def _init_layers(self):
self.levels = len(self.anchor_strides)
self.relu = nn.ReLU(inplace=True)
self.cls_convs = nn.ModuleList()
self.reg_convs = nn.ModuleList()
self.cls_bn_levels = nn.ModuleList()
self.reg_bn_levels = nn.ModuleList()
for i in range(self.stacked_convs):
chn = self.in_channels if i == 0 else self.feat_channels
self.cls_convs.append(
SeparableConv(chn, self.feat_channels, bias=True, relu=False))
cls_bn_level = nn.ModuleList()
for l in range(self.levels):
_, bn_layer = build_norm_layer(self.norm_cfg,
self.feat_channels)
cls_bn_level.append(nn.Sequential(bn_layer, Swish()))
self.cls_bn_levels.append(cls_bn_level)
self.reg_convs.append(
SeparableConv(chn, self.feat_channels, bias=True, relu=False))
reg_bn_level = nn.ModuleList()
for l in range(self.levels):
_, bn_layer = build_norm_layer(self.norm_cfg,
self.feat_channels)
reg_bn_level.append(nn.Sequential(bn_layer, Swish()))
self.reg_bn_levels.append(reg_bn_level)
self.retina_cls = SeparableConv(self.feat_channels,
self.num_anchors *
self.cls_out_channels,
bias=True,
relu=False)
self.retina_reg = SeparableConv(self.feat_channels,
self.num_anchors * 4,
bias=True,
relu=False)
def _make_stem_layer(self, in_channels):
self.conv1 = nn.Sequential(
build_conv_layer(self.conv_cfg,
in_channels,
32,
kernel_size=3,
stride=2,
padding=1,
bias=False),
build_norm_layer(self.norm_cfg, 32)[1],
nn.ReLU(inplace=True),
build_conv_layer(self.conv_cfg,
32,
32,
kernel_size=3,
stride=1,
padding=1,
bias=False),
build_norm_layer(self.norm_cfg, 32)[1],
nn.ReLU(inplace=True),
build_conv_layer(self.conv_cfg,
32,
64,
kernel_size=3,
stride=1,
padding=1,
bias=False),
)
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
self.add_module(self.norm1_name, norm1)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
def _make_fuse_layers(self):
if self.num_branches == 1:
return None
num_branches = self.num_branches
in_channels = self.in_channels
fuse_layers = []
num_out_branches = num_branches if self.multiscale_output else 1
for i in range(num_out_branches):
fuse_layer = []
for j in range(num_branches):
if j > i:
fuse_layer.append(
nn.Sequential(
build_conv_layer(self.conv_cfg,
in_channels[j],
in_channels[i],
kernel_size=1,
stride=1,
padding=0,
bias=False),
build_norm_layer(self.norm_cfg, in_channels[i])[1],
nn.Upsample(scale_factor=2**(j - i),
mode='nearest')))
elif j == i:
fuse_layer.append(None)
else:
conv_downsamples = []
for k in range(i - j):
if k == i - j - 1:
conv_downsamples.append(
nn.Sequential(
build_conv_layer(self.conv_cfg,
in_channels[j],
in_channels[i],
kernel_size=3,
stride=2,
padding=1,
bias=False),
build_norm_layer(self.norm_cfg,
in_channels[i])[1]))
else:
conv_downsamples.append(
nn.Sequential(
build_conv_layer(self.conv_cfg,
in_channels[j],
in_channels[j],
kernel_size=3,
stride=2,
padding=1,
bias=False),
build_norm_layer(self.norm_cfg,
in_channels[j])[1],
nn.ReLU(inplace=False)))
fuse_layer.append(nn.Sequential(*conv_downsamples))
fuse_layers.append(nn.ModuleList(fuse_layer))
return nn.ModuleList(fuse_layers)
def _make_transition_layer(self, num_channels_pre_layer,
num_channels_cur_layer):
num_branches_cur = len(num_channels_cur_layer)
num_branches_pre = len(num_channels_pre_layer)
transition_layers = []
for i in range(num_branches_cur):
if i < num_branches_pre:
if num_channels_cur_layer[i] != num_channels_pre_layer[i]:
transition_layers.append(
nn.Sequential(
build_conv_layer(
self.conv_cfg,
num_channels_pre_layer[i],
num_channels_cur_layer[i],
kernel_size=3,
stride=1,
padding=1,
bias=False,
),
build_norm_layer(self.norm_cfg,
num_channels_cur_layer[i])[1],
nn.ReLU(inplace=True),
))
else:
transition_layers.append(None)
else:
conv_downsamples = []
for j in range(i + 1 - num_branches_pre):
in_channels = num_channels_pre_layer[-1]
out_channels = (num_channels_cur_layer[i] if j == i -
num_branches_pre else in_channels)
conv_downsamples.append(
nn.Sequential(
build_conv_layer(
self.conv_cfg,
in_channels,
out_channels,
kernel_size=3,
stride=2,
padding=1,
bias=False,
),
build_norm_layer(self.norm_cfg, out_channels)[1],
nn.ReLU(inplace=True),
))
transition_layers.append(nn.Sequential(*conv_downsamples))
return nn.ModuleList(transition_layers)
def __init__(self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
gen_attention=None):
super(BasicBlock, self).__init__()
assert dcn is None, 'Not implemented yet.'
assert gen_attention is None, 'Not implemented yet.'
assert gcb is None, 'Not implemented yet.'
