def WideResnet(n_blocks=3, widen_factor=1, n_output_classes=10, bn_momentum=0.9,
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.
bn_momentum: float, momentum in BatchNorm.
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.Serial(
tl.ToFloat(),
tl.Conv(16, (3, 3), padding='SAME'),
WideResnetGroup(n_blocks, 16 * widen_factor, bn_momentum=bn_momentum,
mode=mode),
WideResnetGroup(n_blocks, 32 * widen_factor, (2, 2),
bn_momentum=bn_momentum, mode=mode),
WideResnetGroup(n_blocks, 64 * widen_factor, (2, 2),
bn_momentum=bn_momentum, mode=mode),
tl.BatchNorm(momentum=bn_momentum, mode=mode),
tl.Relu(),
tl.AvgPool(pool_size=(8, 8)),
tl.Flatten(),
tl.Dense(n_output_classes),
tl.LogSoftmax(),
)
def Resnet50(d_hidden=64,
n_output_classes=1001,
mode='train',
norm=tl.BatchNorm,
non_linearity=tl.Relu):
"""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.
norm: `Layer` used for normalization, Ex: BatchNorm or
FilterResponseNorm.
non_linearity: `Layer` used as a non-linearity, Ex: If norm is
BatchNorm then this is a Relu, otherwise for FilterResponseNorm this
should be ThresholdedLinearUnit.
Returns:
The list of layers comprising a ResNet model with the given parameters.
"""
# A ConvBlock configured with the given norm, non-linearity and mode.
def Resnet50ConvBlock(filter_multiplier=1, strides=(2, 2)):
filters = ([
filter_multiplier * dim
for dim in [d_hidden, d_hidden, 4 * d_hidden]
])
return ConvBlock(3, filters, strides, norm, non_linearity, mode)
# Same as above for IdentityBlock.
def Resnet50IdentityBlock(filter_multiplier=1):
filters = ([
filter_multiplier * dim
for dim in [d_hidden, d_hidden, 4 * d_hidden]
])
return IdentityBlock(3, filters, norm, non_linearity, mode)
return tl.Serial(
tl.ToFloat(),
tl.Conv(d_hidden, (7, 7), (2, 2), 'SAME'),
norm(mode=mode),
non_linearity(),
tl.MaxPool(pool_size=(3, 3), strides=(2, 2)),
Resnet50ConvBlock(strides=(1, 1)),
[Resnet50IdentityBlock() for _ in range(2)],
Resnet50ConvBlock(2),
[Resnet50IdentityBlock(2) for _ in range(3)],
Resnet50ConvBlock(4),
[Resnet50IdentityBlock(4) for _ in range(5)],
Resnet50ConvBlock(8),
[Resnet50IdentityBlock(8) for _ in range(2)],
tl.AvgPool(pool_size=(7, 7)),
tl.Flatten(),
tl.Dense(n_output_classes),
tl.LogSoftmax(),
)
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 get_model(n_output_classes=10):
"""
Simple CNN to classify Fashion MNIST
"""
model = tl.Serial(
tl.ToFloat(),
tl.Conv(32, (3, 3), (1, 1), "SAME"),
tl.LayerNorm(),
tl.Relu(),
tl.MaxPool(),
tl.Conv(64, (3, 3), (1, 1), "SAME"),
tl.LayerNorm(),
tl.Relu(),
tl.MaxPool(),
tl.Flatten(),
tl.Dense(n_output_classes),
)
return model
def AtariCnn(n_frames=4, hidden_sizes=(32, 32), output_size=128, mode='train'):
"""An Atari CNN."""
del mode
# TODO(jonni): Include link to paper?
# Input shape: (B, T, H, W, C)
# Output shape: (B, T, output_size)
return tl.Model(
tl.ToFloat(),
tl.Div(divisor=255.0),
# Set up n_frames successive game frames, concatenated on the last axis.
FrameStack(n_frames=n_frames), # (B, T, H, W, 4C)
tl.Conv(hidden_sizes[0], (5, 5), (2, 2), 'SAME'),
tl.Relu(),
tl.Conv(hidden_sizes[1], (5, 5), (2, 2), 'SAME'),
tl.Relu(),
tl.Flatten(n_axes_to_keep=2), # B, T and rest.
tl.Dense(output_size),
tl.Relu(),
)