def WideResnetBlock(channels, strides=(1, 1), channel_mismatch=False):
"""WideResnet convolutational block."""
main = tl.Serial(tl.BatchNorm(), tl.Relu(),
tl.Conv(channels, (3, 3), strides, padding='SAME'),
tl.BatchNorm(), tl.Relu(),
tl.Conv(channels, (3, 3), padding='SAME'))
shortcut = tl.Copy() if not channel_mismatch else tl.Conv(
channels, (3, 3), strides, padding='SAME')
return tl.Residual(main, shortcut=shortcut)
def IdentityBlock(kernel_size, filters):
"""ResNet identical size block."""
ks = kernel_size
filters1, filters2, filters3 = filters
main = tl.Serial(tl.Conv(filters1, (1, 1)), tl.BatchNorm(), tl.Relu(),
tl.Conv(filters2, (ks, ks), padding='SAME'),
tl.BatchNorm(), tl.Relu(), tl.Conv(filters3, (1, 1)),
tl.BatchNorm())
return tl.Serial(tl.Residual(main), tl.Relu())
def WideResnetBlock(channels, strides=(1, 1), mode='train'):
"""WideResnet convolutional block."""
return [
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.Conv(channels, (3, 3), strides, padding='SAME'),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.Conv(channels, (3, 3), padding='SAME'),
]
def ConvBlock(kernel_size, filters, strides):
"""ResNet convolutional striding block."""
ks = kernel_size
filters1, filters2, filters3 = filters
main = tl.Serial(tl.Conv(filters1, (1, 1), strides), tl.BatchNorm(),
tl.Relu(), tl.Conv(filters2, (ks, ks), padding='SAME'),
tl.BatchNorm(), tl.Relu(), tl.Conv(filters3, (1, 1)),
tl.BatchNorm())
shortcut = tl.Serial(tl.Conv(filters3, (1, 1), strides), tl.BatchNorm())
return tl.Serial(tl.Residual(main, shortcut=shortcut), tl.Relu())
def WideResnetBlock(channels, strides=(1, 1), channel_mismatch=False):
"""WideResnet convolutational block."""
main = layers.Serial(
layers.BatchNorm(), layers.Relu(),
layers.Conv(channels, (3, 3), strides, padding='SAME'),
layers.BatchNorm(), layers.Relu(),
layers.Conv(channels, (3, 3), padding='SAME'))
shortcut = layers.Identity() if not channel_mismatch else layers.Conv(
channels, (3, 3), strides, padding='SAME')
return layers.Serial(layers.Branch(), layers.Parallel(main, shortcut),
layers.SumBranches())
def IdentityBlock(kernel_size, filters):
"""ResNet identical size block."""
ks = kernel_size
filters1, filters2, filters3 = filters
main = layers.Serial(layers.Conv(filters1, (1, 1)), layers.BatchNorm(),
layers.Relu(),
layers.Conv(filters2, (ks, ks), padding='SAME'),
layers.BatchNorm(), layers.Relu(),
layers.Conv(filters3, (1, 1)), layers.BatchNorm())
return layers.Serial(layers.Branch(),
layers.Parallel(main, layers.Identity()),
layers.SumBranches(), layers.Relu())
def ConvBlock(kernel_size, filters, strides):
"""ResNet convolutional striding block."""
ks = kernel_size
filters1, filters2, filters3 = filters
main = layers.Serial(layers.Conv(filters1, (1, 1), strides),
layers.BatchNorm(), layers.Relu(),
layers.Conv(filters2, (ks, ks), padding='SAME'),
layers.BatchNorm(), layers.Relu(),
layers.Conv(filters3, (1, 1)), layers.BatchNorm())
shortcut = layers.Serial(layers.Conv(filters3, (1, 1), strides),
layers.BatchNorm())
return layers.Serial(layers.Branch(), layers.Parallel(main, shortcut),
layers.SumBranches(), layers.Relu())
def WideResnet(n_blocks=3, widen_factor=1, n_output_classes=10, mode='train'):
"""WideResnet from https://arxiv.org/pdf/1605.07146.pdf.
Args:
n_blocks: int, number of blocks in a group. total layers = 6n + 4.
widen_factor: int, widening factor of each group. k=1 is vanilla resnet.
n_output_classes: int, number of distinct output classes.
mode: Whether we are training or evaluating or doing inference.
Returns:
The list of layers comprising a WideResnet model with the given parameters.