self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
self.activation = None
if 'inplace' not in norm_cfg['type']:
self.activation = nn.ReLU(inplace=True)
else:
norm2.activation = "identity"
self.conv1 = build_conv_layer(conv_cfg,
inplanes,
planes,
3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False)
self.add_module(self.norm1_name, norm1)
self.conv2 = build_conv_layer(conv_cfg,
planes,
planes,
3,
padding=1,
bias=False)
self.add_module(self.norm2_name, norm2)
self.downsample = downsample
self.stride = stride
self.dilation = dilation
assert not with_cp
def __init__(self,
in_channels,
out_channels,
num_ins,
conv_cfg=None,
norm_cfg=None,
separable_conv=True,
act_cfg=None,
eps=0.0001):
super(WeightedInputConv_V2, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.num_ins = num_ins
self.eps = eps
if separable_conv:
_, bn_layer = build_norm_layer(norm_cfg, out_channels)
self.conv_op = nn.Sequential(
SeparableConv(in_channels, out_channels, bias=True,
relu=False), bn_layer)
else:
self.conv_op = ConvModule(in_channels,
out_channels,
3,
padding=1,
conv_cfg=None,
norm_cfg=norm_cfg,
act_cfg=None,
inplace=False)
# edge weight and swish
self.weight = nn.Parameter(torch.Tensor(self.num_ins).fill_(1.0))
self._swish = Swish()
def _make_layer(self, block, inplanes, planes, blocks, stride=1):
downsample = None
if stride != 1 or inplanes != planes * block.expansion:
downsample = nn.Sequential(
build_conv_layer(self.conv_cfg,
inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False),
build_norm_layer(self.norm_cfg, planes * block.expansion)[1])
layers = []
layers.append(
block(inplanes,
planes,
stride,
downsample=downsample,
with_cp=self.with_cp,
norm_cfg=self.norm_cfg,
conv_cfg=self.conv_cfg))
inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(inplanes,
planes,
with_cp=self.with_cp,
norm_cfg=self.norm_cfg,
conv_cfg=self.conv_cfg))
return nn.Sequential(*layers)
def __init__(
self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style="pytorch",
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type="BN"),
dcn=None,
gcb=None,
gen_attention=None,
):
super(BasicBlock, self).__init__()
assert dcn is None, "Not implemented yet."
assert gen_attention is None, "Not implemented yet."
assert gcb is None, "Not implemented yet."
self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
self.conv1 = build_conv_layer(
conv_cfg,
inplanes,
planes,
3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False,
)
self.add_module(self.norm1_name, norm1)
self.conv2 = build_conv_layer(conv_cfg,
planes,
planes,
3,
padding=1,
bias=False)
self.add_module(self.norm2_name, norm2)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
self.dilation = dilation
assert not with_cp
def make_res_layer(block,
inplanes,
planes,
blocks,
stride=1,
dilation=1,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
gen_attention=None,
gen_attention_blocks=[]):
downsample = None
if stride != 1 or inplanes != planes * block.expansion:
downsample = nn.Sequential(
build_conv_layer(conv_cfg,
inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False),
build_norm_layer(norm_cfg, planes * block.expansion)[1],
)
layers = []
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=stride,
dilation=dilation,
downsample=downsample,
style=style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention if
(0 in gen_attention_blocks) else None))
inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=1,
dilation=dilation,
style=style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention if
(i in gen_attention_blocks) else None))
return nn.Sequential(*layers)
def _make_stem_layer(self, in_channels):
conv1 = build_conv_layer(
self.conv_cfg,
in_channels,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
relu = nn.ReLU(inplace=True)
maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer0 = nn.Sequential(*[conv1, norm1, relu, maxpool])
def __init__(self, block_args, global_params, norm_cfg):
super().__init__()
self._block_args = block_args
self._bn_mom = 1 - global_params.batch_norm_momentum
self._bn_eps = global_params.batch_norm_epsilon
self.has_se = (self._block_args.se_ratio is not None) and (0 < self._block_args.se_ratio <= 1)
self.id_skip = block_args.id_skip # skip connection and drop connect
# Get static or dynamic convolution depending on image size
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
# Expansion phase
inp = self._block_args.input_filters # number of input channels
oup = self._block_args.input_filters * self._block_args.expand_ratio # number of output channels
if self._block_args.expand_ratio != 1:
self._expand_conv = Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)
self._bn0 = build_norm_layer(norm_cfg, num_features=oup, postfix=0)[1]
# Depthwise convolution phase
k = self._block_args.kernel_size
s = self._block_args.stride
self._depthwise_conv = Conv2d(
in_channels=oup, out_channels=oup, groups=oup, # groups makes it depthwise
kernel_size=k, stride=s, bias=False)
self._bn1 = build_norm_layer(norm_cfg, num_features=oup, postfix=1)[1]
# Squeeze and Excitation layer, if desired
if self.has_se:
num_squeezed_channels = max(1, int(self._block_args.input_filters * self._block_args.se_ratio))
self._se_reduce = Conv2d(in_channels=oup, out_channels=num_squeezed_channels, kernel_size=1)
self._se_expand = Conv2d(in_channels=num_squeezed_channels, out_channels=oup, kernel_size=1)
# Output phase
final_oup = self._block_args.output_filters
self._project_conv = Conv2d(in_channels=oup, out_channels=final_oup, kernel_size=1, bias=False)
self._bn2 = build_norm_layer(norm_cfg, num_features=final_oup, postfix=2)[1]
self._swish = MemoryEfficientSwish()
def _make_stem_layer(self, in_channels):
self.conv1 = build_conv_layer(self.conv_cfg,
in_channels,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
self.add_module(self.norm1_name, norm1)
self.activation = None
if 'InPlace' not in self.norm_cfg['type']:
self.activation = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
def _make_one_branch(self,
branch_index,
block,
num_blocks,
num_channels,
stride=1,
dcn=None,
gcb=None):
downsample = None
if stride != 1 or \
self.in_channels[branch_index] != \
num_channels[branch_index] * block.expansion:
downsample = nn.Sequential(
build_conv_layer(self.conv_cfg,
self.in_channels[branch_index],
num_channels[branch_index] * block.expansion,
kernel_size=1,
stride=stride,
bias=False),
build_norm_layer(self.norm_cfg, num_channels[branch_index] *
block.expansion)[1])
layers = []
layers.append(
block(self.in_channels[branch_index],
num_channels[branch_index],
stride,
downsample=downsample,
with_cp=self.with_cp,
norm_cfg=self.norm_cfg,
conv_cfg=self.conv_cfg,
dcn=dcn,
gcb=gcb), )
self.in_channels[branch_index] = \
num_channels[branch_index] * block.expansion
for i in range(1, num_blocks[branch_index]):
layers.append(
block(self.in_channels[branch_index],
num_channels[branch_index],
with_cp=self.with_cp,
norm_cfg=self.norm_cfg,
conv_cfg=self.conv_cfg,
dcn=dcn,
gcb=gcb))
return nn.Sequential(*layers)
def __init__(self,
num_repeat,
connect_norm_eval,
use_act=True,
backbone_type='ResNet',
**backbone_args):
super(CBNet, self).__init__()
assert backbone_type in self.arch_cfg
self.backbone_type = backbone_type
self.num_repeat = num_repeat
self.connect_norm_eval = connect_norm_eval
self.backbone_names = []
self.connect_op_names = []
self.use_act = use_act
assert num_repeat >= 2
for i in range(1, num_repeat + 1):
backbone = self.arch_cfg[backbone_type](**backbone_args)
backbone_name = 'cb{}'.format(i)
self.add_module(backbone_name, backbone)
self.backbone_names.append(backbone_name)
left_out_channels = [256, 512, 1024, 2048]
right_in_channels = [64, 256, 512, 1024]
for i, _ in enumerate(left_out_channels):
conv = build_conv_layer(backbone.conv_cfg,
left_out_channels[i],
right_in_channels[i],
kernel_size=1,
padding=0,
stride=1,
bias=False)
# constant_init(conv, 0)
_, norm = build_norm_layer(backbone.norm_cfg, right_in_channels[i])
