def __init__(self,
in_channels=[128],
tasks=None,
train_cfg=None,
test_cfg=None,
bbox_coder=None,
common_heads=dict(),
loss_cls=dict(type='GaussianFocalLoss', reduction='mean'),
loss_bbox=dict(type='L1Loss',
reduction='none',
loss_weight=0.25),
seperate_head=dict(type='SeparateHead',
init_bias=-2.19,
final_kernel=3),
feature_channel=64,
num_heatmap_convs=2,
conv_cfg=dict(type='Conv2d'),
norm_cfg=dict(type='BN2d'),
bias='auto',
norm_bbox=True,
pred_mode=['voxel'],
num_head_convs=2,
num_classes=17,
ignore_label=0,
grid_size=[480, 360, 32]):
super(SelectivesegHead, self).__init__()
self.class_names = [t['class_names'] for t in tasks]
self.train_cfg = train_cfg
self.test_cfg = test_cfg
self.in_channels = in_channels
self.num_classes = num_classes
self.norm_bbox = norm_bbox
self.loss_cls = build_loss(loss_cls)
self.loss_bbox = build_loss(loss_bbox)
self.bbox_coder = build_bbox_coder(bbox_coder)
self.fp16_enabled = False
self.pred_mode = pred_mode
assert ('voxel' in self.pred_mode) or ('point' in self.pred_mode) or (
'pillar' in self.pred_mode)
self.ignore_label = ignore_label
self.grid_size = grid_size
semantic_conv = []
for i in range(num_head_convs):
if i > 0:
in_channels = feature_channel
conv = ConvModule(in_channels,
feature_channel,
kernel_size=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
bias=False)
semantic_conv.append(conv)
self.semantic_conv = nn.Sequential(*semantic_conv)
self.semantic_cls = nn.Conv1d(feature_channel, self.num_classes, 1)
self.semantic_CE_loss = torch.nn.CrossEntropyLoss(
ignore_index=ignore_label)
class MobileNetV2(nn.Module):
"""MobileNetV2 backbone.
Args:
widen_factor (float): Width multiplier, multiply number of
channels in each layer by this amount. Default: 1.0.
strides (Sequence[int], optional): Strides of the first block of each
layer. If not specified, default config in ``arch_setting`` will
be used.
dilations (Sequence[int]): Dilation of each layer.
out_indices (None or Sequence[int]): Output from which stages.
Default: (7, ).
frozen_stages (int): Stages to be frozen (all param fixed).
Default: -1, which means not freezing any parameters.
conv_cfg (dict): Config dict for convolution layer.
Default: None, which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN').
act_cfg (dict): Config dict for activation layer.
Default: dict(type='ReLU6').
norm_eval (bool): Whether to set norm layers to eval mode, namely,
freeze running stats (mean and var). Note: Effect on Batch Norm
and its variants only. Default: False.
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
memory while slowing down the training speed. Default: False.
"""
# Parameters to build layers. 3 parameters are needed to construct a
# layer, from left to right: expand_ratio, channel, num_blocks.
arch_settings = [[1, 16, 1], [6, 24, 2], [6, 32, 3], [6, 64, 4],
[6, 96, 3], [6, 160, 3], [6, 320, 1]]
def __init__(self,
widen_factor=1.,
strides=(1, 2, 2, 2, 1, 2, 1),
dilations=(1, 1, 1, 1, 1, 1, 1),
out_indices=(1, 2, 4, 6),
frozen_stages=-1,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU6'),
norm_eval=False,
with_cp=False):
super(MobileNetV2, self).__init__()
self.widen_factor = widen_factor
self.strides = strides
self.dilations = dilations
assert len(strides) == len(dilations) == len(self.arch_settings)
self.out_indices = out_indices
for index in out_indices:
if index not in range(0, 7):
raise ValueError('the item in out_indices must in '
f'range(0, 7). But received {index}')
if frozen_stages not in range(-1, 7):
raise ValueError('frozen_stages must be in range(-1, 7). '
f'But received {frozen_stages}')
self.out_indices = out_indices
self.frozen_stages = frozen_stages
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.in_channels = make_divisible(32 * widen_factor, 8)
self.conv1 = ConvModule(in_channels=3,
out_channels=self.in_channels,
kernel_size=3,
stride=2,
padding=1,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
self.layers = []
for i, layer_cfg in enumerate(self.arch_settings):
expand_ratio, channel, num_blocks = layer_cfg
stride = self.strides[i]
dilation = self.dilations[i]
out_channels = make_divisible(channel * widen_factor, 8)
inverted_res_layer = self.make_layer(out_channels=out_channels,
num_blocks=num_blocks,
stride=stride,
dilation=dilation,
expand_ratio=expand_ratio)
layer_name = f'layer{i + 1}'
self.add_module(layer_name, inverted_res_layer)
self.layers.append(layer_name)
def make_layer(self, out_channels, num_blocks, stride, dilation,
expand_ratio):
"""Stack InvertedResidual blocks to build a layer for MobileNetV2.
Args:
out_channels (int): out_channels of block.
num_blocks (int): Number of blocks.
stride (int): Stride of the first block.
dilation (int): Dilation of the first block.
expand_ratio (int): Expand the number of channels of the
hidden layer in InvertedResidual by this ratio.
"""
layers = []
for i in range(num_blocks):
layers.append(
InvertedResidual(self.in_channels,
out_channels,
stride if i == 0 else 1,
expand_ratio=expand_ratio,
dilation=dilation if i == 0 else 1,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg,
with_cp=self.with_cp))
self.in_channels = out_channels
return nn.Sequential(*layers)
def init_weights(self, pretrained=None):
if isinstance(pretrained, str):
logger = logging.getLogger()
load_checkpoint(self, pretrained, strict=False, logger=logger)
elif pretrained is None:
for m in self.modules():
if isinstance(m, nn.Conv2d):
kaiming_init(m)
elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
constant_init(m, 1)
else:
raise TypeError('pretrained must be a str or None')
def forward(self, x):
x = self.conv1(x)
outs = []
for i, layer_name in enumerate(self.layers):
layer = getattr(self, layer_name)
x = layer(x)
if i in self.out_indices:
outs.append(x)
if len(outs) == 1:
return outs[0]
else:
return tuple(outs)
def _freeze_stages(self):
if self.frozen_stages >= 0:
for param in self.conv1.parameters():
param.requires_grad = False
for i in range(1, self.frozen_stages + 1):
layer = getattr(self, f'layer{i}')
layer.eval()
for param in layer.parameters():
param.requires_grad = False
def train(self, mode=True):
super(MobileNetV2, self).train(mode)
self._freeze_stages()
if mode and self.norm_eval:
for m in self.modules():
if isinstance(m, _BatchNorm):
m.eval()
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
add_extra_convs=False,
extra_convs_on_inputs=False,
relu_before_extra_convs=False,
no_norm_on_lateral=False,
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
upsample_cfg=dict(mode='nearest')):
super(FPN_involution, self).__init__()
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels)
self.num_outs = num_outs
self.relu_before_extra_convs = relu_before_extra_convs
self.no_norm_on_lateral = no_norm_on_lateral
self.fp16_enabled = False
self.upsample_cfg = upsample_cfg.copy()
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.add_extra_convs = add_extra_convs
assert isinstance(add_extra_convs, (str, bool))
if isinstance(add_extra_convs, str):
# Extra_convs_source choices: 'on_input', 'on_lateral', 'on_output'
assert add_extra_convs in ('on_input', 'on_lateral', 'on_output')
elif add_extra_convs: # True
if extra_convs_on_inputs:
# For compatibility with previous release
# TODO: deprecate `extra_convs_on_inputs`
self.add_extra_convs = 'on_input'
else:
self.add_extra_convs = 'on_output'
self.lateral_convs = nn.ModuleList()
self.fpn_convs = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(
in_channels[i],
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
fpn_conv = involution(out_channels, 7, 1)
#ConvModule(
#out_channels,
#out_channels,
#3,
#padding=1,
#conv_cfg=conv_cfg,
#norm_cfg=norm_cfg,
#act_cfg=act_cfg,
#inplace=False)
self.lateral_convs.append(l_conv)
self.fpn_convs.append(fpn_conv)
# add extra conv layers (e.g., RetinaNet)
extra_levels = num_outs - self.backbone_end_level + self.start_level
if self.add_extra_convs and extra_levels >= 1:
for i in range(extra_levels):
if i == 0 and self.add_extra_convs == 'on_input':
in_channels = self.in_channels[self.backbone_end_level - 1]
else:
in_channels = out_channels
extra_fpn_conv = ConvModule(in_channels,
out_channels,
3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.fpn_convs.append(extra_fpn_conv)
class X3D(nn.Module):
"""X3D backbone. https://arxiv.org/pdf/2004.04730.pdf.
Args:
gamma_w (float): Global channel width expansion factor. Default: 1.
gamma_b (float): Bottleneck channel width expansion factor. Default: 1.
gamma_d (float): Network depth expansion factor. Default: 1.
pretrained (str | None): Name of pretrained model. Default: None.
in_channels (int): Channel num of input features. Default: 3.
num_stages (int): Resnet stages. Default: 4.
spatial_strides (Sequence[int]):
Spatial strides of residual blocks of each stage.