"""
return tl.Model(
tl.ToFloat(),
tl.Conv(16, (3, 3), padding='SAME'),
WideResnetGroup(n_blocks, 16 * widen_factor, mode=mode),
WideResnetGroup(n_blocks, 32 * widen_factor, (2, 2), mode=mode),
WideResnetGroup(n_blocks, 64 * widen_factor, (2, 2), mode=mode),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.AvgPool(pool_size=(8, 8)),
tl.Flatten(),
tl.Dense(n_output_classes),
tl.LogSoftmax(),
)
def model_fn(mode="train"):
return layers.Model(
layers.Dropout(mode=mode, rate=0.1),
layers.BatchNorm(mode=mode),
models.MLP(d_hidden=16,
n_output_classes=n_classes,
mode=mode))
def WideResnet(n_blocks=3, d_hidden=64, n_output_classes=10, mode='train'):
"""WideResnet from https://arxiv.org/pdf/1605.07146.pdf.
Args:
n_blocks: int, number of blocks in a group.
d_hidden: Dimensionality of the first hidden layer (multiplied later).
n_output_classes: int, number of distinct output classes.
mode: Whether we are training or evaluating or doing inference.
Returns:
The list of layers comprising a WideResnet model with the given parameters.
"""
del mode
return tl.Model(
tl.ToFloat(),
tl.Conv(d_hidden, (3, 3), padding='SAME'),
WideResnetGroup(n_blocks, d_hidden),
WideResnetGroup(n_blocks, d_hidden * 2, (2, 2)),
WideResnetGroup(n_blocks, d_hidden * 4, (2, 2)),
tl.BatchNorm(),
tl.Relu(),
tl.AvgPool(pool_size=(8, 8)),
tl.Flatten(),
tl.Dense(n_output_classes),
tl.LogSoftmax(),
)
def WideResnet(num_blocks=3, hidden_size=64, num_output_classes=10,
mode='train'):
"""WideResnet from https://arxiv.org/pdf/1605.07146.pdf.
Args:
num_blocks: int, number of blocks in a group.
hidden_size: the size of the first hidden layer (multiplied later).
num_output_classes: int, number of classes to distinguish.
mode: is it training or eval.
Returns:
The WideResnet model with given layer and output sizes.
"""
del mode
return tl.Serial(
tl.Conv(hidden_size, (3, 3), padding='SAME'),
WideResnetGroup(num_blocks, hidden_size),
WideResnetGroup(num_blocks, hidden_size * 2, (2, 2)),
WideResnetGroup(num_blocks, hidden_size * 4, (2, 2)),
tl.BatchNorm(),
tl.Relu(),
tl.AvgPool(pool_size=(8, 8)),
tl.Flatten(),
tl.Dense(num_output_classes),
tl.LogSoftmax()
)
def IdentityBlock(kernel_size, filters, mode='train'):
"""ResNet identical size block."""
# TODO(jonni): Use good defaults so Resnet50 code is cleaner / less redundant.
ks = kernel_size
filters1, filters2, filters3 = filters
main = [
tl.Conv(filters1, (1, 1)),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.Conv(filters2, (ks, ks), padding='SAME'),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.Conv(filters3, (1, 1)),
tl.BatchNorm(mode=mode),
]
return [
tl.Residual(main),
tl.Relu(),
]
def Resnet50(d_hidden=64, n_output_classes=1001, mode='train'):
"""ResNet.
Args:
d_hidden: Dimensionality of the first hidden layer (multiplied later).
n_output_classes: Number of distinct output classes.
mode: Whether we are training or evaluating or doing inference.
Returns:
The list of layers comprising a ResNet model with the given parameters.
