def __init__(self, num_blocks=4, nc32=8, nc16=14, nc8=28, k=1,
drop_ratio=0.0):
"""
:param num_blocks: the number of resnet blocks per stage. There are 3
stages, for feature map width 32, 16, 8.
Total number of layers is 6 * num_blocks + 2
:param nc32: the number of feature maps in the first stage (where
feature maps are 32x32)
:param nc16: the number of feature maps in the second stage (where
feature maps are 16x16)
:param nc8: the number of feature maps in the third stage (where
feature maps are 8x8)
"""
ws = sqrt(2.) # This makes the initialization equal to He et al.
super(P6WideResNet, self).__init__()
# The first layer is always a convolution.
self.add_link(
P6ConvZ2(in_channels=3, out_channels=nc32, ksize=3, stride=1,
pad=1, wscale=ws)
)
# Add num_blocks ResBlocks (2n layers) for the size 32x32 feature maps
for i in range(num_blocks):
nc_in = nc32 * k if i > 0 else nc32
self.add_link(
WideResBlock2D(
in_channels=nc_in, out_channels=nc32 * k, wscale=ws,
downsample=False, ConvLink=P6ConvP6,
drop_ratio=drop_ratio))
# Add num_blocks ResBlocks (2n layers) for the size 16x16 feature maps
# The first convolution uses stride 2
for i in range(num_blocks):
nc_in = nc16 * k if i > 0 else nc32 * k
downsample = i == 0
self.add_link(
WideResBlock2D(
in_channels=nc_in, out_channels=nc16 * k, wscale=ws,
downsample=downsample, ConvLink=P6ConvP6,
drop_ratio=drop_ratio))
# Add num_blocks ResBlocks (2n layers) for the size 8x8 feature maps
for i in range(num_blocks):
nc_in = nc8 * k if i > 0 else nc16 * k
downsample = i == 0
self.add_link(
WideResBlock2D(
in_channels=nc_in, out_channels=nc8 * k, wscale=ws,
downsample=downsample, ConvLink=P6ConvP6,
drop_ratio=drop_ratio))
# Add BN and final layer
# We do ReLU and average pooling between BN and final layer, but these
# don't require a link.
self.add_link(F.BatchNormalization(size=nc8 * k))
self.add_link(
L.Convolution2D(in_channels=nc8 * k * 6, out_channels=10, ksize=1,
wscale=ws))
def __init__(self, num_blocks=4, nc32=17, nc16=35, nc8=69, k=1,
drop_ratio=0.0):
"""
:param num_blocks: the number of resnet blocks per stage. There are 3
stages, for feature map width 32, 16, 8. Total number of layers is
6 * num_blocks + 2
:param nc32: the number of feature maps in the first stage (where
feature maps are 32x32)
:param nc16: the number of feature maps in the second stage (where
feature maps are 16x16)
:param nc8: the number of feature maps in the third stage (where
feature maps are 8x8)
:param k: the widening factor
"""
ws = np.sqrt(2)
super(WideResNetHex, self).__init__()
# The first layer is always a convolution.
self.add_link(
HexConv2D(in_channels=3, ksize=3, pad=1, stride=1,
out_channels=nc32, wscale=ws))
# Add num_blocks ResBlocks (2n layers) for the size 32x32 feature maps
for i in range(num_blocks):
nc_in = nc32 * k if i > 0 else nc32
self.add_link(
WideResBlock2D(
in_channels=nc_in, out_channels=nc32 * k, wscale=ws,
downsample=False, ConvLink=HexConv2D, drop_ratio=drop_ratio))
# Add num_blocks ResBlocks (2n layers) for the size 16x16 feature maps
# The first convolution uses stride 2
for i in range(num_blocks):
nc_in = nc16 * k if i > 0 else nc32 * k
downsample = i == 0
self.add_link(
WideResBlock2D(
in_channels=nc_in, out_channels=nc16 * k, wscale=ws,
downsample=downsample, ConvLink=HexConv2D,
drop_ratio=drop_ratio))
# Add num_blocks ResBlocks (2n layers) for the size 8x8 feature maps
for i in range(num_blocks):
nc_in = nc8 * k if i > 0 else nc16 * k
downsample = i == 0
self.add_link(
WideResBlock2D(
in_channels=nc_in, out_channels=nc8 * k, wscale=ws,
downsample=downsample, ConvLink=HexConv2D,
drop_ratio=drop_ratio))
# Add BN and final layer
# We do ReLU and average pooling between BN and final layer, but
# these don't require a link.
self.add_link(F.BatchNormalization(size=nc8 * k))
self.add_link(
L.Convolution2D(
in_channels=nc8 * k, ksize=1, out_channels=10, wscale=ws))