def conv(input_shape,
num_classes,
dense_classifier=False,
pretrained=False,
L=3,
N=32,
nonlinearity=nn.ReLU()):
channels, width, height = input_shape
# Convolutional feature extractor
modules = []
modules.append(layers.Conv2d(channels, N, kernel_size=3, padding=3 // 2))
modules.append(nonlinearity)
for i in range(L - 2):
modules.append(layers.Conv2d(N, N, kernel_size=3, padding=3 // 2))
modules.append(nonlinearity)
# Linear classifier
modules.append(nn.Flatten())
if dense_classifier:
modules.append(nn.Linear(N * width * height, num_classes))
else:
modules.append(layers.Linear(N * width * height, num_classes))
model = nn.Sequential(*modules)
# Pretrained model
if pretrained:
print("WARNING: this model does not have pretrained weights.")
return model
def __init__(self, in_channels, out_channels, stride=1, base_width=64):
super().__init__()
width = int(out_channels * (base_width / 64.))
self.residual_function = nn.Sequential(
layers.Conv2d(in_channels, width, kernel_size=1, bias=False),
layers.BatchNorm2d(width),
nn.ReLU(inplace=True),
layers.Conv2d(width,
width,
stride=stride,
kernel_size=3,
padding=1,
bias=False),
layers.BatchNorm2d(width),
nn.ReLU(inplace=True),
layers.Conv2d(width,
out_channels * BottleNeck.expansion,
kernel_size=1,
bias=False),
layers.BatchNorm2d(out_channels * BottleNeck.expansion),
)
self.shortcut = layers.Identity2d(in_channels)
if stride != 1 or in_channels != out_channels * BottleNeck.expansion:
self.shortcut = nn.Sequential(
layers.Conv2d(in_channels,
out_channels * BottleNeck.expansion,
stride=stride,
kernel_size=1,
bias=False),
layers.BatchNorm2d(out_channels * BottleNeck.expansion))
def __init__(self, f_in: int, f_out: int, downsample=False):
super(Block, self).__init__()
stride = 2 if downsample else 1
self.conv1 = layers.Conv2d(f_in,
f_out,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
self.bn1 = layers.BatchNorm2d(f_out)
self.conv2 = layers.Conv2d(f_out,
f_out,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn2 = layers.BatchNorm2d(f_out)
# No parameters for shortcut connections.
if downsample or f_in != f_out:
self.shortcut = nn.Sequential(
layers.Conv2d(f_in, f_out, kernel_size=1, stride=2,
bias=False), layers.BatchNorm2d(f_out))
else:
self.shortcut = layers.Identity2d(f_in)
def __init__(self, in_channels, out_channels, stride=1, base_width=64):
super().__init__()
#residual function
self.residual_function = nn.Sequential(
layers.Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=stride,
padding=1,
bias=False), layers.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
layers.Conv2d(out_channels,
out_channels * BasicBlock.expansion,
kernel_size=3,
padding=1,
bias=False),
layers.BatchNorm2d(out_channels * BasicBlock.expansion))
#shortcut
self.shortcut = layers.Identity2d(in_channels)
#the shortcut output dimension is not the same with residual function
#use 1*1 convolution to match the dimension
if stride != 1 or in_channels != BasicBlock.expansion * out_channels:
self.shortcut = nn.Sequential(
layers.Conv2d(in_channels,
out_channels * BasicBlock.expansion,
kernel_size=1,
stride=stride,
bias=False),
layers.BatchNorm2d(out_channels * BasicBlock.expansion))
def __init__(self,
in_channels,
out_channels,
stride=1,
base_width=64,
batch_norm=True):
super().__init__()
self.batch_norm = batch_norm
# residual function
layer_list = [
layers.Conv2d(
in_channels,
out_channels,
kernel_size=3,
stride=stride,
padding=1,
bias=False,
),
]
if self.batch_norm:
layer_list.append(layers.BatchNorm2d(out_channels))
layer_list += [
nn.ReLU(inplace=True),
layers.Conv2d(
out_channels,
out_channels * BasicBlock.expansion,
kernel_size=3,
padding=1,
bias=False,
),
]
if self.batch_norm:
layer_list.append(
layers.BatchNorm2d(out_channels * BasicBlock.expansion))
self.residual_function = nn.Sequential(*layer_list)
# shortcut
self.shortcut = layers.Identity2d(in_channels)
# the shortcut output dimension is not the same with residual function
# use 1*1 convolution to match the dimension
if stride != 1 or in_channels != BasicBlock.expansion * out_channels:
layer_list = [
layers.Conv2d(
in_channels,
out_channels * BasicBlock.expansion,
kernel_size=1,
stride=stride,
bias=False,
)
]
if self.batch_norm:
layer_list.append(
layers.BatchNorm2d(out_channels * BasicBlock.expansion))
self.shortcut = nn.Sequential(*layer_list)
def __init__(self,
in_channels,
out_channels,
stride=1,
base_width=64,
batch_norm=True):
super().__init__()
self.batch_norm = batch_norm
width = int(out_channels * (base_width / 64.0))
layer_list = [
layers.Conv2d(in_channels, width, kernel_size=1, bias=False),
]
if self.batch_norm:
layer_list.append(layers.BatchNorm2d(width))
layer_list += [
nn.ReLU(inplace=True),
layers.Conv2d(width,
width,
stride=stride,
kernel_size=3,
padding=1,
bias=False),
]
if self.batch_norm:
layer_list.append(layers.BatchNorm2d(width))
layer_list += [
nn.ReLU(inplace=True),
layers.Conv2d(width,
out_channels * BottleNeck.expansion,
kernel_size=1,
bias=False),
]
if self.batch_norm:
layer_list.append(
layers.BatchNorm2d(out_channels * BottleNeck.expansion))
self.residual_function = nn.Sequential(*layer_list)
self.shortcut = layers.Identity2d(in_channels)
if stride != 1 or in_channels != out_channels * BottleNeck.expansion:
layer_list = [
layers.Conv2d(
in_channels,
out_channels * BottleNeck.expansion,
stride=stride,
kernel_size=1,
bias=False,
),
]
if self.batch_norm:
layer_list.append(
layers.BatchNorm2d(out_channels * BottleNeck.expansion))
self.shortcut = nn.Sequential(*layer_list)
def conv1x1(in_planes, out_planes, stride=1):
"""1x1 convolution"""
return layers.Conv2d(in_planes,
out_planes,
kernel_size=1,
stride=stride,
bias=False)
def __init__(self, plan, num_classes, dense_classifier):
super(ResNet, self).__init__()