Default: ``(1, 2, 2, 2)``.
frozen_stages (int): Stages to be frozen (all param fixed). If set to
-1, it means not freezing any parameters. Default: -1.
se_style (str): The style of inserting SE modules into BlockX3D, 'half'
denotes insert into half of the blocks, while 'all' denotes insert
into all blocks. Default: 'half'.
se_ratio (float | None): The reduction ratio of squeeze and excitation
unit. If set as None, it means not using SE unit. Default: 1 / 16.
use_swish (bool): Whether to use swish as the activation function
before and after the 3x3x3 conv. Default: True.
conv_cfg (dict): Config for conv layers. required keys are ``type``
Default: ``dict(type='Conv3d')``.
norm_cfg (dict): Config for norm layers. required keys are ``type`` and
``requires_grad``.
Default: ``dict(type='BN3d', requires_grad=True)``.
act_cfg (dict): Config dict for activation layer.
Default: ``dict(type='ReLU', inplace=True)``.
norm_eval (bool): Whether to set BN layers to eval mode, namely, freeze
running stats (mean and var). Default: False.
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
memory while slowing down the training speed. Default: False.
zero_init_residual (bool):
Whether to use zero initialization for residual block,
Default: True.
kwargs (dict, optional): Key arguments for "make_res_layer".
"""
def __init__(self,
gamma_w=1.0,
gamma_b=1.0,
gamma_d=1.0,
pretrained=None,
in_channels=3,
num_stages=4,
spatial_strides=(2, 2, 2, 2),
frozen_stages=-1,
se_style='half',
se_ratio=1 / 16,
use_swish=True,
conv_cfg=dict(type='Conv3d'),
norm_cfg=dict(type='BN3d', requires_grad=True),
act_cfg=dict(type='ReLU', inplace=True),
norm_eval=False,
with_cp=False,
zero_init_residual=True,
**kwargs):
super().__init__()
self.gamma_w = gamma_w
self.gamma_b = gamma_b
self.gamma_d = gamma_d
self.pretrained = pretrained
self.in_channels = in_channels
# Hard coded, can be changed by gamma_w
self.base_channels = 24
self.stage_blocks = [1, 2, 5, 3]
# apply parameters gamma_w and gamma_d
self.base_channels = self._round_width(self.base_channels,
self.gamma_w)
self.stage_blocks = [
self._round_repeats(x, self.gamma_d) for x in self.stage_blocks
]
self.num_stages = num_stages
assert 1 <= num_stages <= 4
self.spatial_strides = spatial_strides
assert len(spatial_strides) == num_stages
self.frozen_stages = frozen_stages
self.se_style = se_style
assert self.se_style in ['all', 'half']
self.se_ratio = se_ratio
assert (self.se_ratio is None) or (self.se_ratio > 0)
self.use_swish = use_swish
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.zero_init_residual = zero_init_residual
self.block = BlockX3D
self.stage_blocks = self.stage_blocks[:num_stages]
self.layer_inplanes = self.base_channels
self._make_stem_layer()
self.res_layers = []
for i, num_blocks in enumerate(self.stage_blocks):
spatial_stride = spatial_strides[i]
inplanes = self.base_channels * 2**i
planes = int(inplanes * self.gamma_b)
res_layer = self.make_res_layer(self.block,
self.layer_inplanes,
inplanes,
planes,
num_blocks,
spatial_stride=spatial_stride,
se_style=self.se_style,
se_ratio=self.se_ratio,
use_swish=self.use_swish,
norm_cfg=self.norm_cfg,
conv_cfg=self.conv_cfg,
act_cfg=self.act_cfg,
with_cp=with_cp,
**kwargs)
self.layer_inplanes = inplanes
layer_name = f'layer{i + 1}'
self.add_module(layer_name, res_layer)
self.res_layers.append(layer_name)
self.feat_dim = self.base_channels * 2**(len(self.stage_blocks) - 1)
self.conv5 = ConvModule(self.feat_dim,
int(self.feat_dim * self.gamma_b),
kernel_size=1,
stride=1,
padding=0,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
self.feat_dim = int(self.feat_dim * self.gamma_b)
@staticmethod
def _round_width(width, multiplier, min_depth=8, divisor=8):
"""Round width of filters based on width multiplier."""
if not multiplier:
return width
width *= multiplier
min_depth = min_depth or divisor
new_filters = max(min_depth,
int(width + divisor / 2) // divisor * divisor)
if new_filters < 0.9 * width:
new_filters += divisor
return int(new_filters)
@staticmethod
def _round_repeats(repeats, multiplier):
"""Round number of layers based on depth multiplier."""
if not multiplier:
return repeats
return int(math.ceil(multiplier * repeats))
# the module is parameterized with gamma_b
# no temporal_stride
def make_res_layer(self,
block,
layer_inplanes,
inplanes,
planes,
blocks,
spatial_stride=1,
se_style='half',
se_ratio=None,
use_swish=True,
norm_cfg=None,
act_cfg=None,
conv_cfg=None,
with_cp=False,
**kwargs):
"""Build residual layer for ResNet3D.
Args:
block (nn.Module): Residual module to be built.
layer_inplanes (int): Number of channels for the input feature
of the res layer.
inplanes (int): Number of channels for the input feature in each
block, which equals to base_channels * gamma_w.
planes (int): Number of channels for the output feature in each
block, which equals to base_channel * gamma_w * gamma_b.
blocks (int): Number of residual blocks.
spatial_stride (int): Spatial strides in residual and conv layers.
Default: 1.
se_style (str): The style of inserting SE modules into BlockX3D,
'half' denotes insert into half of the blocks, while 'all'
denotes insert into all blocks. Default: 'half'.
se_ratio (float | None): The reduction ratio of squeeze and
excitation unit. If set as None, it means not using SE unit.
Default: None.
use_swish (bool): Whether to use swish as the activation function
before and after the 3x3x3 conv. Default: True.
conv_cfg (dict | None): Config for norm layers. Default: None.
norm_cfg (dict | None): Config for norm layers. Default: None.
act_cfg (dict | None): Config for activate layers. Default: None.
with_cp (bool | None): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
Returns:
nn.Module: A residual layer for the given config.
"""
downsample = None
if spatial_stride != 1 or layer_inplanes != inplanes:
downsample = ConvModule(layer_inplanes,
inplanes,
kernel_size=1,
stride=(1, spatial_stride, spatial_stride),
padding=0,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
use_se = [False] * blocks
if self.se_style == 'all':
use_se = [True] * blocks
elif self.se_style == 'half':
use_se = [i % 2 == 0 for i in range(blocks)]
else:
raise NotImplementedError
layers = []
layers.append(
block(layer_inplanes,
planes,
inplanes,
spatial_stride=spatial_stride,
downsample=downsample,
se_ratio=se_ratio if use_se[0] else None,
use_swish=use_swish,
norm_cfg=norm_cfg,
conv_cfg=conv_cfg,
act_cfg=act_cfg,
with_cp=with_cp,
**kwargs))
for i in range(1, blocks):
layers.append(
block(inplanes,
planes,
inplanes,
spatial_stride=1,
se_ratio=se_ratio if use_se[i] else None,
use_swish=use_swish,
norm_cfg=norm_cfg,
conv_cfg=conv_cfg,
act_cfg=act_cfg,
with_cp=with_cp,
**kwargs))
return nn.Sequential(*layers)
def _make_stem_layer(self):
"""Construct the stem layers consists of a conv+norm+act module and a
pooling layer."""
self.conv1_s = ConvModule(self.in_channels,
self.base_channels,
kernel_size=(1, 3, 3),
stride=(1, 2, 2),
padding=(0, 1, 1),
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=None,
act_cfg=None)
self.conv1_t = ConvModule(self.base_channels,
self.base_channels,
kernel_size=(5, 1, 1),
stride=(1, 1, 1),
padding=(2, 0, 0),
groups=self.base_channels,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
def _freeze_stages(self):
"""Prevent all the parameters from being optimized before
``self.frozen_stages``."""
if self.frozen_stages >= 0:
self.conv1_s.eval()
self.conv1_t.eval()
for param in self.conv1_s.parameters():
param.requires_grad = False
for param in self.conv1_t.parameters():
param.requires_grad = False
for i in range(1, self.frozen_stages + 1):
m = getattr(self, f'layer{i}')
m.eval()
for param in m.parameters():
param.requires_grad = False
def init_weights(self):
"""Initiate the parameters either from existing checkpoint or from
scratch."""
if isinstance(self.pretrained, str):
logger = get_root_logger()
logger.info(f'load model from: {self.pretrained}')
load_checkpoint(self, self.pretrained, strict=False, logger=logger)
elif self.pretrained is None:
for m in self.modules():
if isinstance(m, nn.Conv3d):
kaiming_init(m)
elif isinstance(m, _BatchNorm):
constant_init(m, 1)
if self.zero_init_residual:
for m in self.modules():
if isinstance(m, BlockX3D):
constant_init(m.conv3.bn, 0)
else:
raise TypeError('pretrained must be a str or None')
def forward(self, x):
"""Defines the computation performed at every call.
Args:
x (torch.Tensor): The input data.
Returns:
torch.Tensor: The feature of the input
samples extracted by the backbone.
"""
x = self.conv1_s(x)
x = self.conv1_t(x)
for layer_name in self.res_layers:
res_layer = getattr(self, layer_name)
x = res_layer(x)
x = self.conv5(x)
return x
def train(self, mode=True):
"""Set the optimization status when training."""
super().train(mode)
self._freeze_stages()
if mode and self.norm_eval:
for m in self.modules():
if isinstance(m, _BatchNorm):
m.eval()
def make_res_layer(self,
block,
layer_inplanes,
inplanes,
planes,
blocks,
spatial_stride=1,
se_style='half',
se_ratio=None,
use_swish=True,
norm_cfg=None,
act_cfg=None,
conv_cfg=None,
with_cp=False,
**kwargs):
"""Build residual layer for ResNet3D.