# BatchNorm
# default value of running_mean is 0
# default value of running_var is 1
# default weight is 1
# default bias is 0
# default ((x - 0)/1)* weight + bias = x,
# in eval mode, running_mean and running_var will not be updated.
# if requires_grad, weight and bias will be updated.
connect_op = nn.Sequential(conv, norm)
connect_op_name = 'connect_op{}'.format(i + 1)
self.add_module(connect_op_name, connect_op)
self.connect_op_names.append(connect_op_name)
def _make_dynamic_gate_layers(self, num_backbones=1):
if self.num_branches < 4:
return None
num_branches = self.num_branches
gate_num = num_branches
in_channels = self.in_channels
gate_layers_backone = []
for num_backbone in range(num_backbones):
gate_layers = []
for i in range(num_branches):
# if i==0 or i==num_branches-1: # only keep and down
# gate_num=2
# else:
# gate_num=3
gate_layers.append(
nn.Sequential(
build_conv_layer(
self.conv_cfg,
in_channels[i],
in_channels[i] // 2, # //2
kernel_size=1,
stride=1,
padding=0,
bias=False),
build_norm_layer(self.norm_cfg,
in_channels[i] // 2)[1], # //2
nn.ReLU(inplace=False),
nn.AdaptiveAvgPool2d((1, 1)),
build_conv_layer(
self.conv_cfg,
in_channels[i] // 2, # //2
gate_num,
kernel_size=1,
stride=1,
padding=0,
bias=True)))
gate_layers_backone.append(nn.ModuleList(gate_layers))
return nn.ModuleList(gate_layers_backone)
def _make_fuse_layers(self):
if self.num_branches == 1:
return None
num_branches = self.num_branches
in_channels = self.in_channels
fuse_layers = []
num_out_branches = num_branches if self.multiscale_output else 1
for i in range(num_out_branches): # i指融合后，j指融合前
fuse_layer = []
for j in range(num_branches):
if j > i: # resolution-up and dim-down 1x1降维再升高分辨率
fuse_layer.append(
nn.Sequential(
build_conv_layer(self.conv_cfg,
in_channels[j],
in_channels[i],
kernel_size=1,
stride=1,
padding=0,
bias=False),
build_norm_layer(self.norm_cfg, in_channels[i])[1],
nn.Upsample(scale_factor=2**(j - i),
mode='nearest')))
elif j == i: # 同层直接运算
fuse_layer.append(None)
else: # j<i
conv_downsamples = []
for k in range(i - j):
if k == i - j - 1:
conv_downsamples.append(
nn.Sequential(
build_conv_layer( # 3x3 conv+stride=2
self.conv_cfg,
in_channels[j],
in_channels[i],
kernel_size=3,
stride=2,
padding=1,
bias=False),
build_norm_layer(self.norm_cfg,
in_channels[i])[1]))
else:
conv_downsamples.append(
nn.Sequential(
build_conv_layer(
self.conv_cfg,
in_channels[j],
in_channels[
j], # downsample but dim not change -> lose some info
kernel_size=3,
stride=2,
padding=1,
bias=False),
build_norm_layer(self.norm_cfg,
in_channels[j])[1],
nn.ReLU(inplace=False)))
fuse_layer.append(nn.Sequential(*conv_downsamples))
fuse_layers.append(nn.ModuleList(fuse_layer))
return nn.ModuleList(fuse_layers)
def __init__(self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
gen_attention=None):
"""Bottleneck block for ResNet.
If style is "pytorch", the stride-two layer is the 3x3 conv layer, if
it is "caffe", the stride-two layer is the first 1x1 conv layer.
