def __init__(self, in_planes, planes, bits, stride=1):
super(QuantBottleneck, self).__init__()
first_bit = bits.pop(0)
self.conv1 = DSConv2d(in_planes,
planes,
1,
32,
bit=first_bit,
bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.activation1 = BFPActivation(bits[0], 32)
self.conv2 = DSConv2d(
planes,
planes,
kernel_size=3,
block_size=32,
bit=bits.pop(0),
stride=stride,
padding=1,
bias=False,
)
self.bn2 = nn.BatchNorm2d(planes)
self.activation2 = BFPActivation(bits[0], 32)
self.conv3 = DSConv2d(
planes,
self.expansion * planes,
kernel_size=1,
block_size=32,
bit=bits.pop(0),
bias=False,
)
self.bn3 = nn.BatchNorm2d(self.expansion * planes)
self.activation3 = (BFPActivation(bits[0], 32)
if len(bits) > 1 else nn.Identity())
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
DSConv2d(
in_planes,
self.expansion * planes,
1,
block_size=32,
bit=first_bit,
stride=stride,
bias=False,
),
nn.BatchNorm2d(self.expansion * planes),
)
def __init__(self, in_planes, planes, bits, stride=1):
super(QuantBasicBlock, self).__init__()
first_bit = bits.pop(0)
self.conv1 = DSConv2d(
in_planes,
planes,
3,
block_size=32,
bit=first_bit,
stride=stride,
padding=1,
bias=False,
)
self.bn1 = nn.BatchNorm2d(planes)
self.activation1 = BFPActivation(bits[0], 32)
self.conv2 = DSConv2d(
planes,
planes,
kernel_size=3,
block_size=32,
bit=bits.pop(0),
stride=1,
padding=1,
bias=False,
)
self.bn2 = nn.BatchNorm2d(planes)
# This is just because the last layer doesn't need to quantize the activation
self.activation2 = (BFPActivation(bits[0], 32)
if len(bits) > 0 else nn.Identity())
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
DSConv2d(
in_planes,
self.expansion * planes,
1,
block_size=32,
bit=first_bit,
stride=stride,
bias=False,
),
nn.BatchNorm2d(self.expansion * planes),
)
def conv1x1(in_planes, out_planes, block_size, bit, stride=1):
"""1x1 convolution"""
return DSConv2d(
in_planes,
out_planes,
kernel_size=1,
block_size=block_size,
bit=bit,
stride=stride,
bias=False,
)
def __init__(self, bits):
super(CNNX, self).__init__(bits, self.number_bits)
self.block_size = 32
bit = self.bits.pop(0)
self.conv1 = DSConv2d(3, 64, (3, 3), block_size=32, bit=bit, padding=1)
self.bn1 = torch.nn.BatchNorm2d(64)
self.features1, outch = self.__make_layers__(64, 3)
self.max_pool1 = torch.nn.MaxPool2d(2, stride=2)
self.features2, outch = self.__make_layers__(outch, 3)
self.max_pool2 = torch.nn.MaxPool2d(2, stride=2)
self.features3, outch = self.__make_layers__(outch, 3)
self.avg_pool = torch.nn.AvgPool2d(8)
self.linear = torch.nn.Linear(outch, 10)
def conv3x3(in_planes,
out_planes,
block_size,
bit,
stride=1,
groups=1,
dilation=1):
"""3x3 convolution with padding"""
return DSConv2d(
in_planes,
out_planes,
kernel_size=3,
block_size=block_size,
bit=bit,
stride=stride,
padding=dilation,
groups=groups,
bias=False,
dilation=dilation,
)
def __init__(self, block, num_blocks, bits, num_classes=10):
super(QUANTIZED_ResNet, self).__init__()
self.in_planes = 64
_bits_ = bits.copy()
self.conv1 = DSConv2d(
3,
64,
kernel_size=3,
block_size=32,
bit=_bits_.pop(0),
stride=1,
padding=1,
bias=False,
)
self.activation = BFPActivation(_bits_[0], 32)
self.bn1 = nn.BatchNorm2d(64)
self.layer1 = self._make_layer(block,
64,
num_blocks[0],
stride=1,
bits=_bits_)
self.layer2 = self._make_layer(block,
128,
num_blocks[1],
stride=2,
bits=_bits_)
self.layer3 = self._make_layer(block,
256,
num_blocks[2],
stride=2,
bits=_bits_)
self.layer4 = self._make_layer(block,
512,
num_blocks[3],
stride=2,
bits=_bits_)
self.linear = nn.Linear(512 * block.expansion, num_classes)
def __init__(self, in_planes, out_planes, kernel, block_size, bit, **kwargs):
super(BaseConv, self).__init__()
self.activation = BFPActivation(bit, 7, block_size)
self.conv = DSConv2d(in_planes, out_planes, kernel, block_size, bit, **kwargs)
self.bn = torch.nn.BatchNorm2d(out_planes)
def __init__(
self,
block,
layers,
bits,
number_bits,
num_classes=1000,
zero_init_residual=False,
groups=1,
width_per_group=64,
replace_stride_with_dilation=None,
norm_layer=None,
):
super(ResNet, self).__init__(bits, number_bits)
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
bit = self.bits.pop(0)
self.conv1 = DSConv2d(
3,
self.inplanes,
kernel_size=7,
block_size=32,
bit=bit,
stride=2,
padding=3,
bias=False,
)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
bit = self.bits[0]
self.activation1 = BFPActivation(bit, 32)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block,
128,
layers[1],
stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block,
256,
layers[2],
stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(
block,
512,
layers[3],
stride=2,
dilate=replace_stride_with_dilation[2],
final=True,
)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(512 * block.expansion, num_classes)
for mod in self.modules():
if isinstance(mod, DSConv2d):
nn.init.kaiming_normal_(mod.weight,
mode="fan_out",
nonlinearity="relu")
elif isinstance(mod, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(mod.weight, 1)
nn.init.constant_(mod.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 mod in self.modules():
if isinstance(mod, Bottleneck):
nn.init.constant_(mod.bn3.weight, 0)
elif isinstance(mod, BasicBlock):
nn.init.constant_(mod.bn2.weight, 0)