Args:
block (nn.Module): Residual module to be built.
layer_inplanes (int): Number of channels for the input feature
of the res layer.
inplanes (int): Number of channels for the input feature in each
block, which equals to base_channels * gamma_w.
planes (int): Number of channels for the output feature in each
block, which equals to base_channel * gamma_w * gamma_b.
blocks (int): Number of residual blocks.
spatial_stride (int): Spatial strides in residual and conv layers.
Default: 1.
se_style (str): The style of inserting SE modules into BlockX3D,
'half' denotes insert into half of the blocks, while 'all'
denotes insert into all blocks. Default: 'half'.
se_ratio (float | None): The reduction ratio of squeeze and
excitation unit. If set as None, it means not using SE unit.
Default: None.
use_swish (bool): Whether to use swish as the activation function
before and after the 3x3x3 conv. Default: True.
conv_cfg (dict | None): Config for norm layers. Default: None.
norm_cfg (dict | None): Config for norm layers. Default: None.
act_cfg (dict | None): Config for activate layers. Default: None.
with_cp (bool | None): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
Returns:
nn.Module: A residual layer for the given config.
"""
downsample = None
if spatial_stride != 1 or layer_inplanes != inplanes:
downsample = ConvModule(layer_inplanes,
inplanes,
kernel_size=1,
stride=(1, spatial_stride, spatial_stride),
padding=0,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
use_se = [False] * blocks
if self.se_style == 'all':
use_se = [True] * blocks
elif self.se_style == 'half':
use_se = [i % 2 == 0 for i in range(blocks)]
else:
raise NotImplementedError
layers = []
layers.append(
block(layer_inplanes,
planes,
inplanes,
spatial_stride=spatial_stride,
downsample=downsample,
se_ratio=se_ratio if use_se[0] else None,
use_swish=use_swish,
norm_cfg=norm_cfg,
conv_cfg=conv_cfg,
act_cfg=act_cfg,
with_cp=with_cp,
**kwargs))
for i in range(1, blocks):
layers.append(
block(inplanes,
planes,
inplanes,
spatial_stride=1,
se_ratio=se_ratio if use_se[i] else None,
use_swish=use_swish,
norm_cfg=norm_cfg,
conv_cfg=conv_cfg,
act_cfg=act_cfg,
with_cp=with_cp,
**kwargs))
return nn.Sequential(*layers)
def make_res_layer(block,
inplanes,
planes,
blocks,
stride=1,
dilation=1,
style='pytorch',
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
with_cp=False):
"""Build residual layer for ResNet.
Args:
block: (nn.Module): Residual module to be built.
inplanes (int): Number of channels for the input feature in each block.
planes (int): Number of channels for the output feature in each block.
blocks (int): Number of residual blocks.
stride (int): Stride in the conv layer. Default: 1.
dilation (int): Spacing between kernel elements. Default: 1.
style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
layer is the 3x3 conv layer, otherwise the stride-two layer is
the first 1x1 conv layer. Default: 'pytorch'.
conv_cfg (dict | None): Config for norm layers. Default: None.
norm_cfg (dict | None): Config for norm layers. Default: None.
act_cfg (dict | None): Config for activate layers. Default: None.
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
memory while slowing down the training speed. Default: False.
Returns:
nn.Module: A residual layer for the given config.
"""
downsample = None
if stride != 1 or inplanes != planes * block.expansion:
downsample = ConvModule(
inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
layers = []
layers.append(
block(
inplanes,
planes,
stride,
dilation,
downsample,
style=style,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
with_cp=with_cp))
inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(
block(
inplanes,
planes,
1,
dilation,
style=style,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
with_cp=with_cp))
return nn.Sequential(*layers)
def __init__(self,
img_channels,
pts_channels,
mid_channels,
out_channels,
img_levels=3,
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
activate_out=True,
fuse_out=False,
dropout_ratio=0,
aligned=True,
align_corners=True,
padding_mode='zeros',
lateral_conv=True):
super(PointFusion, self).__init__()
if isinstance(img_levels, int):
img_levels = [img_levels]
if isinstance(img_channels, int):
img_channels = [img_channels] * len(img_levels)
assert isinstance(img_levels, list)
assert isinstance(img_channels, list)
assert len(img_channels) == len(img_levels)
self.img_levels = img_levels
self.act_cfg = act_cfg
self.activate_out = activate_out
self.fuse_out = fuse_out
self.dropout_ratio = dropout_ratio
self.img_channels = img_channels
self.aligned = aligned
self.align_corners = align_corners
self.padding_mode = padding_mode
self.lateral_convs = None
if lateral_conv:
self.lateral_convs = nn.ModuleList()
for i in range(len(img_channels)):
l_conv = ConvModule(img_channels[i],
mid_channels,
3,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=self.act_cfg,
inplace=False)
self.lateral_convs.append(l_conv)
self.img_transform = nn.Sequential(
nn.Linear(mid_channels * len(img_channels), out_channels),
nn.BatchNorm1d(out_channels, eps=1e-3, momentum=0.01),
)
else:
self.img_transform = nn.Sequential(
nn.Linear(sum(img_channels), out_channels),
nn.BatchNorm1d(out_channels, eps=1e-3, momentum=0.01),
)
self.pts_transform = nn.Sequential(
nn.Linear(pts_channels, out_channels),
nn.BatchNorm1d(out_channels, eps=1e-3, momentum=0.01),
)
if self.fuse_out:
self.fuse_conv = nn.Sequential(
nn.Linear(mid_channels, out_channels),
# For pts the BN is initialized differently by default
# TODO: check whether this is necessary
nn.BatchNorm1d(out_channels, eps=1e-3, momentum=0.01),
nn.ReLU(inplace=False))
self.init_weights()
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
add_extra_convs=False,
extra_convs_on_inputs=True,
relu_before_extra_convs=False,
no_norm_on_lateral=False,
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
upsample_cfg=dict(mode='nearest')):
super(DyFPN_B_CNNGate, self).__init__(in_channels, out_channels,
num_outs)
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels)
self.num_outs = num_outs
self.relu_before_extra_convs = relu_before_extra_convs
self.no_norm_on_lateral = no_norm_on_lateral
self.fp16_enabled = False
self.upsample_cfg = upsample_cfg.copy()
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.add_extra_convs = add_extra_convs
self.gate_0 = Attention_SEblock(channels=in_channels[0],
reduction=4,
temperature=1)
self.gate_1 = Attention_SEblock(channels=in_channels[1],
reduction=4,
temperature=1)
self.gate_2 = Attention_SEblock(channels=in_channels[2],
reduction=4,
temperature=1)
self.gate_3 = Attention_SEblock(channels=in_channels[3],
reduction=4,
temperature=1)
assert isinstance(add_extra_convs, (str, bool))
if isinstance(add_extra_convs, str):
# Extra_convs_source choices: 'on_input', 'on_lateral', 'on_output'
assert add_extra_convs in ('on_input', 'on_lateral', 'on_output')
elif add_extra_convs: # True
if extra_convs_on_inputs:
# For compatibility with previous release
# TODO: deprecate `extra_convs_on_inputs`
self.add_extra_convs = 'on_input'
else:
self.add_extra_convs = 'on_output'
self.lateral_convs = nn.ModuleList()
self.fpn_convs = nn.ModuleList()
self.skip_lateral_convs = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
l_conv_11 = ConvModule(
in_channels[i],
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
l_conv_33 = ConvModule(
in_channels[i],
out_channels,
3,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
l_conv_33_dilation_2 = ConvModule(
in_channels[i],
out_channels,
3,
padding=2,
dilation=2,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
l_conv_33_dilation_3 = ConvModule(
in_channels[i],
out_channels,
3,
padding=3,
dilation=3,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
l_conv_55 = ConvModule(
in_channels[i],
out_channels,
5,
padding=2,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
l_conv_55_dilation_2 = ConvModule(
in_channels[i],
out_channels,
5,
padding=4,
dilation=2,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
l_conv_55_dilation_3 = ConvModule(
in_channels[i],
out_channels,
5,
padding=6,
dilation=3,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
skip_l_conv_11 = ConvModule(
in_channels[i],
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
self.lateral_convs.append(l_conv_11)
self.lateral_convs.append(l_conv_33)
self.lateral_convs.append(l_conv_33_dilation_2)
self.lateral_convs.append(l_conv_33_dilation_3)
self.lateral_convs.append(l_conv_55)
self.lateral_convs.append(l_conv_55_dilation_2)
self.lateral_convs.append(l_conv_55_dilation_3)
self.skip_lateral_convs.append(skip_l_conv_11)
fpn_conv = ConvModule(out_channels,
out_channels,
3,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.fpn_convs.append(fpn_conv)
# add extra conv layers (e.g., RetinaNet)
extra_levels = num_outs - self.backbone_end_level + self.start_level
if self.add_extra_convs and extra_levels >= 1:
for i in range(extra_levels):
if i == 0 and self.add_extra_convs == 'on_input':
in_channels = self.in_channels[self.backbone_end_level - 1]
else:
in_channels = out_channels
extra_fpn_conv = ConvModule(in_channels,
out_channels,
3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.fpn_convs.append(extra_fpn_conv)
def __init__(self,
num_dims,
num_classes,
primitive_mode,
train_cfg=None,
test_cfg=None,
vote_moudule_cfg=None,
vote_aggregation_cfg=None,
feat_channels=(128, 128),
upper_thresh=100.0,
surface_thresh=0.5,
conv_cfg=dict(type='Conv1d'),
norm_cfg=dict(type='BN1d'),
objectness_loss=None,
center_loss=None,
semantic_reg_loss=None,
semantic_cls_loss=None):
super(PrimitiveHead, self).__init__()
assert primitive_mode in ['z', 'xy', 'line']
