def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
lv_init=-5,
var_p=-1):
super(Bottleneck, self).__init__()
self.bn1 = nn.BatchNorm2d(inplanes)
self.conv1 = vi.BayesConv2d(inplanes,
planes,
kernel_size=1,
bias=False,
lv_init=lv_init,
var_p=var_p)
self.bn2 = nn.BatchNorm2d(planes)
self.conv2 = vi.BayesConv2d(planes,
planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False,
lv_init=lv_init,
var_p=var_p)
self.bn3 = nn.BatchNorm2d(planes)
self.conv3 = vi.BayesConv2d(planes,
planes * 4,
kernel_size=1,
bias=False,
lv_init=lv_init,
var_p=var_p)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def __init__(self, in_planes, planes, dropout_rate, stride=1, lv_init=-5, var_p=-1):
super(WideBasic, self).__init__()
self.bn1 = nn.BatchNorm2d(in_planes)
self.conv1 = vi.BayesConv2d(in_planes, planes, kernel_size=3, padding=1, bias=True, lv_init=lv_init, var_p=var_p)
self.dropout = nn.Dropout(p=dropout_rate)
self.bn2 = nn.BatchNorm2d(planes)
self.conv2 = vi.BayesConv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True, lv_init=lv_init, var_p=var_p)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != planes:
self.shortcut = nn.Sequential(
vi.BayesConv2d(in_planes, planes, kernel_size=1, stride=stride, bias=True, lv_init=lv_init, var_p=var_p),
)
def conv1x1(in_planes, out_planes, stride=1):
"""1x1 convolution"""
return vi.BayesConv2d(in_planes,
out_planes,
kernel_size=1,
stride=stride,
bias=False)
def _make_layer(self,
block,
planes,
blocks,
stride=1,
lv_init=-5,
var_p=-1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
vi.BayesConv2d(self.inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False,
lv_init=lv_init,
var_p=var_p), )
layers = list()
layers.append(
block(self.inplanes,
planes,
stride,
downsample,
lv_init=lv_init,
var_p=var_p))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(self.inplanes, planes, lv_init=lv_init, var_p=var_p))
return nn.Sequential(*layers)
def conv3x3(in_planes, out_planes, stride=1, lv_init=-6, var_p=-1):
return vi.BayesConv2d(in_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=1,
bias=True,
lv_init=lv_init,
var_p=var_p)
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
"""3x3 convolution with padding"""
return vi.BayesConv2d(in_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=dilation,
groups=groups,
bias=False,
dilation=dilation)
def __init__(self, num_classes=10, depth=110, lv_init=-5, var_p=-1):
super(PreResNet, self).__init__()
if depth >= 44:
assert (depth - 2) % 9 == 0, 'depth should be 9n+2'
n = (depth - 2) // 9
block = Bottleneck
else:
assert (depth - 2) % 6 == 0, 'depth should be 6n+2'
n = (depth - 2) // 6
block = BasicBlock
self.var_p = var_p
self.inplanes = 16
self.conv1 = vi.BayesConv2d(3,
16,
kernel_size=3,
padding=1,
bias=False,
var_p=var_p)
self.layer1 = self._make_layer(block,
16,
n,
lv_init=lv_init,
var_p=var_p)
self.layer2 = self._make_layer(block,
32,
n,
stride=2,
lv_init=lv_init,
var_p=var_p)
self.layer3 = self._make_layer(block,
64,
n,
stride=2,
lv_init=lv_init,
var_p=var_p)
self.bn = nn.BatchNorm2d(64 * block.expansion)
self.relu = nn.ReLU(inplace=True)
self.avgpool = nn.AvgPool2d(8)
self.fc = vi.BayesLinear(64 * block.expansion,
num_classes,
var_p=var_p)
for m in self.modules():
if isinstance(m, vi.BayesConv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def __init__(self, block, layers, num_classes=1000, zero_init_residual=False,
groups=1, width_per_group=64, replace_stride_with_dilation=None,
norm_layer=None, lv_init=-1, var_p=-1):
super(ResNet, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.inplanes = 64
self.dilation = 1
if replace_stride_with_dilation is None:
# each element in the tuple indicates if we should replace
# the 2x2 stride with a dilated convolution instead
replace_stride_with_dilation = [False, False, False]
if len(replace_stride_with_dilation) != 3:
raise ValueError("replace_stride_with_dilation should be None "
"or a 3-element tuple, got {}".format(replace_stride_with_dilation))
self.groups = groups
self.base_width = width_per_group
self.conv1 = vi.BayesConv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
bias=False, lv_init=lv_init, var_p=var_p)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0], lv_init=lv_init, var_p=var_p)
self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
dilate=replace_stride_with_dilation[0], lv_init=lv_init, var_p=var_p)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
dilate=replace_stride_with_dilation[1], lv_init=lv_init, var_p=var_p)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
dilate=replace_stride_with_dilation[2], lv_init=lv_init, var_p=var_p)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = vi.BayesLinear(512 * block.expansion, num_classes, lv_init=lv_init, var_p=var_p)
for m in self.modules():
if isinstance(m, vi.BayesConv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
def make_layers(cfg, batch_norm=False, lv_init=-5, var_p=-1):
layers = list()
in_channels = 3
#print('make_layers', lv_init)
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = vi.BayesConv2d(in_channels,
v,
kernel_size=3,
padding=1,
lv_init=lv_init,
var_p=var_p)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)]
else:
layers += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return nn.Sequential(*layers)
def conv1x1(in_planes, out_planes, stride=1, lv_init=-1, var_p=-1):
"""1x1 convolution"""
return vi.BayesConv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False, lv_init=lv_init, var_p=var_p)