def __init__(self,
channels,
in_channels=3,
hidden_size=4096,
dropout_rate=.5,
num_classes=1000,
data_format='channels_last',
**kwargs):
super(VGG, self).__init__()
features = tf.keras.Sequential(name='feature_extractor')
for i, channels_per_stage in enumerate(channels):
for j, out_channels in enumerate(channels_per_stage):
features.add(
ConvBlock(in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
strides=1,
padding=1,
data_format=data_format,
name='features/stage{}/conv{}'.format(i, j)))
in_channels = out_channels
features.add(
MaxPool2d(pool_size=2,
strides=2,
padding=0,
ceil_mode=False,
data_format=data_format,
name='features/stage{}/pool'.format(i)))
features.add(Flatten(data_format=data_format))
self.features = features
self.classifier = Classifier(hidden_size=hidden_size,
dropout_rate=dropout_rate,
num_classes=num_classes)
def __init__(self,
channels,
residuals,
init_kernel_size,
init_channels,
in_channels=3,
input_shape=(224, 224),
num_classes=1000,
dropout_rate=.5,
data_format='channels_last',
name='squeezenet'):
super(SqueezeNet, self).__init__()
self.features = tf.keras.Sequential(name=name + '/features')
# init conv
self.features.add(
Conv2d(in_channels=in_channels,
out_channels=init_channels,
kernel_size=init_kernel_size,
strides=2,
use_bias=True,
data_format=data_format,
name=name + '/features/init_conv'))
self.features.add(ReLU(name=name + '/features/init_activ'))
# stages of fire blocks
in_channels = init_channels
for i, channels_per_stage in enumerate(channels):
self.features.add(
MaxPool2d(pool_size=3,
strides=2,
ceil_mode=True,
data_format=data_format,
name=name + '/features/stage{}/pool'.format(i)))
for j, out_channels in enumerate(channels_per_stage):
squeeze_channels = out_channels // 8
expand_channels = out_channels // 2
self.features.add(
Fire(in_channels=in_channels,
squeeze_channels=squeeze_channels,
expand1x1_channels=expand_channels,
expand3x3_channels=expand_channels,
residual=(residuals is not None)
and (residuals[i][j] == 1),
data_format=data_format,
name=name + '/features/stage{}/fire{}'.format(i, j)))
in_channels = out_channels
# classifier head
self.classifier = tf.keras.Sequential([
Dropout(rate=dropout_rate, name=name + '/classifier/dropout'),
Conv2d(in_channels=in_channels,
out_channels=num_classes,
kernel_size=1,
data_format=data_format,
name=name + '/classifier/final_conv'),
ReLU(name=name + '/classifier/activ'),
GlobalAveragePooling2D(name=name + '/classifier/pool'),
Flatten(name=name + '/classifier/flatten')
])
def __init__(self, bn_size=4, growth_rate=32, blocks=(6, 12, 24, 16), in_channels=3, init_channels=64,
dropout_rate=.1, num_classes=1000, data_format='channels_last', name='densenet'):
super(DenseNet, self).__init__(name=name)
self.init = tf.keras.Sequential(
layers=[
Conv2d(in_channels=in_channels, out_channels=init_channels, kernel_size=7, strides=2, padding=3,
use_bias=False, data_format=data_format, name=name + '/init/conv0'),
BatchNormalization(axis=-1 if data_format == 'channels_last' else 1, name=name + '/init/bn'),
ReLU(name=name + '/init/activ'),
MaxPool2d(pool_size=3, strides=2, padding=1, data_format=data_format, name=name + '/init/pool')
], name=name + '/init'
)
self.dense_blocks = tf.keras.Sequential(name=name + '/dense_blocks')
num_channels = init_channels
for i, num_layers in enumerate(blocks, 1):
dense_block = DenseBlock(
in_channels=num_channels,
num_layers=num_layers,
bn_size=bn_size,
growth_rate=growth_rate,
dropout_rate=dropout_rate,
data_format=data_format,