"""
super(Bottleneck, self).__init__()
assert style in ['pytorch', 'caffe']
assert dcn is None or isinstance(dcn, dict)
assert gcb is None or isinstance(gcb, dict)
assert gen_attention is None or isinstance(gen_attention, dict)
self.inplanes = inplanes
self.planes = planes
self.stride = stride
self.dilation = dilation
self.style = style
self.with_cp = with_cp
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.dcn = dcn
self.with_dcn = dcn is not None
self.gcb = gcb
self.with_gcb = gcb is not None
self.gen_attention = gen_attention
self.with_gen_attention = gen_attention is not None
if self.style == 'pytorch':
self.conv1_stride = 1
self.conv2_stride = stride
else:
self.conv1_stride = stride
self.conv2_stride = 1
self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
self.norm3_name, norm3 = build_norm_layer(norm_cfg,
planes * self.expansion,
postfix=3)
self.conv1 = build_conv_layer(conv_cfg,
inplanes,
planes,
kernel_size=1,
stride=self.conv1_stride,
bias=False)
self.add_module(self.norm1_name, norm1)
fallback_on_stride = False
if self.with_dcn:
fallback_on_stride = dcn.pop('fallback_on_stride', False)
if not self.with_dcn or fallback_on_stride:
self.conv2 = build_conv_layer(conv_cfg,
planes,
planes,
kernel_size=3,
stride=self.conv2_stride,
padding=dilation,
dilation=dilation,
bias=False)
else:
assert self.conv_cfg is None, 'conv_cfg cannot be None for DCN'
self.conv2 = build_conv_layer(dcn,
planes,
planes,
kernel_size=3,
stride=self.conv2_stride,
padding=dilation,
dilation=dilation,
bias=False)
self.add_module(self.norm2_name, norm2)
self.conv3 = build_conv_layer(conv_cfg,
planes,
planes * self.expansion,
kernel_size=1,
bias=False)
self.add_module(self.norm3_name, norm3)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
if self.with_gcb:
gcb_inplanes = planes * self.expansion
self.context_block = ContextBlock(inplanes=gcb_inplanes, **gcb)
# gen_attention
if self.with_gen_attention:
self.gen_attention_block = GeneralizedAttention(
planes, **gen_attention)
def __init__(self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
gen_attention=None):
# super(BasicBlockPlugins, self).__init__()
super().__init__()
assert not with_cp
self.dcn = dcn
self.with_dcn = dcn is not None
self.gcb = gcb
self.with_gcb = gcb is not None
# self.gen_attention = gen_attention
self.with_gen_attention = gen_attention is not None
self.downsample = downsample
self.stride = stride
self.dilation = dilation
self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
fallback_on_stride = False
if self.with_dcn:
fallback_on_stride = dcn.pop('fallback_on_stride', False)
if not self.with_dcn or fallback_on_stride:
self.conv1 = build_conv_layer(conv_cfg,
inplanes,
planes,
3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False)
else:
assert conv_cfg is None
self.conv1 = build_conv_layer(dcn,
inplanes,
planes,
kernel_size=3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False)
self.add_module(self.norm1_name, norm1)
self.conv2 = build_conv_layer(conv_cfg,
planes,
planes,
3,
padding=1,
bias=False)
self.add_module(self.norm2_name, norm2)
self.relu = nn.ReLU(inplace=True)
if self.with_gcb:
self.context_block = ContextBlock(inplanes=planes, **gcb)
def __init__(self, inplanes, planes, groups=1, base_width=4, **kwargs):
"""Bottleneck block for ResNeXt.
If style is "pytorch", the stride-two layer is the 3x3 conv layer, if
it is "caffe", the stride-two layer is the first 1x1 conv layer.
"""
super(Bottleneck, self).__init__(inplanes, planes, **kwargs)
if groups == 1:
width = self.planes
else:
width = math.floor(self.planes * (base_width / 64)) * groups
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg,
width,
postfix=1)
self.norm2_name, norm2 = build_norm_layer(self.norm_cfg,
width,
postfix=2)
self.norm3_name, norm3 = build_norm_layer(self.norm_cfg,
self.planes * self.expansion,
postfix=3)
self.conv1 = build_conv_layer(self.conv_cfg,
self.inplanes,
width,
kernel_size=1,
stride=self.conv1_stride,
bias=False)
self.add_module(self.norm1_name, norm1)
fallback_on_stride = False
self.with_modulated_dcn = False
if self.with_dcn:
fallback_on_stride = self.dcn.pop('fallback_on_stride', False)
if not self.with_dcn or fallback_on_stride:
self.conv2 = build_conv_layer(self.conv_cfg,
width,
width,
kernel_size=3,
stride=self.conv2_stride,
padding=self.dilation,
dilation=self.dilation,
groups=groups,
bias=False)
else:
assert self.conv_cfg is None, 'conv_cfg must be None for DCN'
self.conv2 = build_conv_layer(self.dcn,
width,
width,
kernel_size=3,
stride=self.conv2_stride,
padding=self.dilation,
dilation=self.dilation,
groups=groups,
bias=False)
self.add_module(self.norm2_name, norm2)
self.conv3 = build_conv_layer(self.conv_cfg,
width,
self.planes * self.expansion,
kernel_size=1,
bias=False)
self.add_module(self.norm3_name, norm3)
def __init__(self,
arch='efficientnet-b0',
out_levels=[3, 4, 5],
norm_cfg=dict(type='SyncBN',
momentum=0.01,
eps=1e-3,
requires_grad=False),
norm_eval=True,
override_params=None):
super(EfficientNet, self).__init__()
self._check_model_name_is_valid(arch)
blocks_args, global_params = get_model_params(arch, override_params)
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
self.norm_eval = norm_eval
self.out_levels = out_levels
self.norm_cfg = norm_cfg
# Get stem static or dynamic convolution depending on image size
image_size = global_params.image_size
Conv2d = get_same_padding_conv2d(image_size=image_size)
# Batch norm parameters
bn_mom = 1 - self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
# Stem
in_channels = 3 # rgb
out_channels = round_filters(
32, self._global_params) # number of output channels
self._conv_stem = Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=2,
bias=False)
# self._bn0 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
_, self._bn0 = build_norm_layer(norm_cfg, out_channels)
image_size = calculate_output_image_size(image_size, 2)
# Build blocks
self._blocks = nn.ModuleList([])
for block_args in self._blocks_args:
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters,
self._global_params),
output_filters=round_filters(block_args.output_filters,
self._global_params),
num_repeat=round_repeats(block_args.num_repeat,
self._global_params))