# The dimension of primitive semantic information.
self.num_dims = num_dims
self.num_classes = num_classes
self.primitive_mode = primitive_mode
self.train_cfg = train_cfg
self.test_cfg = test_cfg
self.gt_per_seed = vote_moudule_cfg['gt_per_seed']
self.num_proposal = vote_aggregation_cfg['num_point']
self.upper_thresh = upper_thresh
self.surface_thresh = surface_thresh
self.objectness_loss = build_loss(objectness_loss)
self.center_loss = build_loss(center_loss)
self.semantic_reg_loss = build_loss(semantic_reg_loss)
self.semantic_cls_loss = build_loss(semantic_cls_loss)
assert vote_aggregation_cfg['mlp_channels'][0] == vote_moudule_cfg[
'in_channels']
# Primitive existence flag prediction
self.flag_conv = ConvModule(vote_moudule_cfg['conv_channels'][-1],
vote_moudule_cfg['conv_channels'][-1] // 2,
1,
padding=0,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
bias=True,
inplace=True)
self.flag_pred = torch.nn.Conv1d(
vote_moudule_cfg['conv_channels'][-1] // 2, 2, 1)
self.vote_module = VoteModule(**vote_moudule_cfg)
self.vote_aggregation = PointSAModule(**vote_aggregation_cfg)
prev_channel = vote_aggregation_cfg['mlp_channels'][-1]
conv_pred_list = list()
for k in range(len(feat_channels)):
conv_pred_list.append(
ConvModule(prev_channel,
feat_channels[k],
1,
padding=0,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
bias=True,
inplace=True))
prev_channel = feat_channels[k]
self.conv_pred = nn.Sequential(*conv_pred_list)
conv_out_channel = 3 + num_dims + num_classes
self.conv_pred.add_module('conv_out',
nn.Conv1d(prev_channel, conv_out_channel, 1))
def __init__(self,
in_channels,
stem_channels=1024,
num_blocks=2,
kernel_sizes=(3, 3, 3),
dropout=0.25,
causal=False,
residual=True,
use_stride_conv=False,
conv_cfg=dict(type='Conv1d'),
norm_cfg=dict(type='BN1d'),
max_norm=None):
# Protect mutable default arguments
conv_cfg = copy.deepcopy(conv_cfg)
norm_cfg = copy.deepcopy(norm_cfg)
super().__init__()
self.in_channels = in_channels
self.stem_channels = stem_channels
self.num_blocks = num_blocks
self.kernel_sizes = kernel_sizes
self.dropout = dropout
self.causal = causal
self.residual = residual
self.use_stride_conv = use_stride_conv
self.max_norm = max_norm
assert num_blocks == len(kernel_sizes) - 1
for ks in kernel_sizes:
assert ks % 2 == 1, 'Only odd filter widths are supported.'
self.expand_conv = ConvModule(
in_channels,
stem_channels,
kernel_size=kernel_sizes[0],
stride=kernel_sizes[0] if use_stride_conv else 1,
bias='auto',
conv_cfg=conv_cfg,
norm_cfg=norm_cfg)
dilation = kernel_sizes[0]
self.tcn_blocks = nn.ModuleList()
for i in range(1, num_blocks + 1):
self.tcn_blocks.append(
BasicTemporalBlock(in_channels=stem_channels,
out_channels=stem_channels,
mid_channels=stem_channels,
kernel_size=kernel_sizes[i],
dilation=dilation,
dropout=dropout,
causal=causal,
residual=residual,
use_stride_conv=use_stride_conv,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg))
dilation *= kernel_sizes[i]
if self.max_norm is not None:
# Apply weight norm clip to conv layers
weight_clip = WeightNormClipHook(self.max_norm)
for module in self.modules():
if isinstance(module, nn.modules.conv._ConvNd):
weight_clip.register(module)
self.dropout = nn.Dropout(dropout) if dropout > 0 else None
def __init__(self,
num_classes,
seg_in_channels,
part_in_channels,
seg_conv_channels=None,
part_conv_channels=None,
merge_conv_channels=None,
down_conv_channels=None,
shared_fc_channels=None,
cls_channels=None,
reg_channels=None,
dropout_ratio=0.1,
roi_feat_size=14,
with_corner_loss=True,
bbox_coder=dict(type='DeltaXYZWLHRBBoxCoder'),
conv_cfg=dict(type='Conv1d'),
norm_cfg=dict(type='BN1d', eps=1e-3, momentum=0.01),
loss_bbox=dict(type='SmoothL1Loss',
beta=1.0 / 9.0,
loss_weight=2.0),
loss_cls=dict(type='CrossEntropyLoss',
use_sigmoid=True,
reduction='none',
loss_weight=1.0),
init_cfg=None):
super(PartA2BboxHead, self).__init__(init_cfg=init_cfg)
self.num_classes = num_classes
self.with_corner_loss = with_corner_loss
self.bbox_coder = build_bbox_coder(bbox_coder)
self.loss_bbox = build_loss(loss_bbox)
self.loss_cls = build_loss(loss_cls)
self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
assert down_conv_channels[-1] == shared_fc_channels[0]
# init layers
part_channel_last = part_in_channels
part_conv = []
for i, channel in enumerate(part_conv_channels):
part_conv.append(
make_sparse_convmodule(part_channel_last,
channel,
3,
padding=1,
norm_cfg=norm_cfg,
indice_key=f'rcnn_part{i}',
conv_type='SubMConv3d'))
part_channel_last = channel
self.part_conv = spconv.SparseSequential(*part_conv)
seg_channel_last = seg_in_channels
seg_conv = []
for i, channel in enumerate(seg_conv_channels):
seg_conv.append(
make_sparse_convmodule(seg_channel_last,
channel,
3,
padding=1,
norm_cfg=norm_cfg,
indice_key=f'rcnn_seg{i}',
conv_type='SubMConv3d'))
seg_channel_last = channel
self.seg_conv = spconv.SparseSequential(*seg_conv)
self.conv_down = spconv.SparseSequential()
merge_conv_channel_last = part_channel_last + seg_channel_last
merge_conv = []
for i, channel in enumerate(merge_conv_channels):
merge_conv.append(
make_sparse_convmodule(merge_conv_channel_last,
channel,
3,
padding=1,
norm_cfg=norm_cfg,
indice_key='rcnn_down0'))
merge_conv_channel_last = channel
down_conv_channel_last = merge_conv_channel_last
conv_down = []
for i, channel in enumerate(down_conv_channels):
conv_down.append(
make_sparse_convmodule(down_conv_channel_last,
channel,
3,
padding=1,
norm_cfg=norm_cfg,
indice_key='rcnn_down1'))
down_conv_channel_last = channel
self.conv_down.add_module('merge_conv',
spconv.SparseSequential(*merge_conv))
self.conv_down.add_module(
'max_pool3d', spconv.SparseMaxPool3d(kernel_size=2, stride=2))
self.conv_down.add_module('down_conv',
spconv.SparseSequential(*conv_down))
shared_fc_list = []
pool_size = roi_feat_size // 2
pre_channel = shared_fc_channels[0] * pool_size**3
for k in range(1, len(shared_fc_channels)):
shared_fc_list.append(
ConvModule(pre_channel,
shared_fc_channels[k],
1,
padding=0,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
inplace=True))
pre_channel = shared_fc_channels[k]
if k != len(shared_fc_channels) - 1 and dropout_ratio > 0:
shared_fc_list.append(nn.Dropout(dropout_ratio))
self.shared_fc = nn.Sequential(*shared_fc_list)
# Classification layer
channel_in = shared_fc_channels[-1]
cls_channel = 1
cls_layers = []
pre_channel = channel_in
for k in range(0, len(cls_channels)):
cls_layers.append(
ConvModule(pre_channel,
cls_channels[k],
1,
padding=0,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
inplace=True))
pre_channel = cls_channels[k]
cls_layers.append(
ConvModule(pre_channel,
cls_channel,
1,
padding=0,
conv_cfg=conv_cfg,
act_cfg=None))
if dropout_ratio >= 0:
cls_layers.insert(1, nn.Dropout(dropout_ratio))
self.conv_cls = nn.Sequential(*cls_layers)
# Regression layer
reg_layers = []
pre_channel = channel_in
for k in range(0, len(reg_channels)):
reg_layers.append(
ConvModule(pre_channel,
reg_channels[k],
1,
padding=0,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
inplace=True))
pre_channel = reg_channels[k]
reg_layers.append(
ConvModule(pre_channel,
self.bbox_coder.code_size,
1,
padding=0,
conv_cfg=conv_cfg,
act_cfg=None))
if dropout_ratio >= 0:
reg_layers.insert(1, nn.Dropout(dropout_ratio))
self.conv_reg = nn.Sequential(*reg_layers)
if init_cfg is None:
self.init_cfg = dict(type='Xavier',
layer=['Conv2d', 'Conv1d'],
distribution='uniform')
def __init__(self,
out_channels=128,
align_corners=False,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
init_cfg=None):
super(BGALayer, self).__init__(init_cfg=init_cfg)
self.out_channels = out_channels
self.align_corners = align_corners
self.detail_dwconv = nn.Sequential(
DepthwiseSeparableConvModule(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
dw_norm_cfg=norm_cfg,
dw_act_cfg=None,
pw_norm_cfg=None,
pw_act_cfg=None,
))
self.detail_down = nn.Sequential(
ConvModule(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
stride=2,
padding=1,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None),
nn.AvgPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=False))
self.semantic_conv = nn.Sequential(
ConvModule(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None))
