in models/imagenet/resnet_cnsn.py [0:0]
def __init__(self, layers, num_classes=1000, zero_init_residual=False,
groups=1, width_per_group=64, replace_stride_with_dilation=None,
active_num=1, pos=None, beta=None, crop=None, cnsn_type=None):
super(ResNet, self).__init__()
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 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
if beta is not None:
print('beta: {}'.format(beta))
if crop is not None:
print('crop mode: {}'.format(crop))
self.layer1 = self._make_layer_custom(BottleneckCustom, 64, layers[0],
pos=pos, beta=beta,
crop=crop, cnsn_type=cnsn_type)
self.layer2 = self._make_layer_custom(BottleneckCustom, 128, layers[1],
pos=pos, beta=beta,
crop=crop, cnsn_type=cnsn_type,
stride=2, dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer_custom(BottleneckCustom, 256, layers[2],
pos=pos, beta=beta,
crop=crop, cnsn_type=cnsn_type,
stride=2, dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer_custom(BottleneckCustom, 512, layers[3],
pos=pos, beta=beta,
crop=crop, cnsn_type=cnsn_type,
stride=2, dilate=replace_stride_with_dilation[2])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(512 * BottleneckCustom.expansion, num_classes)
self.cn_modules = []
for m in self.modules():
if isinstance(m, nn.Conv2d):
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)
elif isinstance(m, CrossNorm):
self.cn_modules.append(m)
if cnsn_type is not None and 'cn' in cnsn_type:
self.cn_num = len(self.cn_modules)
assert self.cn_num > 0
print('cn_num: {}'.format(self.cn_num))
self.active_num = active_num
assert self.active_num > 0
print('active_num: {}'.format(self.active_num))
# 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, BottleneckCustom):
nn.init.constant_(m.bn3.weight, 0)