# The first block needs to take care of stride and filter size increase.
self._blocks.append(
MBConvBlock(block_args,
self._global_params,
image_size=image_size,
norm_cfg=self.norm_cfg))
image_size = calculate_output_image_size(image_size,
block_args.stride)
if block_args.num_repeat > 1: # modify block_args to keep same output size
block_args = block_args._replace(
input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(
MBConvBlock(block_args,
self._global_params,
image_size=image_size,
norm_cfg=self.norm_cfg))
# last_stage_idx
self._swish = MemoryEfficientSwish()
def __init__(self,
block_args,
global_params,
image_size=None,
norm_cfg=dict(type="BN",
momentum=0.01,
eps=1e-3,
requires_grad=False)):
super().__init__()
self._block_args = block_args
# print(block_args)
self._bn_mom = 1 - global_params.batch_norm_momentum # pytorch's difference from tensorflow
self._bn_eps = global_params.batch_norm_epsilon
self.has_se = (self._block_args.se_ratio
is not None) and (0 < self._block_args.se_ratio <= 1)
self.id_skip = block_args.id_skip # whether to use skip connection and drop connect
self.norm_cfg = norm_cfg
# Expansion phase (Inverted Bottleneck)
inp = self._block_args.input_filters # number of input channels
oup = self._block_args.input_filters * self._block_args.expand_ratio # number of output channels
if self._block_args.expand_ratio != 1:
Conv2d = get_same_padding_conv2d(image_size=image_size)
self._expand_conv = Conv2d(in_channels=inp,
out_channels=oup,
kernel_size=1,
bias=False)
# self._bn0 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
_, self._bn0 = build_norm_layer(norm_cfg, oup)
# image_size = calculate_output_image_size(image_size, 1) <-- this wouldn't modify image_size
# Depthwise convolution phase
k = self._block_args.kernel_size
s = self._block_args.stride
Conv2d = get_same_padding_conv2d(image_size=image_size)
self._depthwise_conv = Conv2d(
in_channels=oup,
out_channels=oup,
groups=oup, # groups makes it depthwise
kernel_size=k,
stride=s,
bias=False)
# self._bn1 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
_, self._bn1 = build_norm_layer(norm_cfg, oup)
image_size = calculate_output_image_size(image_size, s)
# Squeeze and Excitation layer, if desired
if self.has_se:
Conv2d = get_same_padding_conv2d(image_size=(1, 1))
num_squeezed_channels = max(
1,
int(self._block_args.input_filters *
self._block_args.se_ratio))
self._se_reduce = Conv2d(in_channels=oup,
out_channels=num_squeezed_channels,
kernel_size=1)
self._se_expand = Conv2d(in_channels=num_squeezed_channels,
out_channels=oup,
kernel_size=1)
# Pointwise convolution phase
final_oup = self._block_args.output_filters
Conv2d = get_same_padding_conv2d(image_size=image_size)
self._project_conv = Conv2d(in_channels=oup,
out_channels=final_oup,
kernel_size=1,
bias=False)
# self._bn2 = nn.BatchNorm2d(num_features=final_oup, momentum=self._bn_mom, eps=self._bn_eps)
_, self._bn2 = build_norm_layer(norm_cfg, final_oup)
self._swish = MemoryEfficientSwish()
def __init__(self,
extra,
in_channels=3,
conv_cfg=None,
norm_cfg=dict(type='BN'),
norm_eval=True,
with_cp=False,
zero_init_residual=False,
with_dcn=None,
with_gcb=None,
dcn=None,
gcb=None):
super(HRNet, self).__init__()
self.extra = extra
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.zero_init_residual = zero_init_residual
# stem net
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
self.norm2_name, norm2 = build_norm_layer(self.norm_cfg, 64, postfix=2)
self.conv1 = build_conv_layer(self.conv_cfg,
in_channels,
64,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.add_module(self.norm1_name, norm1)
self.conv2 = build_conv_layer(self.conv_cfg,
64,
64,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.add_module(self.norm2_name, norm2)
self.relu = nn.ReLU(inplace=True)
# stage 1
self.stage1_cfg = self.extra['stage1']
num_channels = self.stage1_cfg['num_channels'][0]
block_type = self.stage1_cfg['block']
num_blocks = self.stage1_cfg['num_blocks'][0]
block = self.blocks_dict[block_type]
stage1_out_channels = num_channels * block.expansion
self.layer1 = self._make_layer(block, 64, num_channels, num_blocks)
# stage 2
self.stage2_cfg = self.extra['stage2']
num_channels = self.stage2_cfg['num_channels']
block_type = self.stage2_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition1 = self._make_transition_layer([stage1_out_channels],
num_channels)
self.stage2, pre_stage_channels = self._make_stage(
self.stage2_cfg, num_channels)
# stage 3
self.stage3_cfg = self.extra['stage3']
num_channels = self.stage3_cfg['num_channels']
block_type = self.stage3_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition2 = self._make_transition_layer(pre_stage_channels,
num_channels)
self.stage3, pre_stage_channels = self._make_stage(
self.stage3_cfg, num_channels)
# stage 4
self.stage4_cfg = self.extra['stage4']
num_channels = self.stage4_cfg['num_channels']
block_type = self.stage4_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition3 = self._make_transition_layer(pre_stage_channels,
num_channels)
dcn_branches_repeats, gcb_branches_repeats = None, None
if with_dcn is not None:
assert len(with_dcn) == self.stage4_cfg['num_modules']
dcn_branches_repeats = [
[] for _ in range(self.stage4_cfg['num_modules'])
]
for repeatid, dcn_branches in enumerate(with_dcn):
assert len(dcn_branches) == self.stage4_cfg['num_branches']
for use_dcn in dcn_branches:
if use_dcn is False:
dcn_branches_repeats[repeatid].append(None)
else:
assert isinstance(dcn, dict)
dcn_branches_repeats[repeatid].append(dcn)
if with_gcb is not None:
assert len(with_gcb) == self.stage4_cfg['num_modules']
gcb_branches_repeats = [
[] for _ in range(self.stage4_cfg['num_modules'])
]
for repeatid, gcb_branches in enumerate(gcb_branches_repeats):
assert len(gcb_branches) == self.stage4_cfg['num_branches']
for use_gcb in gcb_branches:
if use_gcb is False:
gcb_branches_repeats[repeatid].append(None)
else:
assert isinstance(gcb, dict)
gcb_branches_repeats[repeatid].append(gcb)
self.stage4, pre_stage_channels = self._make_stage(
self.stage4_cfg,
num_channels,
dcn_branches_repeats=dcn_branches_repeats,
gcb_branches_repeats=gcb_branches_repeats) # NOTE
def make_res_layer(block,
inplanes,
planes,
blocks,
stride=1,
dilation=1,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
gen_attention=None,
gen_attention_blocks=[],
scale=4,
baseWidth=26):
downsample = None
if stride != 1 or inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.AvgPool2d(kernel_size=stride,
stride=stride,
ceil_mode=True,
count_include_pad=False),
build_conv_layer(conv_cfg,
inplanes,
planes * block.expansion,
kernel_size=1,
stride=1,
bias=False),
build_norm_layer(norm_cfg, planes * block.expansion)[1],
)
layers = []
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=stride,
dilation=dilation,
downsample=downsample,
style=style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention if
(0 in gen_attention_blocks) else None,
scale=scale,
baseWidth=baseWidth,
stype='stage'))
inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=1,
dilation=dilation,
style=style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention if
(i in gen_attention_blocks) else None,
scale=scale,
baseWidth=baseWidth))
return nn.Sequential(*layers)
def __init__(self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
gen_attention=None,
scale=4,
baseWidth=26,
stype='normal'):
"""Bottle2neck block for Res2Net.