self.semantic_dwconv = nn.Sequential(
DepthwiseSeparableConvModule(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
dw_norm_cfg=norm_cfg,
dw_act_cfg=None,
pw_norm_cfg=None,
pw_act_cfg=None,
))
self.conv = ConvModule(
in_channels=self.out_channels,
out_channels=self.out_channels,
kernel_size=3,
stride=1,
padding=1,
inplace=True,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
)
def __init__(self,
detail_channels=(64, 64, 128),
in_channels=3,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
init_cfg=None):
super(DetailBranch, self).__init__(init_cfg=init_cfg)
detail_branch = []
for i in range(len(detail_channels)):
if i == 0:
detail_branch.append(
nn.Sequential(
ConvModule(
in_channels=in_channels,
out_channels=detail_channels[i],
kernel_size=3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg),
ConvModule(
in_channels=detail_channels[i],
out_channels=detail_channels[i],
kernel_size=3,
stride=1,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)))
else:
detail_branch.append(
nn.Sequential(
ConvModule(
in_channels=detail_channels[i - 1],
out_channels=detail_channels[i],
kernel_size=3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg),
ConvModule(
in_channels=detail_channels[i],
out_channels=detail_channels[i],
kernel_size=3,
stride=1,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg),
ConvModule(
in_channels=detail_channels[i],
out_channels=detail_channels[i],
kernel_size=3,
stride=1,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)))
self.detail_branch = nn.ModuleList(detail_branch)
def __init__(self,
in_channels,
out_channels,
exp_ratio=6,
stride=1,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
init_cfg=None):
super(GELayer, self).__init__(init_cfg=init_cfg)
mid_channel = in_channels * exp_ratio
self.conv1 = ConvModule(
in_channels=in_channels,
out_channels=in_channels,
kernel_size=3,
stride=1,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
if stride == 1:
self.dwconv = nn.Sequential(
# ReLU in ConvModule not shown in paper
ConvModule(
in_channels=in_channels,
out_channels=mid_channel,
kernel_size=3,
stride=stride,
padding=1,
groups=in_channels,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg))
self.shortcut = None
else:
self.dwconv = nn.Sequential(
ConvModule(
in_channels=in_channels,
out_channels=mid_channel,
kernel_size=3,
stride=stride,
padding=1,
groups=in_channels,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None),
# ReLU in ConvModule not shown in paper
ConvModule(
in_channels=mid_channel,
out_channels=mid_channel,
kernel_size=3,
stride=1,
padding=1,
groups=mid_channel,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg),
)
self.shortcut = nn.Sequential(
DepthwiseSeparableConvModule(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=3,
stride=stride,
padding=1,
dw_norm_cfg=norm_cfg,
dw_act_cfg=None,
pw_norm_cfg=norm_cfg,
pw_act_cfg=None,
))
self.conv2 = nn.Sequential(
ConvModule(
in_channels=mid_channel,
out_channels=out_channels,
kernel_size=1,
stride=1,
padding=0,
bias=False,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None,
))
self.act = build_activation_layer(act_cfg)
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
norm_cfg=None,
act_cfg=None,
order=('conv', 'norm', 'act'),
upsample_cfg=dict(type='carafe',
up_kernel=5,
up_group=1,
encoder_kernel=3,
encoder_dilation=1)):
super(FPN_CARAFE_3_3_pow_norm, self).__init__()
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels)
self.num_outs = num_outs
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.with_bias = norm_cfg is None
self.upsample_cfg = upsample_cfg.copy()
self.upsample = self.upsample_cfg.get('type')
self.relu = nn.ReLU(inplace=False)
self.order = order
assert order in [('conv', 'norm', 'act'), ('act', 'conv', 'norm')]
assert self.upsample in [
'nearest', 'bilinear', 'deconv', 'pixel_shuffle', 'carafe', None
]
if self.upsample in ['deconv', 'pixel_shuffle']:
assert hasattr(
self.upsample_cfg,
'upsample_kernel') and self.upsample_cfg.upsample_kernel > 0
self.upsample_kernel = self.upsample_cfg.pop('upsample_kernel')
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.lateral_convs = nn.ModuleList()
self.fpn_convs = nn.ModuleList()
self.upsample_modules = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(in_channels[i],
out_channels,
1,
norm_cfg=norm_cfg,
bias=self.with_bias,
act_cfg=act_cfg,
inplace=False,
order=self.order)
fpn_conv = ConvModule(out_channels,
out_channels,
3,
padding=1,
norm_cfg=self.norm_cfg,
bias=self.with_bias,
act_cfg=act_cfg,
inplace=False,
order=self.order)
if i != self.backbone_end_level - 1:
upsample_cfg_ = self.upsample_cfg.copy()
if self.upsample == 'deconv':
upsample_cfg_.update(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=self.upsample_kernel,
stride=2,
padding=(self.upsample_kernel - 1) // 2,
output_padding=(self.upsample_kernel - 1) // 2)
elif self.upsample == 'pixel_shuffle':
upsample_cfg_.update(in_channels=out_channels,
out_channels=out_channels,
scale_factor=2,
upsample_kernel=self.upsample_kernel)
elif self.upsample == 'carafe':
upsample_cfg_.update(channels=out_channels, scale_factor=2)
else:
# suppress warnings
align_corners = (None
if self.upsample == 'nearest' else False)
upsample_cfg_.update(scale_factor=2,
mode=self.upsample,
align_corners=align_corners)
upsample_module = build_upsample_layer(upsample_cfg_)
self.upsample_modules.append(upsample_module)
self.lateral_convs.append(l_conv)
self.fpn_convs.append(fpn_conv)
# add extra conv layers (e.g., RetinaNet)
extra_out_levels = (num_outs - self.backbone_end_level +
self.start_level)
if extra_out_levels >= 1:
for i in range(extra_out_levels):
in_channels = (self.in_channels[self.backbone_end_level -
1] if i == 0 else out_channels)
extra_l_conv = ConvModule(in_channels,
out_channels,
3,
stride=2,
padding=1,
norm_cfg=norm_cfg,
bias=self.with_bias,
act_cfg=act_cfg,
inplace=False,
order=self.order)
if self.upsample == 'deconv':
upsampler_cfg_ = dict(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=self.upsample_kernel,
stride=2,
padding=(self.upsample_kernel - 1) // 2,
output_padding=(self.upsample_kernel - 1) // 2)
elif self.upsample == 'pixel_shuffle':
upsampler_cfg_ = dict(in_channels=out_channels,
out_channels=out_channels,
scale_factor=2,
upsample_kernel=self.upsample_kernel)
elif self.upsample == 'carafe':
upsampler_cfg_ = dict(channels=out_channels,
scale_factor=2,
**self.upsample_cfg)
else:
# suppress warnings
align_corners = (None
if self.upsample == 'nearest' else False)
upsampler_cfg_ = dict(scale_factor=2,
mode=self.upsample,
align_corners=align_corners)
upsampler_cfg_['type'] = self.upsample
upsample_module = build_upsample_layer(upsampler_cfg_)
extra_fpn_conv = ConvModule(out_channels,
out_channels,
3,
padding=1,
norm_cfg=self.norm_cfg,
bias=self.with_bias,
act_cfg=act_cfg,
inplace=False,
order=self.order)
self.upsample_modules.append(upsample_module)
self.fpn_convs.append(extra_fpn_conv)
self.lateral_convs.append(extra_l_conv)
def __init__(self,
ind,
num_units,
in_channels,
unit_channels=256,
gen_skip=False,
gen_cross_conv=False,
norm_cfg=dict(type='BN'),
out_channels=64):
# Protect mutable default arguments
norm_cfg = cp.deepcopy(norm_cfg)
super().__init__()
self.num_units = num_units
self.norm_cfg = norm_cfg
self.in_skip = ConvModule(
in_channels,
unit_channels,
kernel_size=1,
stride=1,
padding=0,
norm_cfg=self.norm_cfg,
act_cfg=None,
inplace=True)
self.relu = nn.ReLU(inplace=True)
self.ind = ind
if self.ind > 0:
self.up_conv = ConvModule(
unit_channels,
unit_channels,
kernel_size=1,
stride=1,
padding=0,
norm_cfg=self.norm_cfg,
act_cfg=None,
inplace=True)
self.gen_skip = gen_skip
if self.gen_skip:
self.out_skip1 = ConvModule(
in_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0,
norm_cfg=self.norm_cfg,
inplace=True)
self.out_skip2 = ConvModule(
unit_channels,
in_channels,
kernel_size=1,
stride=1,
padding=0,
norm_cfg=self.norm_cfg,
inplace=True)
self.gen_cross_conv = gen_cross_conv
if self.ind == num_units - 1 and self.gen_cross_conv:
self.cross_conv = ConvModule(
unit_channels,
out_channels,
kernel_size=1,
stride=1,
padding=0,
norm_cfg=self.norm_cfg,
inplace=True)
def __init__(self,
in_channels,
out_channels,
num_outs,
stack_times,
start_level=0,
end_level=-1,
