def __init__(self,
inplanes,
outplanes,
innerplanes,
stride=1,
dilation=1,
group=1,
downsample=None,
cfg_prune=None):
super().__init__()
# In original resnet, stride=2 is on 1x1.
# In fb.torch resnet, stride=2 is on 3x3.
(str1x1, str3x3) = (stride, 1) if cfg.RESNETS.STRIDE_1X1 else (1,
stride)
self.stride = stride
if cfg_prune is None:
self.conv1 = nn.Conv2d(inplanes,
innerplanes,
kernel_size=3,
stride=str1x1,
padding=1,
bias=False)
if cfg.REID.BN_COMPLETE:
self.bn1 = nn.BatchNorm2d(innerplanes)
else:
self.bn1 = mynn.AffineChannel2d(innerplanes)
self.conv2 = nn.Conv2d(innerplanes,
innerplanes,
kernel_size=3,
stride=str3x3,
bias=False,
padding=1,
dilation=dilation,
groups=group)
if cfg.REID.BN_COMPLETE:
self.bn2 = nn.BatchNorm2d(innerplanes)
else:
self.bn2 = mynn.AffineChannel2d(innerplanes)
else:
self.conv1 = nn.Conv2d(inplanes,
cfg_prune,
kernel_size=3,
stride=str1x1,
padding=1,
bias=False)
self.bn1 = mynn.AffineChannel2d(cfg_prune)
self.conv2 = nn.Conv2d(cfg_prune,
innerplanes,
kernel_size=3,
stride=str3x3,
bias=False,
padding=1,
dilation=dilation,
groups=group)
self.bn2 = mynn.AffineChannel2d(innerplanes)
self.downsample = downsample
self.relu = nn.ReLU(inplace=True)
def __init__(self, inplanes, outplanes, innerplanes, stride=1, dilation=1, group=1,
downsample=None, attention=False):
super().__init__()
# In original resnet, stride=2 is on 1x1.
# In fb.torch resnet, stride=2 is on 3x3.
(str1x1, str3x3) = (stride, 1) if cfg.RESNETS.STRIDE_1X1 else (1, stride)
self.stride = stride
self.attention = attention
self.conv1 = nn.Conv2d(
inplanes, innerplanes, kernel_size=1, stride=str1x1, bias=False)
self.bn1 = mynn.AffineChannel2d(innerplanes)
self.conv2 = nn.Conv2d(
innerplanes, innerplanes, kernel_size=3, stride=str3x3, bias=False,
padding=1 * dilation, dilation=dilation, groups=group)
self.bn2 = mynn.AffineChannel2d(innerplanes)
self.conv3 = nn.Conv2d(
innerplanes, outplanes, kernel_size=1, stride=1, bias=False)
self.bn3 = mynn.AffineChannel2d(outplanes)
self.downsample = downsample
self.relu = nn.ReLU(inplace=True)
if self.attention:
self.cbam = CBAM(innerplanes*4, 16)
def conv_bn(inp, oup, stride):
return nn.Sequential(
nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
mynn.AffineChannel2d(oup),
# nn.BatchNorm2d(oup),
nn.ReLU6(inplace=True),
)
def add_stage(inplanes, planes, nblocks, stride=1, dilation=1):
"""Make a stage consist of `nblocks` residual blocks.