If style is "pytorch", the stride-two layer is the 3x3 conv layer,
if it is "caffe", the stride-two layer is the first 1x1 conv layer.
"""
super(Bottle2neck, self).__init__()
assert style in ['pytorch', 'caffe']
assert dcn is None or isinstance(dcn, dict)
assert gcb is None or isinstance(gcb, dict)
assert gen_attention is None or isinstance(gen_attention, dict)
width = int(math.floor(planes * (baseWidth / 64.0)))
self.inplanes = inplanes
self.planes = planes
self.stride = stride
self.dilation = dilation
self.style = style
self.with_cp = with_cp
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.dcn = dcn
self.with_dcn = dcn is not None
self.gcb = gcb
self.with_gcb = gcb is not None
self.gen_attention = gen_attention
self.with_gen_attention = gen_attention is not None
if self.style == 'pytorch':
self.conv1_stride = 1
self.conv2_stride = stride
else:
self.conv1_stride = stride
self.conv2_stride = 1
self.norm1_name, norm1 = build_norm_layer(norm_cfg,
width * scale,
postfix=1)
self.norm3_name, norm3 = build_norm_layer(norm_cfg,
planes * self.expansion,
postfix=3)
self.conv1 = build_conv_layer(conv_cfg,
inplanes,
width * scale,
kernel_size=1,
stride=self.conv1_stride,
bias=False)
self.add_module(self.norm1_name, norm1)
if scale == 1:
self.nums = 1
else:
self.nums = scale - 1
if stype == 'stage':
self.pool = nn.AvgPool2d(kernel_size=3, stride=stride, padding=1)
convs = []
bns = []
fallback_on_stride = False
if self.with_dcn:
fallback_on_stride = dcn.pop('fallback_on_stride', False)
if not self.with_dcn or fallback_on_stride:
for i in range(self.nums):
convs.append(
build_conv_layer(conv_cfg,
width,
width,
kernel_size=3,
stride=self.conv2_stride,
padding=dilation,
dilation=dilation,
bias=False))
bns.append(build_norm_layer(norm_cfg, width, postfix=i + 1)[1])
self.convs = nn.ModuleList(convs)
self.bns = nn.ModuleList(bns)
else:
assert self.conv_cfg is None, 'conv_cfg cannot be None for DCN'
for i in range(self.nums):
convs.append(
build_conv_layer(dcn,
width,
width,
kernel_size=3,
stride=self.conv2_stride,
padding=dilation,
dilation=dilation,
bias=False))
bns.append(build_norm_layer(norm_cfg, width, postfix=i + 1)[1])
self.convs = nn.ModuleList(convs)
self.bns = nn.ModuleList(bns)
self.conv3 = build_conv_layer(conv_cfg,
width * scale,
planes * self.expansion,
kernel_size=1,
bias=False)
self.add_module(self.norm3_name, norm3)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stype = stype
self.scale = scale
self.width = width
if self.with_gcb:
gcb_inplanes = planes * self.expansion
self.context_block = ContextBlock(inplanes=gcb_inplanes, **gcb)
# gen_attention
if self.with_gen_attention:
self.gen_attention_block = GeneralizedAttention(
planes, **gen_attention)
def __init__(self,
depth,
in_channels=3,
num_stages=4,
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1),
out_indices=(0, 1, 2, 3),
style='pytorch',
frozen_stages=-1,
conv_cfg=None,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=True,
dcn=None,
stage_with_dcn=(False, False, False, False),
gcb=None,
stage_with_gcb=(False, False, False, False),
gen_attention=None,
stage_with_gen_attention=((), (), (), ()),
with_cp=False,
zero_init_residual=True):
super(TB_ResNet, self).__init__()
if depth not in self.arch_settings:
raise KeyError('invalid depth {} for resnet'.format(depth))
self.depth = depth
self.num_stages = num_stages
assert num_stages >= 1 and num_stages <= 4
self.strides = strides
self.dilations = dilations
assert len(strides) == len(dilations) == num_stages
self.out_indices = out_indices
assert max(out_indices) < num_stages
self.style = style
self.frozen_stages = frozen_stages
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.with_cp = with_cp
self.norm_eval = norm_eval
self.dcn = dcn
self.stage_with_dcn = stage_with_dcn
if dcn is not None:
assert len(stage_with_dcn) == num_stages
self.gen_attention = gen_attention
self.gcb = gcb
self.stage_with_gcb = stage_with_gcb
if gcb is not None:
assert len(stage_with_gcb) == num_stages
self.zero_init_residual = zero_init_residual
self.block, stage_blocks = self.arch_settings[depth]
self.stage_blocks = stage_blocks[:num_stages]
self.inplanes = 64
self._make_stem_layer(in_channels)
self.res_layers = []
### add db modules
self.second_res_layers = []
self.eb_second_conv1 = nn.Conv2d(
256,
64,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
self.eb_second_conv2 = nn.Conv2d(
512,
256,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
self.eb_second_conv3 = nn.Conv2d(
1024,
512,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
self.eb_second_conv4 = nn.Conv2d(
2048,
1024,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
#### add tb modules
self.third_res_layers = []
self.eb_third_conv1 = nn.Conv2d(
256,
64,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
self.eb_third_conv2 = nn.Conv2d(
512,
256,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
self.eb_third_conv3 = nn.Conv2d(
1024,
512,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
self.eb_third_conv4 = nn.Conv2d(
2048,
1024,
kernel_size=1,
stride=1,
padding=0,
dilation=1, bias=False)
######### add bn for db
normalize = self.norm_cfg
self.eb_second_bn1_name, eb_second_bn1 = build_norm_layer(normalize, 64, postfix='eb_second_bn1')
self.add_module(self.eb_second_bn1_name, eb_second_bn1)
self.eb_second_bn2_name, eb_second_bn2 = build_norm_layer(normalize, 256, postfix='eb_second_bn2')
self.add_module(self.eb_second_bn2_name, eb_second_bn2)
self.eb_second_bn3_name, eb_second_bn3 = build_norm_layer(normalize, 512, postfix='eb_second_bn3')
self.add_module(self.eb_second_bn3_name, eb_second_bn3)
self.eb_second_bn4_name, eb_second_bn4 = build_norm_layer(normalize, 1024, postfix='eb_second_bn4')
self.add_module(self.eb_second_bn4_name, eb_second_bn4)
######### add bn for tb
self.eb_third_bn1_name, eb_third_bn1 = build_norm_layer(normalize, 64, postfix='eb_third_bn1')
self.add_module(self.eb_third_bn1_name, eb_third_bn1)