add_extra_convs=False,
norm_cfg=None):
super(NASFPN, self).__init__()
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels) # num of input feature levels
self.num_outs = num_outs # num of output feature levels
self.stack_times = stack_times
self.norm_cfg = norm_cfg
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.add_extra_convs = add_extra_convs
# add lateral connections
self.lateral_convs = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(in_channels[i],
out_channels,
1,
norm_cfg=norm_cfg,
act_cfg=None)
self.lateral_convs.append(l_conv)
# add extra downsample layers (stride-2 pooling or conv)
extra_levels = num_outs - self.backbone_end_level + self.start_level
self.extra_downsamples = nn.ModuleList()
for i in range(extra_levels):
extra_conv = ConvModule(out_channels,
out_channels,
1,
norm_cfg=norm_cfg,
act_cfg=None)
self.extra_downsamples.append(
nn.Sequential(extra_conv, nn.MaxPool2d(2, 2)))
# add NAS FPN connections
self.fpn_stages = nn.ModuleList()
for _ in range(self.stack_times):
stage = nn.ModuleDict()
# gp(p6, p4) -> p4_1
stage['gp_64_4'] = GlobalPoolingCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p4_1, p4) -> p4_2
stage['sum_44_4'] = SumCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p4_2, p3) -> p3_out
stage['sum_43_3'] = SumCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p3_out, p4_2) -> p4_out
stage['sum_34_4'] = SumCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p5, gp(p4_out, p3_out)) -> p5_out
stage['gp_43_5'] = GlobalPoolingCell(with_out_conv=False)
stage['sum_55_5'] = SumCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# sum(p7, gp(p5_out, p4_2)) -> p7_out
stage['gp_54_7'] = GlobalPoolingCell(with_out_conv=False)
stage['sum_77_7'] = SumCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
# gp(p7_out, p5_out) -> p6_out
stage['gp_75_6'] = GlobalPoolingCell(in_channels=out_channels,
out_channels=out_channels,
out_norm_cfg=norm_cfg)
self.fpn_stages.append(stage)
def __init__(self,
st_feat_channels,
lt_feat_channels,
latent_channels,
num_st_feat,
num_lt_feat,
use_scale=True,
pre_activate=True,
pre_activate_with_ln=True,
conv_cfg=None,
norm_cfg=None,
dropout_ratio=0.2,
zero_init_out_conv=False):
super().__init__()
if conv_cfg is None:
conv_cfg = dict(type='Conv3d')
self.st_feat_channels = st_feat_channels
self.lt_feat_channels = lt_feat_channels
self.latent_channels = latent_channels
self.num_st_feat = num_st_feat
self.num_lt_feat = num_lt_feat
self.use_scale = use_scale
self.pre_activate = pre_activate
self.pre_activate_with_ln = pre_activate_with_ln
self.dropout_ratio = dropout_ratio
self.zero_init_out_conv = zero_init_out_conv
self.st_feat_conv = ConvModule(self.st_feat_channels,
self.latent_channels,
kernel_size=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
self.lt_feat_conv = ConvModule(self.lt_feat_channels,
self.latent_channels,
kernel_size=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
self.global_conv = ConvModule(self.lt_feat_channels,
self.latent_channels,
kernel_size=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
if pre_activate:
self.ln = nn.LayerNorm([latent_channels, num_st_feat, 1, 1])
else:
self.ln = nn.LayerNorm([st_feat_channels, num_st_feat, 1, 1])
self.relu = nn.ReLU()
self.out_conv = ConvModule(self.latent_channels,
self.st_feat_channels,
kernel_size=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
if self.dropout_ratio > 0:
self.dropout = nn.Dropout(self.dropout_ratio)
class ResNet(nn.Module):
"""ResNet backbone.
Args:
depth (int): Depth of resnet, from {18, 34, 50, 101, 152}.
pretrained (str | None): Name of pretrained model. Default: None.
in_channels (int): Channel num of input features. Default: 3.
num_stages (int): Resnet stages. Default: 4.
strides (Sequence[int]): Strides of the first block of each stage.
out_indices (Sequence[int]): Indices of output feature. Default: (3, ).
dilations (Sequence[int]): Dilation of each stage.
style (str): ``pytorch`` or ``caffe``. If set to "pytorch", the
stride-two layer is the 3x3 conv layer, otherwise the stride-two
layer is the first 1x1 conv layer. Default: ``pytorch``.
frozen_stages (int): Stages to be frozen (all param fixed). -1 means
not freezing any parameters. Default: -1.
conv_cfg (dict): Config for norm layers. Default: dict(type='Conv').
norm_cfg (dict):
Config for norm layers. required keys are `type` and
`requires_grad`. Default: dict(type='BN2d', requires_grad=True).
act_cfg (dict): Config for activate layers.
Default: dict(type='ReLU', inplace=True).
norm_eval (bool): Whether to set BN layers to eval mode, namely, freeze
running stats (mean and var). Default: False.
partial_bn (bool): Whether to use partial bn. Default: False.
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
memory while slowing down the training speed. Default: False.
"""
arch_settings = {
18: (BasicBlock, (2, 2, 2, 2)),
34: (BasicBlock, (3, 4, 6, 3)),
50: (Bottleneck, (3, 4, 6, 3)),
101: (Bottleneck, (3, 4, 23, 3)),
152: (Bottleneck, (3, 8, 36, 3))
}
def __init__(self,
depth,
pretrained=None,
torchvision_pretrain=True,
in_channels=3,
num_stages=4,
out_indices=(3, ),
strides=(1, 2, 2, 2),
dilations=(1, 1, 1, 1),
style='pytorch',
frozen_stages=-1,
conv_cfg=dict(type='Conv'),
norm_cfg=dict(type='BN2d', requires_grad=True),
act_cfg=dict(type='ReLU', inplace=True),
norm_eval=False,
partial_bn=False,
with_cp=False):
super().__init__()
if depth not in self.arch_settings:
raise KeyError(f'invalid depth {depth} for resnet')
self.depth = depth
self.in_channels = in_channels
self.pretrained = pretrained
self.torchvision_pretrain = torchvision_pretrain
self.num_stages = num_stages
assert 1 <= num_stages <= 4
self.out_indices = out_indices
assert max(out_indices) < num_stages
self.strides = strides
self.dilations = dilations
assert len(strides) == len(dilations) == num_stages
self.style = style
self.frozen_stages = frozen_stages
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.partial_bn = partial_bn
self.with_cp = with_cp
self.block, stage_blocks = self.arch_settings[depth]
self.stage_blocks = stage_blocks[:num_stages]
self.inplanes = 64
self._make_stem_layer()
self.res_layers = []
for i, num_blocks in enumerate(self.stage_blocks):
stride = strides[i]
dilation = dilations[i]
planes = 64 * 2**i
res_layer = make_res_layer(
self.block,
self.inplanes,
planes,
num_blocks,
stride=stride,
dilation=dilation,
style=self.style,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
with_cp=with_cp)
self.inplanes = planes * self.block.expansion
layer_name = f'layer{i + 1}'
self.add_module(layer_name, res_layer)
self.res_layers.append(layer_name)
self.feat_dim = self.block.expansion * 64 * 2**(
len(self.stage_blocks) - 1)
def _make_stem_layer(self):
"""Construct the stem layers consists of a conv+norm+act module and a
pooling layer."""
self.conv1 = ConvModule(
self.in_channels,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
@staticmethod
def _load_conv_params(conv, state_dict_tv, module_name_tv,
loaded_param_names):
"""Load the conv parameters of resnet from torchvision.
Args:
conv (nn.Module): The destination conv module.
state_dict_tv (OrderedDict): The state dict of pretrained
torchvision model.
module_name_tv (str): The name of corresponding conv module in the
torchvision model.
loaded_param_names (list[str]): List of parameters that have been
loaded.
"""
weight_tv_name = module_name_tv + '.weight'
if conv.weight.data.shape == state_dict_tv[weight_tv_name].shape:
conv.weight.data.copy_(state_dict_tv[weight_tv_name])
loaded_param_names.append(weight_tv_name)
if getattr(conv, 'bias') is not None:
bias_tv_name = module_name_tv + '.bias'
if conv.bias.data.shape == state_dict_tv[bias_tv_name].shape:
conv.bias.data.copy_(state_dict_tv[bias_tv_name])
loaded_param_names.append(bias_tv_name)
@staticmethod
def _load_bn_params(bn, state_dict_tv, module_name_tv, loaded_param_names):
"""Load the bn parameters of resnet from torchvision.
Args:
bn (nn.Module): The destination bn module.
state_dict_tv (OrderedDict): The state dict of pretrained
torchvision model.
module_name_tv (str): The name of corresponding bn module in the
torchvision model.
loaded_param_names (list[str]): List of parameters that have been
loaded.