Returns:
- stage module: an nn.Sequentail module of residual blocks
- final output dimension
"""
block_func = globals()[cfg.RESNETS.TRANS_FUNC]
downsample = None
if stride != 1 or inplanes != planes * block_func.expansion:
downsample = nn.Sequential(
nn.Conv2d(inplanes,
planes * block_func.expansion,
kernel_size=1,
stride=stride,
bias=False),
mynn.AffineChannel2d(planes * block_func.expansion),
)
layers = []
layers.append(
block_func(inplanes,
planes,
stride,
dilation=dilation,
group=cfg.RESNETS.NUM_GROUPS,
downsample=downsample))
inplanes = planes * block_func.expansion
for i in range(1, nblocks):
layers.append(block_func(inplanes, planes, dilation=dilation))
return nn.Sequential(*layers), inplanes
def __init__(self, block_counts):
super().__init__()
self.block_counts = block_counts
self.convX = len(block_counts) + 1
self.num_layers = (sum(block_counts) + 3 * (self.convX == 4)) * 3 + 2
self.res1 = nn.Sequential(
OrderedDict([('conv1', nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=False)),
('bn1', mynn.AffineChannel2d(64)),
('relu', nn.ReLU(inplace=True)),
# ('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True))]))
('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))]))
dim_in = 64
dim_bottleneck = cfg.RESNETS.NUM_GROUPS * cfg.RESNETS.WIDTH_PER_GROUP
self.res2, dim_in = add_stage(dim_in, dim_bottleneck, block_counts[0])
self.res3, dim_in = add_stage(dim_in, dim_bottleneck * 2, block_counts[1], stride=2)
self.res4, dim_in = add_stage(dim_in, dim_bottleneck * 4, block_counts[2], stride=2)
if len(block_counts) == 4:
if cfg.RESNETS.RES5_DILATION != 1:
stride = 1
else:
stride = 2
self.res5, dim_in = add_stage(dim_in, dim_bottleneck * 8, block_counts[3],
stride, cfg.RESNETS.RES5_DILATION)
self.spatial_scale = 1 / 32 * cfg.RESNETS.RES5_DILATION
else:
self.spatial_scale = 1 / 16 # final feature scale wrt. original image scale
self.dim_out = dim_in
self._init_modules()
def basic_bn_stem():
return nn.Sequential(OrderedDict([
('conv1', nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=False)),
('bn1', mynn.AffineChannel2d(64)),
('relu', nn.ReLU(inplace=True)),
# ('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True))]))
('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))]))
def basic_bn_shortcut(inplanes, outplanes, stride):
if cfg.RESNETS.SSN:
return nn.Sequential(
nn.Conv2d(inplanes,
outplanes,
kernel_size=1,
stride=stride,
bias=False),
SSN2d(outplanes, using_moving_average=True),
)
else:
if cfg.REID.BN_COMPLETE:
return nn.Sequential(
nn.Conv2d(inplanes,
outplanes,
kernel_size=1,
stride=stride,
bias=False),
nn.BatchNorm2d(outplanes),
)
else:
return nn.Sequential(
nn.Conv2d(inplanes,
outplanes,
kernel_size=1,
stride=stride,
bias=False),
mynn.AffineChannel2d(outplanes),
)
def basic_bn_shortcut(inplanes, outplanes, stride):
return nn.Sequential(
nn.Conv2d(inplanes,
outplanes,
kernel_size=1,
stride=stride,
bias=False),
mynn.AffineChannel2d(outplanes),
)
def basic_bn_stem():
#if cfg.LESION.LESION_ENABLED:
#input_dim = cfg.LESION.SLICE_NUM
#else:
input_dim = 3
return nn.Sequential(OrderedDict([
('conv1', nn.Conv2d(input_dim, 64, 7, stride=2, padding=3, bias=False)),
('bn1', mynn.AffineChannel2d(64)),
('relu', nn.ReLU(inplace=True)),
('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))]))
def __init__(self, inp, oup, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
assert stride in [1, 2]
self.use_res_connect = self.stride == 1 and inp == oup
self.conv = nn.Sequential(
# pw
nn.Conv2d(inp, inp * expand_ratio, 1, 1, 0, bias=False),
mynn.AffineChannel2d(inp * expand_ratio),
# nn.BatchNorm2d(inp * expand_ratio),
nn.ReLU6(inplace=True),
# dw
nn.Conv2d(inp * expand_ratio, inp * expand_ratio, 3, stride, 1, groups=inp * expand_ratio, bias=False),
mynn.AffineChannel2d(inp * expand_ratio),
# nn.BatchNorm2d(inp * expand_ratio),
nn.ReLU6(inplace=True),
# pw-linear
nn.Conv2d(inp * expand_ratio, oup, 1, 1, 0, bias=False),
mynn.AffineChannel2d(oup),
# nn.BatchNorm2d(oup),
)
def basic_bn_stem():
if cfg.RESNETS.SSN:
return nn.Sequential(
OrderedDict([
('conv1', nn.Conv2d(3, 64, 7, stride=2, padding=3,
bias=False)),
('bn1', SSN2d(64, using_moving_average=True)),
('relu', nn.ReLU(inplace=True)),
# ('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True))]))
('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))
]))
else:
if cfg.REID.BN_COMPLETE:
return nn.Sequential(
OrderedDict([
('conv1',
nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=False)),
('bn1', nn.BatchNorm2d(64)),
('relu', nn.ReLU(inplace=True)),
# ('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True))]))
('maxpool', nn.MaxPool2d(kernel_size=3,
stride=2,
padding=1))
]))
else:
return nn.Sequential(
OrderedDict([
('conv1',
nn.Conv2d(3, 64, 7, stride=2, padding=3, bias=False)),
('bn1', mynn.AffineChannel2d(64)),
('relu', nn.ReLU(inplace=True)),
# ('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=0, ceil_mode=True))]))
('maxpool', nn.MaxPool2d(kernel_size=3,
stride=2,
padding=1))
]))
def __init__(self,
inplanes,
outplanes,
innerplanes,
stride=1,
dilation=1,
group=1,
downsample=None):
super().__init__()
