def _make_layer(self,
block,
planes,
blocks,
mask_init,
mask_scale,
threshold_fn,
stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nl.ElementWiseConv2d(self.inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False,
mask_init=mask_init,
mask_scale=mask_scale,
threshold_fn=threshold_fn),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(
block(self.inplanes, planes, mask_init, mask_scale, threshold_fn,
stride, downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(self.inplanes, planes, mask_init, mask_scale,
threshold_fn))
return nn.Sequential(*layers)
def __init__(self,
block,
layers,
mask_init,
mask_scale,
threshold_fn,
num_classes=1000):
self.inplanes = 64
super(ResNet, self).__init__()
self.conv1 = nl.ElementWiseConv2d(3,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False,
mask_init=mask_init,
mask_scale=mask_scale,
threshold_fn=threshold_fn)
self.bn1 = nn.BatchNorm2d(64)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0], mask_init,
mask_scale, threshold_fn)
self.layer2 = self._make_layer(block,
128,
layers[1],
mask_init,
mask_scale,
threshold_fn,
stride=2)
self.layer3 = self._make_layer(block,
256,
layers[2],
mask_init,
mask_scale,
threshold_fn,
stride=2)
self.layer4 = self._make_layer(block,
512,
layers[3],
mask_init,
mask_scale,
threshold_fn,
stride=2)
self.avgpool = nn.AvgPool2d(7, stride=1)
self.fc = nn.Linear(512 * block.expansion, num_classes)
for m in self.modules():
if isinstance(m, nl.ElementWiseConv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def __init__(self,
inplanes,
planes,
mask_init,
mask_scale,
threshold_fn,
stride=1,
downsample=None):
super(Bottleneck, self).__init__()
self.conv1 = nl.ElementWiseConv2d(inplanes,
planes,
kernel_size=1,
bias=False,
mask_init=mask_init,
mask_scale=mask_scale,
threshold_fn=threshold_fn)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nl.ElementWiseConv2d(planes,
planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False,
mask_init=mask_init,
mask_scale=mask_scale,
threshold_fn=threshold_fn)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = nl.ElementWiseConv2d(planes,
planes * 4,
kernel_size=1,
bias=False,
mask_init=mask_init,
mask_scale=mask_scale,
threshold_fn=threshold_fn)
self.bn3 = nn.BatchNorm2d(planes * 4)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def conv3x3(in_planes,
out_planes,
mask_init,
mask_scale,
threshold_fn,
stride=1):
"3x3 convolution with padding"
return nl.ElementWiseConv2d(in_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False,
mask_init=mask_init,
mask_scale=mask_scale,
threshold_fn=threshold_fn)
def make_layers(cfg, mask_init, mask_scale, threshold_fn, batch_norm=False):
layers = []
in_channels = 3
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = nl.ElementWiseConv2d(
in_channels, v, kernel_size=3, padding=1,
mask_init=mask_init, mask_scale=mask_scale,
threshold_fn=threshold_fn)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)]
else:
layers += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return nn.Sequential(*layers)