"""
for param_name, param in bn.named_parameters():
param_tv_name = f'{module_name_tv}.{param_name}'
param_tv = state_dict_tv[param_tv_name]
if param.data.shape == param_tv.shape:
param.data.copy_(param_tv)
loaded_param_names.append(param_tv_name)
for param_name, param in bn.named_buffers():
param_tv_name = f'{module_name_tv}.{param_name}'
# some buffers like num_batches_tracked may not exist
if param_tv_name in state_dict_tv:
param_tv = state_dict_tv[param_tv_name]
if param.data.shape == param_tv.shape:
param.data.copy_(param_tv)
loaded_param_names.append(param_tv_name)
def _load_torchvision_checkpoint(self, logger=None):
"""Initiate the parameters from torchvision pretrained checkpoint."""
state_dict_torchvision = _load_checkpoint(self.pretrained)
if 'state_dict' in state_dict_torchvision:
state_dict_torchvision = state_dict_torchvision['state_dict']
loaded_param_names = []
for name, module in self.named_modules():
if isinstance(module, ConvModule):
# we use a ConvModule to wrap conv+bn+relu layers, thus the
# name mapping is needed
if 'downsample' in name:
# layer{X}.{Y}.downsample.conv->layer{X}.{Y}.downsample.0
original_conv_name = name + '.0'
# layer{X}.{Y}.downsample.bn->layer{X}.{Y}.downsample.1
original_bn_name = name + '.1'
else:
# layer{X}.{Y}.conv{n}.conv->layer{X}.{Y}.conv{n}
original_conv_name = name
# layer{X}.{Y}.conv{n}.bn->layer{X}.{Y}.bn{n}
original_bn_name = name.replace('conv', 'bn')
print(name)
self._load_conv_params(module.conv, state_dict_torchvision,
original_conv_name, loaded_param_names)
self._load_bn_params(module.bn, state_dict_torchvision,
original_bn_name, loaded_param_names)
import pdb
pdb.set_trace()
# check if any parameters in the 2d checkpoint are not loaded
remaining_names = set(
state_dict_torchvision.keys()) - set(loaded_param_names)
if remaining_names:
logger.info(
f'These parameters in pretrained checkpoint are not loaded'
f': {remaining_names}')
def init_weights(self):
"""Initiate the parameters either from existing checkpoint or from
scratch."""
if isinstance(self.pretrained, str):
logger = get_root_logger()
if self.torchvision_pretrain:
# torchvision's
self._load_torchvision_checkpoint(logger)
else:
# ours
load_checkpoint(
self, self.pretrained, strict=False, logger=logger)
elif self.pretrained is None:
for m in self.modules():
if isinstance(m, nn.Conv2d):
kaiming_init(m)
elif isinstance(m, nn.BatchNorm2d):
constant_init(m, 1)
else:
raise TypeError('pretrained must be a str or None')
def forward(self, x):
"""Defines the computation performed at every call.
Args:
x (torch.Tensor): The input data.
Returns:
torch.Tensor: The feature of the input samples extracted
by the backbone.
"""
x = self.conv1(x)
x = self.maxpool(x)
outs = []
for i, layer_name in enumerate(self.res_layers):
res_layer = getattr(self, layer_name)
x = res_layer(x)
if i in self.out_indices:
outs.append(x)
if len(outs) == 1:
return outs[0]
return tuple(outs)
def _freeze_stages(self):
"""Prevent all the parameters from being optimized before
``self.frozen_stages``."""
if self.frozen_stages >= 0:
self.conv1.bn.eval()
for m in self.conv1.modules():
for param in m.parameters():
param.requires_grad = False
for i in range(1, self.frozen_stages + 1):
m = getattr(self, f'layer{i}')
m.eval()
for param in m.parameters():
param.requires_grad = False
def _partial_bn(self):
logger = get_root_logger()
logger.info('Freezing BatchNorm2D except the first one.')
count_bn = 0
for m in self.modules():
if isinstance(m, nn.BatchNorm2d):
count_bn += 1
if count_bn >= 2:
m.eval()
# shutdown update in frozen mode
m.weight.requires_grad = False
m.bias.requires_grad = False
def train(self, mode=True):
"""Set the optimization status when training."""
super().train(mode)
self._freeze_stages()
if mode and self.norm_eval:
for m in self.modules():
if isinstance(m, _BatchNorm):
m.eval()
if mode and self.partial_bn:
self._partial_bn()
class ShuffleNetV1(BaseBackbone):
"""ShuffleNetV1 backbone.
Args:
groups (int): The number of groups to be used in grouped 1x1
convolutions in each ShuffleUnit. Default: 3.
widen_factor (float): Width multiplier - adjusts the number
of channels in each layer by this amount. Default: 1.0.
out_indices (Sequence[int]): Output from which stages.
Default: (2, )
frozen_stages (int): Stages to be frozen (all param fixed).
Default: -1, which means not freezing any parameters.
conv_cfg (dict, optional): Config dict for convolution layer.
Default: None, which means using conv2d.
norm_cfg (dict): Config dict for normalization layer.
Default: dict(type='BN').
act_cfg (dict): Config dict for activation layer.
Default: dict(type='ReLU').
norm_eval (bool): Whether to set norm layers to eval mode, namely,
freeze running stats (mean and var). Note: Effect on Batch Norm
and its variants only. Default: False.
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
memory while slowing down the training speed. Default: False.
"""
def __init__(self,
groups=3,
widen_factor=1.0,
out_indices=(2, ),
frozen_stages=-1,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
norm_eval=False,
with_cp=False):
super(ShuffleNetV1, self).__init__()
self.stage_blocks = [4, 8, 4]
self.groups = groups
for index in out_indices:
if index not in range(0, 3):
raise ValueError('the item in out_indices must in '
f'range(0, 3). But received {index}')
if frozen_stages not in range(-1, 3):
raise ValueError('frozen_stages must be in range(-1, 3). '
f'But received {frozen_stages}')
self.out_indices = out_indices
self.frozen_stages = frozen_stages
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
if groups == 1:
channels = (144, 288, 576)
elif groups == 2:
channels = (200, 400, 800)
elif groups == 3:
channels = (240, 480, 960)
elif groups == 4:
channels = (272, 544, 1088)
elif groups == 8:
channels = (384, 768, 1536)
else:
raise ValueError(f'{groups} groups is not supported for 1x1 '
'Grouped Convolutions')
channels = [make_divisible(ch * widen_factor, 8) for ch in channels]
self.in_channels = int(24 * widen_factor)
self.conv1 = ConvModule(in_channels=3,
out_channels=self.in_channels,
kernel_size=3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layers = nn.ModuleList()
for i, num_blocks in enumerate(self.stage_blocks):
first_block = True if i == 0 else False
layer = self.make_layer(channels[i], num_blocks, first_block)
self.layers.append(layer)
def _freeze_stages(self):
if self.frozen_stages >= 0:
for param in self.conv1.parameters():
param.requires_grad = False
for i in range(self.frozen_stages):
layer = self.layers[i]
layer.eval()
for param in layer.parameters():
param.requires_grad = False
def init_weights(self, pretrained=None):
if isinstance(pretrained, str):
logger = logging.getLogger()
load_checkpoint(self, pretrained, strict=False, logger=logger)
elif pretrained is None:
for name, m in self.named_modules():
if isinstance(m, nn.Conv2d):
if 'conv1' in name:
normal_init(m, mean=0, std=0.01)
else:
normal_init(m, mean=0, std=1.0 / m.weight.shape[1])
elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
constant_init(m.weight, val=1, bias=0.0001)
if isinstance(m, _BatchNorm):
if m.running_mean is not None:
nn.init.constant_(m.running_mean, 0)
else:
raise TypeError('pretrained must be a str or None. But received '
f'{type(pretrained)}')
def make_layer(self, out_channels, num_blocks, first_block=False):
""" Stack ShuffleUnit blocks to make a layer.
Args:
out_channels (int): out_channels of the block.
num_blocks (int): Number of blocks.
first_block (bool): Whether is the first ShuffleUnit of a
sequential ShuffleUnits. Default: False, which means not using
the grouped 1x1 convolution.