name=name + 'dense_blocks/dense_block{}'.format(i)
)
self.dense_blocks.add(dense_block)
num_channels += num_layers * growth_rate
if i != len(blocks):
transit_block = TransitBlock(
in_channels=num_channels,
out_channels=num_channels // 2,
data_format=data_format,
name=name + 'dense_blocks/transit_block{}'.format(i)
)
num_channels = num_channels // 2
self.dense_blocks.add(transit_block)
self.features = tf.keras.Sequential(
layers=[
BatchNormalization(axis=-1 if data_format == 'channels_last' else 1, name=name + '/features/bn'),
ReLU(name=name + '/features/activ'),
GlobalAveragePooling2D(data_format=data_format, name=name + '/features/final_pool')
], name=name + '/features'
)
self.classifier = Dense(units=num_classes, name=name + '/classifier')
def __init__(self,
in_channels,
ch1x1,
ch3x3red,
ch3x3,
ch5x5red,
ch5x5,
pool_proj,
data_format='channels_last',
name='inception'):
super(InceptionBlock, self).__init__()
self.branch1 = ConvBlock(in_channels,
ch1x1,
1,
data_format,
name=name + '/1x1conv')
self.branch2 = tf.keras.Sequential([
ConvBlock(in_channels,
ch3x3red,
1,
data_format,
name=name + '/3x3reduction'),
ConvBlock(ch3x3red,
ch3x3,
3,
data_format,
name=name + '/3x3conv',
padding=1)
])
self.branch3 = tf.keras.Sequential([
ConvBlock(in_channels,
ch5x5red,
1,
data_format,
name=name + '/5x5reduction'),
ConvBlock(ch5x5red,
ch5x5,
5,
data_format,
name=name + '/5x5conv',
padding=2)
])
self.branch4 = tf.keras.Sequential([
MaxPool2d(pool_size=3,
strides=1,
padding=1,
ceil_mode=True,
data_format=data_format,
name=name + '/3x3pool'),
ConvBlock(in_channels,
pool_proj,
1,
data_format,
name=name + '/pool_projection')
])
self.concat = tf.keras.layers.Concatenate(
axis=-1 if data_format == 'channels_last' else 1,
name=name + 'concat')
def __init__(self,
in_channels=3,
data_format='channels_last',
aux_logits=True,
num_classes=1000):
super(GoogleNet, self).__init__()
self.aux_logits = aux_logits
self.conv1 = ConvBlock(in_channels,
64,
7,
data_format,
name='conv1',
strides=2,
padding=3)
self.maxpool1 = MaxPool2d(3,
strides=2,
ceil_mode=True,
data_format=data_format,
name='pool1')
self.conv2 = ConvBlock(64, 64, 1, data_format, name='conv2')
self.conv3 = ConvBlock(64, 192, 3, data_format, name='conv3')
self.maxpool2 = MaxPool2d(3,
strides=2,
ceil_mode=True,
data_format=data_format,
name='pool2')
self.inception3a = InceptionBlock(192,
64,
96,
128,
16,
32,
32,
data_format,
name='inception3a')
self.inception3b = InceptionBlock(256,
128,
128,
192,
32,
96,
64,
data_format,
name='inception3b')
self.maxpool3 = MaxPool2d(3,
strides=2,
ceil_mode=True,
data_format=data_format,
name='pool3')
self.inception4a = InceptionBlock(480,
192,
96,
208,
16,
48,
64,
data_format,
name='inception4a')
self.inception4b = InceptionBlock(512,
160,
112,
224,
24,
64,
64,
data_format,
name='inception4b')
self.inception4c = InceptionBlock(512,
128,
128,
256,
24,
64,
64,
data_format,
name='inception4c')
self.inception4d = InceptionBlock(512,
112,
144,
288,
32,
64,
64,
data_format,
name='inception4d')
self.inception4e = InceptionBlock(528,
256,
160,
320,
32,
128,
128,
data_format,
name='inception4e')
self.maxpool4 = MaxPool2d(2,
strides=2,
ceil_mode=True,
data_format=data_format,
name='pool4')
self.inception5a = InceptionBlock(832,
256,
160,
320,
32,
128,
128,
data_format,
name='inception5a')
self.inception5b = InceptionBlock(832,
384,
192,
384,
48,
128,
128,
data_format,
name='inception5b')
self.avgpool = tf.keras.layers.GlobalAveragePooling2D(data_format,