self.eb_third_bn2_name, eb_third_bn2 = build_norm_layer(normalize, 256, postfix='eb_third_bn2')
self.add_module(self.eb_third_bn2_name, eb_third_bn2)
self.eb_third_bn3_name, eb_third_bn3 = build_norm_layer(normalize, 512, postfix='eb_third_bn3')
self.add_module(self.eb_third_bn3_name, eb_third_bn3)
self.eb_third_bn4_name, eb_third_bn4 = build_norm_layer(normalize, 1024, postfix='eb_third_bn4')
self.add_module(self.eb_third_bn4_name, eb_third_bn4)
###### upsample in db
self.eb_upsample = nn.Upsample(
scale_factor=2, mode='nearest')
for i, num_blocks in enumerate(self.stage_blocks):
stride = strides[i]
dilation = dilations[i]
dcn = self.dcn if self.stage_with_dcn[i] else None
gcb = self.gcb if self.stage_with_gcb[i] else None
planes = 64 * 2**i
res_layer = tb_make_res_layer(
self.block,
self.inplanes,
planes,
num_blocks,
stride=stride,
dilation=dilation,
style=self.style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention,
gen_attention_blocks=stage_with_gen_attention[i])
second_res_layer = tb_make_res_layer(
self.block,
self.inplanes,
planes,
num_blocks,
stride=stride,
dilation=dilation,
style=self.style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention,
gen_attention_blocks=stage_with_gen_attention[i])
third_res_layer = tb_make_res_layer(
self.block,
self.inplanes,
planes,
num_blocks,
stride=stride,
dilation=dilation,
style=self.style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
gen_attention=gen_attention,
gen_attention_blocks=stage_with_gen_attention[i])
self.inplanes = planes * self.block.expansion
layer_name = 'layer{}'.format(i + 1)
self.add_module(layer_name, res_layer)
self.res_layers.append(layer_name)
#### add dual backbone
second_layer_name = 'second_layer{}'.format(i + 1)
self.add_module(second_layer_name, second_res_layer)
self.second_res_layers.append(second_layer_name)
#### add triple backbone
third_layer_name = 'third_layer{}'.format(i + 1)
self.add_module(third_layer_name, third_res_layer)
self.third_res_layers.append(third_layer_name)
self._freeze_stages()
self.feat_dim = self.block.expansion * 64 * 2**(
len(self.stage_blocks) - 1)
def __init__(self,
extra,
in_channels=3,
conv_cfg=None,
norm_cfg=dict(type='BN'),
norm_eval=True,
with_cp=False,
zero_init_residual=False):
super(HRNet_upsamp, self).__init__()
self.extra = extra
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.zero_init_residual = zero_init_residual
# stem net
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
self.norm2_name, norm2 = build_norm_layer(self.norm_cfg, 64, postfix=2)
self.conv1 = build_conv_layer(self.conv_cfg,
in_channels,
64,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.add_module(self.norm1_name, norm1)
self.conv2 = build_conv_layer(self.conv_cfg,
64,
64,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.add_module(self.norm2_name, norm2)
self.relu = nn.ReLU(inplace=True)
#self.do_upsample = nn.Upsample(scale_factor=2,mode='bilinear')
# stage 1
self.stage1_cfg = self.extra['stage1']
num_channels = self.stage1_cfg['num_channels'][0]
block_type = self.stage1_cfg['block']
num_blocks = self.stage1_cfg['num_blocks'][0]
block = self.blocks_dict[block_type]
stage1_out_channels = num_channels * block.expansion
self.layer1 = self._make_layer(block, 64, num_channels, num_blocks)
# stage 2
self.stage2_cfg = self.extra['stage2']
num_channels = self.stage2_cfg['num_channels']
block_type = self.stage2_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition1 = self._make_transition_layer([stage1_out_channels],
num_channels)
self.stage2, pre_stage_channels = self._make_stage(
self.stage2_cfg, num_channels)
# stage 3
self.stage3_cfg = self.extra['stage3']
num_channels = self.stage3_cfg['num_channels']
block_type = self.stage3_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition2 = self._make_transition_layer(pre_stage_channels,
num_channels)
self.stage3, pre_stage_channels = self._make_stage(
self.stage3_cfg, num_channels)
# stage 4
deconv_cfg = {}
deconv_cfg["NUM_CHANNELS"] = [32, 64, 128, 256]
deconv_cfg["NUM_DECONVS"] = 1
deconv_cfg["KERNEL_SIZE"] = [3, 3, 3, 3]
deconv_cfg["NUM_BASIC_BLOCKS"] = 1
self.num_deconvs = len(pre_stage_channels) #deconv_cfg["NUM_DECONVS"]
self.deconv_layers = self._make_deconv_layers(deconv_cfg,
pre_stage_channels)
#self.upsample = nn.Upsample(scale_factor=2,mode='bilinear')
self.stage4_cfg = self.extra['stage4']
num_channels = self.stage4_cfg['num_channels']
block_type = self.stage4_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition3 = self._make_transition_layer(pre_stage_channels,
num_channels)
self.stage4, pre_stage_channels = self._make_stage(
self.stage4_cfg, num_channels)
#deconv_cfg["NUM_BASIC_BLOCKS"] = 1
self.num_deconvs = len(pre_stage_channels) #deconv_cfg["NUM_DECONVS"]
self.deconv_layers_2 = self._make_deconv_layers(
deconv_cfg, pre_stage_channels)
def __init__(self,
arch='efficientnet-b0',
out_indices=[4, 5, 6, 7, 8],
norm_cfg=dict(type="BN"),
norm_eval=True,
override_params=None):
super(EfficientNet, self).__init__()
self._check_model_name_is_valid(arch)
blocks_args, global_params = get_model_params(arch, override_params)
assert isinstance(blocks_args, list), 'blocks_args should be a list'
assert len(blocks_args) > 0, 'block args must be greater than 0'
self._global_params = global_params
self._blocks_args = blocks_args
self.out_indices = out_indices
self.norm_eval = norm_eval
# Get static or dynamic convolution depending on image size
Conv2d = get_same_padding_conv2d(image_size=global_params.image_size)
# Batch norm parameters
bn_mom = 1 - self._global_params.batch_norm_momentum
bn_eps = self._global_params.batch_norm_epsilon
# Stem
in_channels = 3 # rgb
out_channels = round_filters(32, self._global_params) # number of output channels
self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)
self._bn0 = build_norm_layer(norm_cfg, num_features=out_channels, postfix=0)[1]
# Build blocks
self._blocks = nn.ModuleList([])
self.per_last_stage_idx = []
cum_idx = 0
for block_args in self._blocks_args:
