def __init__(
self,
inplanes,
planes,
stride=1,
downsample=None,
act_bitwidth=32,
act_quant=False,
act_noise=False,
):
super(BasicBlock, self).__init__()
if isinstance(act_bitwidth, list):
assert (len(act_bitwidth) == 2)
self.act_bitwidth = act_bitwidth
else:
self.act_bitwidth = [act_bitwidth] * 2
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = nn.BatchNorm2d(planes)
self.relu1 = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=act_bitwidth[0])
self.conv2 = conv3x3(planes, planes)
self.bn2 = nn.BatchNorm2d(planes)
self.downsample = downsample
self.relu2 = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=act_bitwidth[1])
self.stride = stride
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride=1,
padding=0,
act_bitwidth=32,
act_quant=False,
act_noise=False,
uniq=True):
super(DepthwiseSeparableFusedConv2d, self).__init__()
if act_bitwidth == 32:
self.components = nn.Sequential(
Conv2d(in_channels,
in_channels,
kernel_size,
stride=stride,
padding=padding,
groups=in_channels),
nn.BatchNorm2d(in_channels),
# nn.ReLU(inplace=True),
actquant.ActQuantBuffers(quant=act_quant,
noise=act_noise,
bitwidth=act_bitwidth,
uniq=uniq),
Conv2d(in_channels, out_channels, 1, bias=False),
nn.BatchNorm2d(out_channels),
# nn.ReLU(inplace=True)
actquant.ActQuantBuffers(quant=act_quant,
noise=act_noise,
bitwidth=act_bitwidth,
uniq=uniq))
else:
self.components = nn.Sequential(
Conv2d(in_channels,
in_channels,
kernel_size,
stride=stride,
padding=padding,
groups=in_channels), nn.BatchNorm2d(in_channels),
actquant.ActQuantBuffers(quant=act_quant,
noise=act_noise,
bitwidth=act_bitwidth,
uniq=uniq),
Conv2d(in_channels, out_channels, 1, bias=False),
nn.BatchNorm2d(out_channels),
actquant.ActQuantBuffers(quant=act_quant,
noise=act_noise,
bitwidth=act_bitwidth,
uniq=uniq))
def get_feature_activation(self):
if self.wrpn:
return actquant.ActQuantWRPN(act_quant=self.act_quant,
act_bitwidth=self.act_bitwidth)
else:
return actquant.ActQuantBuffers(quant=self.act_quant,
bitwidth=self.act_bitwidth)
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
act_bitwidth=32,
act_quant=False):
super(Bottleneck, self).__init__()
if isinstance(act_bitwidth, list):
assert (len(act_bitwidth) == 3)
self.act_bitwidth = act_bitwidth
else:
self.act_bitwidth = [act_bitwidth] * 3
self.conv1 = Conv2d(inplanes, planes, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.relu1 = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=act_bitwidth[0])
self.conv2 = Conv2d(planes,
planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False,
groups=32)
self.bn2 = nn.BatchNorm2d(planes)
self.relu2 = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=act_bitwidth[1])
self.conv3 = Conv2d(planes, planes * 4, kernel_size=1, bias=False)
self.bn3 = nn.BatchNorm2d(planes * 4)
self.relu3 = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=act_bitwidth[2])
self.downsample = downsample
self.stride = stride
def __init__(self,
quant_epoch_step,
quant_start_stage,
alpha=1.0,
shallow=False,
num_classes=1000,
quant=False,
noise=False,
bitwidth=32,
step=2,
quant_edges=True,
act_bitwidth=32,
act_noise=True,
step_setup=[15, 9],
act_quant=False,
uniq=True,
normalize=False):
super(MobileNet, self).__init__(quant_epoch_step=quant_epoch_step,
quant_start_stage=quant_start_stage,
quant=quant,
noise=noise,
bitwidth=bitwidth,
step=step,
quant_edges=quant_edges,
act_noise=act_noise,
step_setup=step_setup,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
uniq=uniq)
# if act_bitwidth == 32:
# act = nn.ReLU(inplace=True)
# else:
self.preprocess = lambda x: x
act = actquant.ActQuantBuffers(quant=act_quant,
noise=act_noise,
bitwidth=act_bitwidth,
uniq=self.uniq)
layers = [
Conv2d(3,
nearby_int(alpha * 32),
kernel_size=3,
stride=2,
padding=1,
bias=False),
nn.BatchNorm2d(nearby_int(alpha * 32)), act,
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 32),
nearby_int(alpha * 64),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 64),
nearby_int(alpha * 128),
kernel_size=3,
stride=2,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 128),
nearby_int(alpha * 128),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 128),
nearby_int(alpha * 256),
kernel_size=3,
stride=2,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 256),
nearby_int(alpha * 256),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 256),
nearby_int(alpha * 512),
kernel_size=3,
stride=2,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq)
]
if not shallow:
# 5x 512->512 DW-separable convolutions
layers += [
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 512),
nearby_int(alpha * 512),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 512),
nearby_int(alpha * 512),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 512),
nearby_int(alpha * 512),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 512),
nearby_int(alpha * 512),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 512),
nearby_int(alpha * 512),
kernel_size=3,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
]