name='avgpool')
self.aux_classifier1 = InceptionAux(512,
num_classes,
data_format,
name='aux1')
self.aux_classifier2 = InceptionAux(528,
num_classes,
data_format,
name='aux2')
self.classifier = tf.keras.Sequential([
tf.keras.layers.Dropout(rate=.2, name='classifier/dropout'),
tf.keras.layers.Dense(num_classes, name='classifier/fc')
])
def __init__(self,
in_channels=3,
num_classes=1000,
n_groups=2,
se_reduction=None,
residual=False,
init_channels=INIT_CHANNELS,
out_channels=OUT_CHANNELS,
c=C,
n=N,
data_format='channels_last',
name='ShuffleNetV2'):
super(ShuffleNetV2, self).__init__(name=name)
self.init_conv = tf.keras.Sequential([
Conv2d(in_channels,
init_channels,
kernel_size=3,
strides=2,
padding=1,
use_bias=False,
data_format=data_format,
name=name + '/init/conv0'),
BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/init/bn0'),
ReLU(name=name + '/init/activ0'),
MaxPool2d(pool_size=3,
strides=2,
padding=1,
data_format=data_format,
name=name + '/init/pool0')
])
self.features = tf.keras.Sequential()
input_channels = init_channels
for i, (nn, nc) in enumerate(zip(n, c)):
stage = tf.keras.Sequential(name=name + '/stage{}'.format(i))
for j in range(nn):
if j == 0:
stage.add(
ShuffleNetV2DownsampleUnit(input_channels, nc,
n_groups, se_reduction,
data_format))
else:
stage.add(
ShuffleNetV2BasicUnit(nc, n_groups, se_reduction,
residual, data_format))
input_channels = nc
self.features.add(stage)
final_conv = tf.keras.Sequential([
Conv2d(input_channels,
out_channels,
kernel_size=1,
strides=1,
padding=0,
use_bias=False,
data_format=data_format,
name=name + '/final_conv'),
BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/final_bn'),
ReLU(name=name + '/final_activ')
])
self.features.add(final_conv)
self.features.add(
AveragePooling2D(pool_size=7,
data_format=data_format,
name='global_avg_pool'))
self.classifier = tf.keras.Sequential([
Conv2d(out_channels,
num_classes,
kernel_size=1,
strides=1,
padding=0,
use_bias=True,
data_format=data_format,
name='classifier'),
Flatten()
])
def __init__(self,
res_block,
layers,
in_channels,
num_classes=1000,
groups=1,
width_per_group=64,
dilation_for_stride=None,
se_reduction=0,
data_format='channels_last',
name='resnet'):
super(ResNet, self).__init__()
self.channels = 64
self.dilation = 1
if dilation_for_stride is None:
dilation_for_stride = [False, False, False]
self.se_reduction = se_reduction
self.groups = groups
self.base_width = width_per_group
self.stage0 = tf.keras.Sequential([
Conv2d(in_channels,
self.channels,
kernel_size=7,
strides=2,
padding=3,
use_bias=False,
name=name + '/stage0/conv1'),
BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/stage0/bn1'),
ReLU(name=name + '/stage0/relu'),
MaxPool2d(pool_size=3,
strides=2,
padding=1,
data_format=data_format,
name=name + '/stage0/maxpool')
])
self.stage1 = self._make_layer(block=res_block,
channels=64,
blocks=layers[0],
se_reduction=se_reduction)
self.stage2 = self._make_layer(block=res_block,
channels=128,
blocks=layers[1],
strides=2,
dilation=dilation_for_stride[0],
se_reduction=se_reduction)
self.stage3 = self._make_layer(block=res_block,
channels=256,
blocks=layers[2],
strides=2,
dilation=dilation_for_stride[1],
se_reduction=se_reduction)
self.stage4 = self._make_layer(block=res_block,
channels=512,
blocks=layers[3],
strides=2,
dilation=dilation_for_stride[2],
se_reduction=se_reduction)
self.globalavgpool = GlobalAveragePooling2D(data_format=data_format,
name='features')
self.fc = Dense(units=num_classes, name='last_linear')