"""
return tl.Model(
tl.ToFloat(),
tl.Conv(d_hidden, (7, 7), (2, 2), 'SAME'),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.MaxPool(pool_size=(3, 3), strides=(2, 2)),
ConvBlock(3, [d_hidden, d_hidden, 4 * d_hidden], (1, 1), mode=mode),
IdentityBlock(3, [d_hidden, d_hidden, 4 * d_hidden], mode=mode),
IdentityBlock(3, [d_hidden, d_hidden, 4 * d_hidden], mode=mode),
ConvBlock(3, [2 * d_hidden, 2 * d_hidden, 8 * d_hidden], (2, 2),
mode=mode),
IdentityBlock(3, [2 * d_hidden, 2 * d_hidden, 8 * d_hidden],
mode=mode),
IdentityBlock(3, [2 * d_hidden, 2 * d_hidden, 8 * d_hidden],
mode=mode),
IdentityBlock(3, [2 * d_hidden, 2 * d_hidden, 8 * d_hidden],
mode=mode),
ConvBlock(3, [4 * d_hidden, 4 * d_hidden, 16 * d_hidden], (2, 2),
mode=mode),
IdentityBlock(3, [4 * d_hidden, 4 * d_hidden, 16 * d_hidden],
mode=mode),
IdentityBlock(3, [4 * d_hidden, 4 * d_hidden, 16 * d_hidden],
mode=mode),
IdentityBlock(3, [4 * d_hidden, 4 * d_hidden, 16 * d_hidden],
mode=mode),
IdentityBlock(3, [4 * d_hidden, 4 * d_hidden, 16 * d_hidden],
mode=mode),
IdentityBlock(3, [4 * d_hidden, 4 * d_hidden, 16 * d_hidden],
mode=mode),
ConvBlock(3, [8 * d_hidden, 8 * d_hidden, 32 * d_hidden], (2, 2),
mode=mode),
IdentityBlock(3, [8 * d_hidden, 8 * d_hidden, 32 * d_hidden],
mode=mode),
IdentityBlock(3, [8 * d_hidden, 8 * d_hidden, 32 * d_hidden],
mode=mode),
tl.AvgPool(pool_size=(7, 7)),
tl.Flatten(),
tl.Dense(n_output_classes),
tl.LogSoftmax(),
)
def ConvBlock(kernel_size, filters, strides, mode='train'):
"""ResNet convolutional striding block."""
# TODO(jonni): Use good defaults so Resnet50 code is cleaner / less redundant.
ks = kernel_size
filters1, filters2, filters3 = filters
main = [
tl.Conv(filters1, (1, 1), strides),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.Conv(filters2, (ks, ks), padding='SAME'),
tl.BatchNorm(mode=mode),
tl.Relu(),
tl.Conv(filters3, (1, 1)),
tl.BatchNorm(mode=mode),
]
shortcut = [
tl.Conv(filters3, (1, 1), strides),
tl.BatchNorm(mode=mode),
]
return [
tl.Residual(main, shortcut=shortcut),
tl.Relu(),
]
def Resnet50(hidden_size=64, num_output_classes=1001, mode='train'):
"""ResNet.
Args:
hidden_size: the size of the first hidden layer (multiplied later).
num_output_classes: how many classes to distinguish.
mode: whether we are training or evaluating or doing inference.
Returns:
The ResNet model with the given layer and output sizes.
"""
del mode
return tl.Serial(
tl.Conv(hidden_size, (7, 7), (2, 2),
'SAME'), tl.BatchNorm(), tl.Relu(),
tl.MaxPool(pool_size=(3, 3), strides=(2, 2)),
ConvBlock(3, [hidden_size, hidden_size, 4 * hidden_size], (1, 1)),
IdentityBlock(3, [hidden_size, hidden_size, 4 * hidden_size]),
IdentityBlock(3, [hidden_size, hidden_size, 4 * hidden_size]),
ConvBlock(3, [2 * hidden_size, 2 * hidden_size, 8 * hidden_size],
(2, 2)),
IdentityBlock(3, [2 * hidden_size, 2 * hidden_size, 8 * hidden_size]),
IdentityBlock(3, [2 * hidden_size, 2 * hidden_size, 8 * hidden_size]),
IdentityBlock(3, [2 * hidden_size, 2 * hidden_size, 8 * hidden_size]),
ConvBlock(3, [4 * hidden_size, 4 * hidden_size, 16 * hidden_size],
(2, 2)),
IdentityBlock(3, [4 * hidden_size, 4 * hidden_size, 16 * hidden_size]),
IdentityBlock(3, [4 * hidden_size, 4 * hidden_size, 16 * hidden_size]),
IdentityBlock(3, [4 * hidden_size, 4 * hidden_size, 16 * hidden_size]),
IdentityBlock(3, [4 * hidden_size, 4 * hidden_size, 16 * hidden_size]),
IdentityBlock(3, [4 * hidden_size, 4 * hidden_size, 16 * hidden_size]),
ConvBlock(3, [8 * hidden_size, 8 * hidden_size, 32 * hidden_size],
(2, 2)),
IdentityBlock(3, [8 * hidden_size, 8 * hidden_size, 32 * hidden_size]),
IdentityBlock(3, [8 * hidden_size, 8 * hidden_size, 32 * hidden_size]),
tl.AvgPool(pool_size=(7, 7)), tl.Flatten(),
tl.Dense(num_output_classes), tl.LogSoftmax())
def model_fn(mode="train"):
return layers.Model(layers.Dropout(mode=mode, rate=0.1),
layers.BatchNorm(mode=mode),
mlp(mode=mode))