layers += [
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 512),
nearby_int(alpha * 1024),
kernel_size=3,
stride=2,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq),
# Paper specifies stride-2, but unchanged size.
# Assume its a typo and use stride-1 convolution
DepthwiseSeparableFusedConv2d(nearby_int(alpha * 1024),
nearby_int(alpha * 1024),
kernel_size=3,
stride=1,
padding=1,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
act_noise=act_noise,
uniq=uniq)
]
self.features = nn.Sequential(*layers)
self.avg_pool = nn.AvgPool2d(7)
self.classifier = Linear(nearby_int(alpha * 1024), num_classes)
self.regime = [
{
'epoch': 0,
'optimizer': 'SGD',
'lr': 1e-3,
# 'weight_decay': 5e-4, 'momentum': 0.9},
'weight_decay': 1e-4,
'momentum': 0.9
},
{
'epoch': 30,
'lr': 5e-3
},
{
'epoch': 60,
'lr': 5e-4,
'weight_decay': 0
},
{
'epoch': 90,
'lr': 5e-5
}
]
def __init__(self,
quant_epoch_step,
quant_start_stage,
num_classes=10,
block=BasicBlock,
depth=18,
quant=False,
noise=False,
bitwidth=32,
step=2,
quant_edges=True,
act_noise=True,
step_setup=[15, 9],
act_bitwidth=32,
act_quant=False,
layers=[2, 2, 2, 2],
normalize=False,
noise_mask=0.05):
super(ResNet_cifar10,
self).__init__(quant_epoch_step=quant_epoch_step,
quant_start_stage=quant_start_stage,
quant=quant,
noise=noise,
bitwidth=bitwidth,
step=step,
quant_edges=quant_edges,
step_setup=step_setup,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
noise_mask=noise_mask)
self.preprocess = lambda x: x
self.inplanes = 64
n = int((depth - 2) / 6)
self.conv1 = Conv2d(3,
64,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn1 = nn.BatchNorm2d(64)
if isinstance(act_bitwidth, list):
if self.quant_edges:
assert (len(act_bitwidth) == depth - 1)
else:
assert (len(act_bitwidth) == depth - 2)
self.act_bitwidth = act_bitwidth
else:
if self.quant_edges:
self.act_bitwidth = [act_bitwidth] * (depth - 1)
else:
self.act_bitwidth = [act_bitwidth] * (depth - 2)
self.relu = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=act_bitwidth)
self.maxpool = lambda x: x
self.layer1 = self._make_layer(block,
64,
layers[0],
act_bitwidth=act_bitwidth,
act_quant=act_quant)
self.layer2 = self._make_layer(block,
128,
layers[1],
stride=2,
act_bitwidth=act_bitwidth,
act_quant=act_quant)
self.layer3 = self._make_layer(block,
256,
layers[2],
stride=2,
act_bitwidth=act_bitwidth,
act_quant=act_quant)
self.layer4 = self._make_layer(block,
512,
layers[3],
stride=2,
act_bitwidth=act_bitwidth,
act_quant=act_quant)
# self.layer4 = lambda x: x
self.avgpool = nn.AvgPool2d(4)
self.fc = Linear(512, num_classes)
init_model(self)
self.regime = [{
'epoch': 0,
'optimizer': 'SGD',
'lr': 1e-1,
'weight_decay': 1e-4,
'momentum': 0.9
}, {
'epoch': 81,
'lr': 1e-2
}, {
'epoch': 122,
'lr': 1e-3,
'weight_decay': 0
}, {
'epoch': 164,
'lr': 1e-4
}]
def __init__(self,
quant_epoch_step,
quant_start_stage,
num_classes=1000,
block=Bottleneck,
layers=[3, 4, 23, 3],
quant=False,
noise=False,
bitwidth=32,
step=2,
quant_edges=True,
step_setup=[15, 9],
act_bitwidth=32,
act_quant=False,
uniq=True,
normalize=False,
noise_mask=0.05):
super(ResNet_imagenet,
self).__init__(quant_epoch_step=quant_epoch_step,
quant_start_stage=quant_start_stage,
quant=quant,
noise=noise,
bitwidth=bitwidth,
step=step,
quant_edges=quant_edges,
step_setup=step_setup,
act_bitwidth=act_bitwidth,
act_quant=act_quant,
noise_mask=noise_mask)
self.inplanes = 64
self.preprocess = lambda x: x
self.conv1 = Conv2d(3,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = nn.BatchNorm2d(64)
block_size = 2 if block is BasicBlock else 3
self.relu = actquant.ActQuantBuffers(quant=act_quant,
bitwidth=self.act_bitwidth)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
act_nums = [1]
for i in range(4):
act_nums.append(act_nums[i] + layers[i] * block_size)
self.layer1 = self._make_layer(block,
64,
layers[0],
act_bitwidth=self.act_bitwidth,
act_quant=act_quant)
self.layer2 = self._make_layer(block,
128,
layers[1],
stride=2,
act_bitwidth=self.act_bitwidth,
act_quant=act_quant)
self.layer3 = self._make_layer(block,
256,
layers[2],
stride=2,
act_bitwidth=self.act_bitwidth,
act_quant=act_quant)
self.layer4 = self._make_layer(block,
512,
layers[3],
stride=2,
act_bitwidth=self.act_bitwidth,
act_quant=act_quant)
self.avgpool = nn.AvgPool2d(7)
self.fc = Linear(512 * block.expansion, num_classes)
init_model(self)
self.regime = [{
'epoch': 0,
'optimizer': 'SGD',
'lr': 5e-2,
'weight_decay': 1e-4,
'momentum': 0.9
}, {
'epoch': 30,
'lr': 5e-3
}, {
'epoch': 60,
'lr': 5e-4,
'weight_decay': 0
}, {
'epoch': 90,
'lr': 5e-5
}]