def __init__(self,
inplanes,
planes,
bitW,
bitA,
stride=1,
downsample=None,
is_last=False):
super(Bottleneck, self).__init__()
self.bitW = bitW
self.bitA = bitA
self.conv1 = QConv2d(inplanes, planes, bitW, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = QConv2d(planes,
planes,
bitW,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = QConv2d(planes,
planes * 4,
bitW,
kernel_size=1,
bias=False)
self.bn3 = nn.BatchNorm2d(planes * 4)
self.downsample = downsample
self.QReLU = QReLU(k=bitA)
self.stride = stride
self.is_last = is_last
def _make_layer(self, block, planes, blocks, stride=1, is_last=False):
downsample_booster = None
downsample_residual = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample_booster = nn.Sequential(
nn.Conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
downsample_residual = nn.Sequential(
QConv2d(self.inplanes, planes * block.expansion, self.bitW,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = nn.ModuleList([])
layers.append(block(self.inplanes, planes, self.bitW, self.bitA, stride, downsample_booster, downsample_residual))
self.inplanes = planes * block.expansion
for i in range(1, blocks-1):
layers.append(block(self.inplanes, planes, self.bitW, self.bitA))
layers.append(block(self.inplanes, planes, self.bitW, self.bitA, is_last=is_last))
return layers
def _make_layer(self, block, planes, blocks, stride=1, is_last=False):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
QConv2d(self.inplanes,
planes * block.expansion,
self.bitW,
kernel_size=1,
stride=stride,
bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(
block(self.inplanes, planes, self.bitW, self.bitA, stride,
downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks - 1):
layers.append(block(self.inplanes, planes, self.bitW, self.bitA))
layers.append(
block(self.inplanes, planes, self.bitW, self.bitA,
is_last=is_last))
return nn.Sequential(*layers)
def __init__(self, inplanes, planes, bitW, bitA, stride=1, downsample_booster=None, downsample_residual=None, is_last=False):
super(Bottleneck, self).__init__()
self.bitW = bitW
self.bitA = bitA
self.conv1 = QConv2d(inplanes, planes, bitW, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = QConv2d(planes, planes, bitW, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = QConv2d(planes, planes * 4, bitW, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * 4)
self.QReLU = QReLU(k=bitA)
self.downsample_booster = downsample_booster
self.downsample = downsample_residual
self.stride = stride
self.auxiliary = nn.Sequential(conv1x1(planes * self.expansion, planes * self.expansion), nn.BatchNorm2d(planes * self.expansion))
self.is_last = is_last
def conv1x1(in_planes, out_planes, bitW=32, stride=1):
"3x3 convolution with padding"
return QConv2d(in_planes,
out_planes,
bitW=32,
kernel_size=1,
stride=stride,
bias=False)
def conv3x3(in_planes, out_planes, bitW, stride=1):
"3x3 convolution with padding"
return QConv2d(in_planes,
out_planes,
bitW,
kernel_size=3,
stride=stride,
padding=1,
bias=False)