import math
from torch import nn
from torch.nn import functional as F
from backbone import MammothBackbone, register_backbone
EPS_BATCH_NORM = 1e-4
[docs]
def conv3x3(in_planes, out_planes, stride=1):
"3x3 convolution with padding"
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
[docs]
class IdentityShortcut(nn.Module):
def __init__(self, planes):
super(IdentityShortcut, self).__init__()
self.planes = planes
self.pad_dimension = planes // 4
[docs]
def forward(self, x):
x = x[:, :, ::2, ::2]
# add padding to the channels dimension to match the output of the residual
return F.pad(x, (0, 0, 0, 0, self.pad_dimension, self.pad_dimension))
[docs]
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, downsample=None, last=False):
super(BasicBlock, self).__init__()
self.last = last
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = nn.BatchNorm2d(planes, eps=EPS_BATCH_NORM)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = nn.BatchNorm2d(planes, eps=EPS_BATCH_NORM)
self.downsample = downsample
self.stride = stride
[docs]
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.relu(out)
out = self.bn1(out)
out = self.conv2(out)
# out = self.relu(out)
out = self.bn2(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
if not self.last:
out = self.relu(out)
return out
[docs]
class ResNet32(MammothBackbone):
def __init__(self, depth=32, num_classes=1000):
super().__init__()
assert (depth - 2) % 6 == 0, 'When use basicblock, depth should be 6n+2, e.g. 20, 32, 44, 56, 110, 1202'
n = (depth - 2) // 6
self.inplanes = 16
self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1, bias=False)
self.bn_initial = nn.BatchNorm2d(16, eps=EPS_BATCH_NORM)
self.layer1 = self._make_layer(BasicBlock, 16, n)
self.layer2 = self._make_layer(BasicBlock, 32, n, stride=2)
self.layer3 = self._make_layer(BasicBlock, 64, n, stride=2, last=True)
self.bn_final = nn.BatchNorm2d(64 * BasicBlock.expansion, eps=EPS_BATCH_NORM)
self.relu = nn.ReLU(inplace=True)
self.avgpool = nn.AvgPool2d(8)
self.classifier = nn.Linear(64 * BasicBlock.expansion, num_classes)
gain = math.sqrt(2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.in_channels # m.kernel_size[0] * m.kernel_size[1] * m.out_channels
# theano init HeNormal with gain='relu' (from Kaiming He et al. (2015): Delving deep into rectifiers: Surpassing human-level performance on imagenet classification)
m.weight.data.normal_(0, gain * math.sqrt(1. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
self.classifier.weight.data.normal_(0, math.sqrt(1. / 64))
self.classifier.bias.data.zero_()
# self.classifier.apply(xavier)
def _make_layer(self, block, planes, blocks, stride=1, last=False):
downsample = None
if stride != 1:
downsample = IdentityShortcut(planes)
# downsample = nn.Sequential(
# nn.Conv2d(self.inplanes, planes * block.expansion,
# kernel_size=1, stride=stride, bias=False),
# )
if last:
blocks -= 1
layers = [block(self.inplanes, planes, stride, downsample)]
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes))
if last:
layers.append(block(self.inplanes, planes, last=True))
return nn.Sequential(*layers)
[docs]
def forward(self, x, returnt='out'):
x = self.bn_initial(self.relu(self.conv1(x)))
x = self.layer1(x) # 32x32
x = self.layer2(x) # 16x16
x = self.layer3(x) # 8x8
x = self.avgpool(x)
features = x.view(x.size(0), -1)
if returnt == 'features':
return features
out = self.classifier(features)
if returnt == 'both':
return (out, features)
return out
[docs]
@register_backbone('resnet32')
def resnet32(num_classes: int, depth: int = 32):
"""
Constructs a ResNet-32 model, as used in 'iCaRL'.
"""
return ResNet32(num_classes=num_classes, depth=depth)