num, axis=-1))
self.model.add(Conv2D(filters * 2, 3))
self.model.add(LeakyReLU(lrelu_alpha))
def basic_Block(x_input, out_filters, strides=(1, 1), with_conv_shortcut=False, final_activation=True, GROUPS = 18):
x = conv3x3(x_input, out_filters, strides)
x = GroupNormalization(axis=3, groups=GROUPS)(x)
x = Activation('relu')(x)
x = conv3x3(x, out_filters)
x = GroupNormalization(axis=3, groups=GROUPS)(x)
if with_conv_shortcut:
residual = Conv2D(out_filters, 1, strides=strides, use_bias=False, kernel_initializer='he_normal')(x_input)
residual = GroupNormalization(axis=3, groups=GROUPS)(residual)
x = add([x, residual])
else:
x = add([x, x_input])
if final_activation:
x = Activation('relu')(x)
return x
def stem_net(x_input, GROUPS = 32):
x = Conv2D(64, 3, strides=(2, 2), padding='same', use_bias=False, kernel_initializer='he_normal')(x_input)
x = GroupNormalization(axis=3, groups=GROUPS)(x)
x = Activation('relu')(x)
x = bottleneck_Block(x, 256, with_conv_shortcut=True, GROUPS=GROUPS)
x = bottleneck_Block(x, 256, with_conv_shortcut=False, GROUPS=GROUPS)
x = bottleneck_Block(x, 256, with_conv_shortcut=False, GROUPS=GROUPS)
x = bottleneck_Block(x, 256, with_conv_shortcut=False, GROUPS=GROUPS)
return x
def __init__(self,
filters,
regularizer,
data_format='channels_first',
name=None,
**kwargs):
super().__init__(**kwargs)
self.hidden = [
Conv3D(filters=filters,
kernel_size=(1, 1, 1),
strides=1,
kernel_regularizer=regularizer,
data_format=data_format,
name=f'Res_{name}' if name else None),
GroupNormalization(
groups=8,
axis=1 if data_format == 'channels_first' else 0,
name=f'GroupNorm_1_{name}' if name else None),
Activation('relu', name=f'Relu_1_{name}' if name else None),
Conv3D(filters=filters,
kernel_size=(3, 3, 3),
strides=1,
padding='same',
kernel_regularizer=regularizer,
data_format=data_format,
name=f'Conv3D_1_{name}' if name else None),
GroupNormalization(
groups=8,
axis=1 if data_format == 'channels_first' else 0,
name=f'GroupNorm_2_{name}' if name else None),
Activation('relu', name=f'Relu_2_{name}' if name else None),
Conv3D(filters=filters,
kernel_size=(3, 3, 3),
strides=1,
padding='same',
kernel_regularizer=regularizer,
data_format=data_format,
name=f'Conv3D_2_{name}' if name else None),
Add(name=f'Out_{name}' if name else None)
]
def _make_transition_layer(num_channels_pre_layer, num_channels_cur_layer,
GN_GROUPS):
num_branches_pre = len(num_channels_pre_layer)
num_branches_cur = len(num_channels_cur_layer)
transition_layers = []
for i in range(num_branches_cur):
if i < num_branches_pre:
if num_channels_cur_layer[i] != num_channels_pre_layer[i]:
transition_layers.append(
Sequential([
Conv2D(filters=num_channels_cur_layer[i],
kernel_size=(3, 3),
strides=(1, 1),
padding="same",
use_bias=False),
GroupNormalization(groups=GN_GROUPS),
ReLU()
]))
else:
transition_layers.append(None)
else:
conv3x3s = []
for j in range(i + 1 - num_branches_pre):
inchannels = num_channels_pre_layer[-1]
outchannels = num_channels_cur_layer[
i] if j == i - num_branches_pre else inchannels
conv3x3s.append(
Sequential([
Conv2D(filters=outchannels,
kernel_size=(3, 3),
strides=(2, 2),
padding="same",
use_bias=False),
GroupNormalization(groups=GN_GROUPS),
ReLU()
]))
transition_layers.append(Sequential(conv3x3s))
return transition_layers
def __init__(self, c, name=None, drop_rate=0.0, groups=32):
super().__init__(name=name)
self.c = c
self.k = Dense(c)
self.norm = GroupNormalization(groups=groups)
self.proj_out = Dense(c)
self.q = Dense(c)
self.v = Dense(c)
if drop_rate > 0.01:
self.drop_fn = Dropout(drop_rate)
else:
self.drop_fn = tf.identity
def bottleneck_Block(x_input, out_filters, strides=(1, 1), with_conv_shortcut=False, GROUPS = 18):
expansion = 4
de_filters = int(out_filters / expansion)
x = Conv2D(de_filters, 1, use_bias=False, kernel_initializer='he_normal')(x_input)
x = GroupNormalization(axis=3, groups=GROUPS)(x)
x = Activation('relu')(x)
x = Conv2D(de_filters, 3, strides=strides, padding='same', use_bias=False, kernel_initializer='he_normal')(x)
x = GroupNormalization(axis=3, groups=GROUPS)(x)
x = Activation('relu')(x)
x = Conv2D(out_filters, 1, use_bias=False, kernel_initializer='he_normal')(x)
x = GroupNormalization(axis=3, groups=GROUPS)(x)
if with_conv_shortcut:
residual = Conv2D(out_filters, 1, strides=strides, use_bias=False, kernel_initializer='he_normal')(x_input)
residual = GroupNormalization(axis=3, groups=GROUPS)(residual)
x = add([x, residual])
else:
x = add([x, x_input])
x = Activation('relu')(x)
return x
def bottleneck_conv_block(channels=32, spatial_dims=2, norm_groups=4):
'''
Notes
-----
pre-activation to keep the residual path clear as described in:
HE, Kaiming, et al. Identity mappings in deep residual networks.
In: European conference on computer vision. Springer, Cham, 2016.
S. 630-645.
'''
Conv = get_nd_conv(spatial_dims)
conv_in = Conv(channels, kernel_size=1, padding='same')
seq = _stack_layers([
Activation(),
GroupNormalization(groups=norm_groups, axis=-1),
Conv(channels // 2, kernel_size=1, padding='same'),
Activation(),
GroupNormalization(groups=norm_groups, axis=-1),
Conv(channels // 2, kernel_size=3, padding='same'),
Activation(),
GroupNormalization(groups=norm_groups, axis=-1),
Conv(channels, kernel_size=1, padding='same'),
])
def block(x):
if x.shape[-1] != channels:
# if needed, brings the number of input channels to the same as output channels
# not strictly a bottleneck anymore
x = Activation()(conv_in(x))
return x + seq(x)
return block
def __init__(self,
filters,
kernel_size,
gp_num=3,
pad_type="constant",
light=False,
**kwargs):
super(UpSampleConv, self).__init__(name="UpSampleConv")
if light:
self.model = tf.keras.models.Sequential([
GroupNormalization(groups=gp_num, axis=-1),
Conv2D(filters, 1),
BasicShuffleUnitV2(filters, pad_type)
])
else:
self.model = ConvBlock(filters, kernel_size, 1, gp_num, pad_type)