# In original resnet, stride=2 is on 1x1.
# In fb.torch resnet, stride=2 is on 3x3.
(str1x1, str3x3) = (stride, 1) if cfg.RESNETS.STRIDE_1X1 else (1,
stride)
self.stride = stride
self.conv1 = nn.Conv2d(inplanes,
innerplanes,
kernel_size=1,
stride=str1x1,
bias=False)
self.bn1 = mynn.AffineChannel2d(innerplanes)
# DCN 使用c3-c5层
if cfg.USE_DCN and innerplanes in cfg.INNERPLANES_LIST:
deformable_groups = 1
offset_channels = 18
conv_op = DeformConv
# self.conv2_deformable
self.conv2_offset = nn.Conv2d(innerplanes,
deformable_groups * offset_channels,
kernel_size=3,
stride=str3x3,
padding=dilation,
dilation=dilation)
self.conv2 = conv_op(innerplanes,
innerplanes,
kernel_size=3,
stride=str3x3,
padding=dilation,
dilation=dilation,
deformable_groups=deformable_groups,
bias=False)
# print(innerplanes, str3x3)
else:
# DCN模式下 c2层加载参数
self.conv2 = nn.Conv2d(innerplanes,
innerplanes,
kernel_size=3,
stride=str3x3,
bias=False,
padding=1 * dilation,
dilation=dilation,
groups=group)
self.bn2 = mynn.AffineChannel2d(innerplanes)
self.conv3 = nn.Conv2d(innerplanes,
outplanes,
kernel_size=1,
stride=1,
bias=False)
self.bn3 = mynn.AffineChannel2d(outplanes)
self.downsample = downsample
self.relu = nn.ReLU(inplace=True)
if cfg.USE_DCN and innerplanes in cfg.INNERPLANES_LIST:
self._init_weights()
def __init__(self,
inplanes,
outplanes,
innerplanes,
stride=1,
dilation=1,
group=1,
downsample=None):
super().__init__()
# In original resnet, stride=2 is on 1x1.
# In fb.torch resnet, stride=2 is on 3x3.
(str1x1, str3x3) = (stride, 1) if cfg.RESNETS.STRIDE_1X1 else (1,
stride)
self.stride = stride
self.conv1 = nn.Conv2d(inplanes,
innerplanes,
kernel_size=1,
stride=str1x1,
bias=False)
if cfg.RESNETS.SSN:
self.bn1 = SSN2d(innerplanes, using_moving_average=True)
else:
if cfg.REID.BN_COMPLETE:
self.bn1 = nn.BatchNorm2d(innerplanes)
else:
self.bn1 = mynn.AffineChannel2d(innerplanes)
self.conv2 = nn.Conv2d(innerplanes,
innerplanes,
kernel_size=3,
stride=str3x3,
bias=False,
padding=1 * dilation,
dilation=dilation,
groups=group)
if cfg.RESNETS.SSN:
self.bn2 = SSN2d(innerplanes, using_moving_average=True)
else:
if cfg.REID.BN_COMPLETE:
self.bn2 = nn.BatchNorm2d(innerplanes)
else:
self.bn2 = mynn.AffineChannel2d(innerplanes)
self.conv3 = nn.Conv2d(innerplanes,
outplanes,
kernel_size=1,
stride=1,
bias=False)
if cfg.RESNETS.SSN:
self.bn3 = SSN2d(outplanes,
using_moving_average=True,
last_gamma=True)
else:
if cfg.REID.BN_COMPLETE:
self.bn3 = nn.BatchNorm2d(outplanes)
else:
self.bn3 = mynn.AffineChannel2d(outplanes)
self.downsample = downsample
self.relu = nn.ReLU(inplace=True)
if cfg.REID.ZERO_GAMMA:
self._init_weights()