# Initial convolution.
current_filters = plan[0][0]
self.conv = layers.Conv2d(
3, current_filters, kernel_size=3, stride=1, padding=1, bias=False
)
self.bn = layers.BatchNorm2d(current_filters)
# The subsequent blocks of the ResNet.
blocks = []
for segment_index, (filters, num_blocks) in enumerate(plan):
for block_index in range(num_blocks):
downsample = segment_index > 0 and block_index == 0
blocks.append(Block(current_filters, filters, downsample))
current_filters = filters
self.blocks = nn.Sequential(*blocks)
self.fc = layers.Linear(plan[-1][0], num_classes)
if dense_classifier:
self.fc = nn.Linear(plan[-1][0], num_classes)
self._initialize_weights()
def __init__(self, in_filters, out_filters):
super(ConvBNModule, self).__init__()
self.conv = layers.Conv2d(in_filters,
out_filters,
kernel_size=3,
padding=1)
self.bn = layers.BatchNorm2d(out_filters)
def __init__(self,
block,
num_block,
base_width,
num_classes=200,
dense_classifier=False):
super().__init__()
self.in_channels = 64
self.conv1 = nn.Sequential(
layers.Conv2d(3, 64, kernel_size=3, padding=1, bias=False),
layers.BatchNorm2d(64), nn.ReLU(inplace=True))
#we use a different inputsize than the original paper
#so conv2_x's stride is 1
self.conv2_x = self._make_layer(block, 64, num_block[0], 1, base_width)
self.conv3_x = self._make_layer(block, 128, num_block[1], 2,
base_width)
self.conv4_x = self._make_layer(block, 256, num_block[2], 2,
base_width)
self.conv5_x = self._make_layer(block, 512, num_block[3], 2,
base_width)
self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = layers.Linear(512 * block.expansion, num_classes)
if dense_classifier:
self.fc = nn.Linear(512 * block.expansion, num_classes)
self._initialize_weights()
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
"""3x3 convolution with padding"""
return layers.Conv2d(in_planes,
out_planes,
kernel_size=3,
stride=stride,
padding=dilation,
groups=groups,
bias=False,
dilation=dilation)
def make_layers(cfg, batch_norm=False):
layer_list = []
in_channels = 3
for v in cfg:
if v == 'M':
layer_list += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = layers.Conv2d(in_channels, v, kernel_size=3, padding=1)
if batch_norm:
layer_list += [
conv2d,
layers.BatchNorm2d(v),
nn.ReLU(inplace=True)
]
else:
layer_list += [conv2d, nn.ReLU(inplace=True)]
in_channels = v
return nn.Sequential(*layer_list)
def make_layers(cfg, batch_norm=False):
layer_list = []
input_channel = 3
for l in cfg:
if l == 'M':
layer_list += [nn.MaxPool2d(kernel_size=2, stride=2)]
continue
layer_list += [
layers.Conv2d(input_channel, l, kernel_size=3, padding=1)
]
if batch_norm:
layer_list += [layers.BatchNorm2d(l)]
layer_list += [nn.ReLU(inplace=True)]
input_channel = l
return nn.Sequential(*layer_list)
def __init__(self,
block,
layer_list,
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__()
if norm_layer is None:
norm_layer = layers.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
self.conv1 = layers.Conv2d(3,
self.inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layer_list[0])
self.layer2 = self._make_layer(block,
128,
layer_list[1],
stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block,
256,
layer_list[2],
stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(block,
512,
layer_list[3],
stride=2,
dilate=replace_stride_with_dilation[2])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = layers.Linear(512 * block.expansion, num_classes)
for m in self.modules():
if isinstance(m, layers.Conv2d):
nn.init.kaiming_normal_(m.weight,
mode='fan_out',
nonlinearity='relu')
elif isinstance(m, (layers.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.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 m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