# Update block input and output filters based on depth multiplier.
block_args = block_args._replace(
input_filters=round_filters(block_args.input_filters, self._global_params),
output_filters=round_filters(block_args.output_filters, self._global_params),
num_repeat=round_repeats(block_args.num_repeat, self._global_params)
)
# The first block needs to take care of stride and filter size increase.
self._blocks.append(MBConvBlock(block_args, self._global_params, norm_cfg))
if block_args.num_repeat > 1:
block_args = block_args._replace(input_filters=block_args.output_filters, stride=1)
for _ in range(block_args.num_repeat - 1):
self._blocks.append(MBConvBlock(block_args, self._global_params, norm_cfg))
# last_stage_idx
cum_idx += block_args.num_repeat
self.per_last_stage_idx.append(cum_idx)
# out_indices_stage_idx
self.out_stage_idx = []
for stage in self.out_indices:
self.out_stage_idx.append(self.per_last_stage_idx[stage - 2] - 1)
# Head
# in_channels = block_args.output_filters # output of final block
# out_channels = round_filters(1280, self._global_params)
# self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
# self._bn1 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
# Final linear layer
# self._avg_pooling = nn.AdaptiveAvgPool2d(1)
# self._dropout = nn.Dropout(self._global_params.dropout_rate)
# self._fc = nn.Linear(out_channels, self._global_params.num_classes)
self._swish = MemoryEfficientSwish()
def make_res_layer(block,
inplanes,
planes,
blocks,
stride=1,
dilation=1,
groups=1,
base_width=4,
style='pytorch',
with_cp=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
dcn=None,
gcb=None,
sac=None,
rfp=None):
downsample = None
if stride != 1 or inplanes != planes * block.expansion:
downsample = nn.Sequential(
build_conv_layer(conv_cfg,
inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False),
build_norm_layer(norm_cfg, planes * block.expansion)[1],
)
layers = []
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=stride,
dilation=dilation,
downsample=downsample,
groups=groups,
base_width=base_width,
style=style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
sac=sac,
rfp=rfp))
inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=1,
dilation=dilation,
groups=groups,
base_width=base_width,
style=style,
with_cp=with_cp,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
dcn=dcn,
gcb=gcb,
sac=sac))
return nn.Sequential(*layers)
def __init__(self,
extra,
in_channels=3,
conv_cfg=None,
norm_cfg=dict(type='BN'),
norm_eval=True,
with_cp=False,
zero_init_residual=False):
super(HRNet, self).__init__()
self.extra = extra
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.zero_init_residual = zero_init_residual
# stem net
self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
self.norm2_name, norm2 = build_norm_layer(self.norm_cfg, 64, postfix=2)
self.conv1 = build_conv_layer(
self.conv_cfg,
in_channels,
64,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.add_module(self.norm1_name, norm1)
self.conv2 = build_conv_layer(
self.conv_cfg,
64,
64,
kernel_size=3,
stride=2,
padding=1,
bias=False)
self.add_module(self.norm2_name, norm2)
self.relu = nn.ReLU(inplace=True)
# stage 1
self.stage1_cfg = self.extra['stage1']
num_channels = self.stage1_cfg['num_channels'][0]
block_type = self.stage1_cfg['block']
num_blocks = self.stage1_cfg['num_blocks'][0]
block = self.blocks_dict[block_type]
stage1_out_channels = num_channels * block.expansion
self.layer1 = self._make_layer(block, 64, num_channels, num_blocks)
# stage 2
self.stage2_cfg = self.extra['stage2']
num_channels = self.stage2_cfg['num_channels']
block_type = self.stage2_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition1 = self._make_transition_layer([stage1_out_channels],
num_channels)
self.stage2, pre_stage_channels = self._make_stage(
self.stage2_cfg, num_channels)
# stage 3
self.stage3_cfg = self.extra['stage3']
num_channels = self.stage3_cfg['num_channels']
block_type = self.stage3_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition2 = self._make_transition_layer(pre_stage_channels,
num_channels)
self.stage3, pre_stage_channels = self._make_stage(
self.stage3_cfg, num_channels)
# stage 4
self.stage4_cfg = self.extra['stage4']
num_channels = self.stage4_cfg['num_channels']
block_type = self.stage4_cfg['block']
block = self.blocks_dict[block_type]
num_channels = [channel * block.expansion for channel in num_channels]
self.transition3 = self._make_transition_layer(pre_stage_channels,
num_channels)
self.stage4, pre_stage_channels = self._make_stage(
self.stage4_cfg, num_channels)