"""
layers = []
for i in range(num_blocks):
first_block = first_block if i == 0 else False
combine_mode = 'concat' if i == 0 else 'add'
layers.append(
ShuffleUnit(self.in_channels,
out_channels,
groups=self.groups,
first_block=first_block,
combine=combine_mode,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg,
with_cp=self.with_cp))
self.in_channels = out_channels
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.maxpool(x)
outs = []
for i, layer in enumerate(self.layers):
x = layer(x)
if i in self.out_indices:
outs.append(x)
if len(outs) == 1:
return outs[0]
else:
return tuple(outs)
def train(self, mode=True):
super(ShuffleNetV1, self).train(mode)
self._freeze_stages()
if mode and self.norm_eval:
for m in self.modules():
if isinstance(m, _BatchNorm):
m.eval()
def __init__(self,
in_channels,
max_channels,
num_convs=5,
fc_in_channels=None,
fc_out_channels=1024,
kernel_size=5,
conv_cfg=None,
norm_cfg=None,
act_cfg=dict(type='ReLU'),
out_act_cfg=dict(type='ReLU'),
with_input_norm=True,
with_out_convs=False,
with_spectral_norm=False,
**kwargs):
super().__init__()
if fc_in_channels is not None:
assert fc_in_channels > 0
self.max_channels = max_channels
self.with_fc = fc_in_channels is not None
self.num_convs = num_convs
self.with_out_act = out_act_cfg is not None
self.with_out_convs = with_out_convs
cur_channels = in_channels
for i in range(num_convs):
out_ch = min(64 * 2**i, max_channels)
norm_cfg_ = norm_cfg
act_cfg_ = act_cfg
if i == 0 and not with_input_norm:
norm_cfg_ = None
elif (i == num_convs - 1 and not self.with_fc
and not self.with_out_convs):
norm_cfg_ = None
act_cfg_ = out_act_cfg
self.add_module(
f'conv{i + 1}',
ConvModule(cur_channels,
out_ch,
kernel_size=kernel_size,
stride=2,
padding=kernel_size // 2,
norm_cfg=norm_cfg_,
act_cfg=act_cfg_,
with_spectral_norm=with_spectral_norm,
**kwargs))
cur_channels = out_ch
if self.with_out_convs:
cur_channels = min(64 * 2**(num_convs - 1), max_channels)
out_ch = min(64 * 2**num_convs, max_channels)
self.add_module(
f'conv{num_convs + 1}',
ConvModule(cur_channels,
out_ch,
kernel_size,
stride=1,
padding=kernel_size // 2,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
with_spectral_norm=with_spectral_norm,
**kwargs))
self.add_module(
f'conv{num_convs + 2}',
ConvModule(out_ch,
1,
kernel_size,
stride=1,
padding=kernel_size // 2,
act_cfg=None,
with_spectral_norm=with_spectral_norm,
**kwargs))
if self.with_fc:
self.fc = LinearModule(fc_in_channels,
fc_out_channels,
bias=True,
act_cfg=out_act_cfg,
with_spectral_norm=with_spectral_norm)
def __init__(self,
groups=3,
widen_factor=1.0,
out_indices=(2, ),
frozen_stages=-1,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
norm_eval=False,
with_cp=False):
super(ShuffleNetV1, self).__init__()
self.stage_blocks = [4, 8, 4]
self.groups = groups
for index in out_indices:
if index not in range(0, 3):
raise ValueError('the item in out_indices must in '
f'range(0, 3). But received {index}')
if frozen_stages not in range(-1, 3):
raise ValueError('frozen_stages must be in range(-1, 3). '
f'But received {frozen_stages}')
self.out_indices = out_indices
self.frozen_stages = frozen_stages
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
if groups == 1:
channels = (144, 288, 576)
elif groups == 2:
channels = (200, 400, 800)
elif groups == 3:
channels = (240, 480, 960)
elif groups == 4:
channels = (272, 544, 1088)
elif groups == 8:
channels = (384, 768, 1536)
else:
raise ValueError(f'{groups} groups is not supported for 1x1 '
'Grouped Convolutions')
channels = [make_divisible(ch * widen_factor, 8) for ch in channels]
self.in_channels = int(24 * widen_factor)
self.conv1 = ConvModule(in_channels=3,
out_channels=self.in_channels,
kernel_size=3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layers = nn.ModuleList()
for i, num_blocks in enumerate(self.stage_blocks):
first_block = True if i == 0 else False
layer = self.make_layer(channels[i], num_blocks, first_block)
self.layers.append(layer)
def __init__(self,
gamma_w=1.0,
gamma_b=1.0,
gamma_d=1.0,
pretrained=None,
in_channels=3,
num_stages=4,
spatial_strides=(2, 2, 2, 2),
frozen_stages=-1,
se_style='half',
se_ratio=1 / 16,
use_swish=True,
conv_cfg=dict(type='Conv3d'),
norm_cfg=dict(type='BN3d', requires_grad=True),
act_cfg=dict(type='ReLU', inplace=True),
norm_eval=False,
with_cp=False,
zero_init_residual=True,
**kwargs):
super().__init__()
self.gamma_w = gamma_w
self.gamma_b = gamma_b
self.gamma_d = gamma_d
self.pretrained = pretrained
self.in_channels = in_channels
# Hard coded, can be changed by gamma_w
self.base_channels = 24
self.stage_blocks = [1, 2, 5, 3]
# apply parameters gamma_w and gamma_d
self.base_channels = self._round_width(self.base_channels,
self.gamma_w)
self.stage_blocks = [
self._round_repeats(x, self.gamma_d) for x in self.stage_blocks
]
self.num_stages = num_stages
assert 1 <= num_stages <= 4
self.spatial_strides = spatial_strides
assert len(spatial_strides) == num_stages
self.frozen_stages = frozen_stages
self.se_style = se_style
assert self.se_style in ['all', 'half']
self.se_ratio = se_ratio
assert (self.se_ratio is None) or (self.se_ratio > 0)
self.use_swish = use_swish
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.zero_init_residual = zero_init_residual
self.block = BlockX3D
self.stage_blocks = self.stage_blocks[:num_stages]
self.layer_inplanes = self.base_channels
self._make_stem_layer()
self.res_layers = []
for i, num_blocks in enumerate(self.stage_blocks):
spatial_stride = spatial_strides[i]
inplanes = self.base_channels * 2**i
planes = int(inplanes * self.gamma_b)
res_layer = self.make_res_layer(self.block,
self.layer_inplanes,
inplanes,
planes,
num_blocks,
spatial_stride=spatial_stride,
se_style=self.se_style,
se_ratio=self.se_ratio,
use_swish=self.use_swish,
norm_cfg=self.norm_cfg,
conv_cfg=self.conv_cfg,
act_cfg=self.act_cfg,
with_cp=with_cp,
**kwargs)
self.layer_inplanes = inplanes
layer_name = f'layer{i + 1}'
self.add_module(layer_name, res_layer)
self.res_layers.append(layer_name)
self.feat_dim = self.base_channels * 2**(len(self.stage_blocks) - 1)
self.conv5 = ConvModule(self.feat_dim,
int(self.feat_dim * self.gamma_b),
kernel_size=1,
stride=1,
padding=0,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
self.feat_dim = int(self.feat_dim * self.gamma_b)
def __init__(self,
in_channels,
out_channels,
groups=3,
first_block=True,
combine='add',
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
with_cp=False):
super(ShuffleUnit, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.first_block = first_block
self.combine = combine
self.groups = groups
self.bottleneck_channels = self.out_channels // 4
self.with_cp = with_cp
if self.combine == 'add':
self.depthwise_stride = 1
self._combine_func = self._add
assert in_channels == out_channels, (
'in_channels must be equal to out_channels when combine '
'is add')
elif self.combine == 'concat':
self.depthwise_stride = 2
self._combine_func = self._concat
self.out_channels -= self.in_channels
self.avgpool = nn.AvgPool2d(kernel_size=3, stride=2, padding=1)
else:
raise ValueError(f'Cannot combine tensors with {self.combine}. '
'Only "add" and "concat" are supported')
self.first_1x1_groups = 1 if first_block else self.groups
self.g_conv_1x1_compress = ConvModule(
in_channels=self.in_channels,
out_channels=self.bottleneck_channels,
kernel_size=1,
groups=self.first_1x1_groups,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.depthwise_conv3x3_bn = ConvModule(
in_channels=self.bottleneck_channels,
out_channels=self.bottleneck_channels,
kernel_size=3,
stride=self.depthwise_stride,
padding=1,
groups=self.bottleneck_channels,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
self.g_conv_1x1_expand = ConvModule(
in_channels=self.bottleneck_channels,
out_channels=self.out_channels,
kernel_size=1,
groups=self.groups,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
self.act = build_activation_layer(act_cfg)
def __init__(self,
inplanes,
planes,
outplanes,
spatial_stride=1,
downsample=None,
se_ratio=None,
use_swish=True,
conv_cfg=dict(type='Conv3d'),
norm_cfg=dict(type='BN3d'),
act_cfg=dict(type='ReLU'),
with_cp=False):
super().__init__()
self.inplanes = inplanes
self.planes = planes
self.outplanes = outplanes
self.spatial_stride = spatial_stride
self.downsample = downsample
self.se_ratio = se_ratio
self.use_swish = use_swish
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.act_cfg_swish = dict(type='Swish')
self.with_cp = with_cp
self.conv1 = ConvModule(in_channels=inplanes,
out_channels=planes,
kernel_size=1,
stride=1,
padding=0,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
# Here we use the channel-wise conv
self.conv2 = ConvModule(in_channels=planes,
out_channels=planes,
kernel_size=3,
stride=(1, self.spatial_stride,
self.spatial_stride),
padding=1,
groups=planes,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=None)
self.swish = Swish()
self.conv3 = ConvModule(in_channels=planes,
out_channels=outplanes,
kernel_size=1,
stride=1,
padding=0,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=None)
if self.se_ratio is not None:
self.se_module = SEModule(planes, self.se_ratio)
self.relu = build_activation_layer(self.act_cfg)
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
add_extra_convs=False,
extra_convs_on_inputs=True,
relu_before_extra_convs=False,
no_norm_on_lateral=False,
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
upsample_cfg=dict(mode='nearest')):
super(NoFPN, self).__init__()
assert isinstance(in_channels, list)
self.add_extra_convs = add_extra_convs
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels)
self.num_outs = num_outs
self.fp16_enabled = False
self.upsample_cfg = upsample_cfg.copy()
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.lateral_convs = nn.ModuleList()
self.fpn_convs = nn.ModuleList()
self.spatail_gates = nn.ModuleList()
self.downsample_convs = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
d_conv = nn.ModuleList()
for j in range(self.start_level, self.backbone_end_level):
if i > j:
d_conv.append(
ConvModule(out_channels,
out_channels,
1 + 2 * (i - j),
stride=2**(i - j),
padding=1,
inplace=False))
self.spatail_gates.append(
SpatialGate(levels=self.num_ins - start_level))
self.downsample_convs.append(d_conv)
# 1, 4
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(in_channels[i], out_channels, 1, inplace=False)
fpn_conv = ConvModule(out_channels,
out_channels,
3,
padding=1,
inplace=False)
self.lateral_convs.append(l_conv)
self.fpn_convs.append(fpn_conv)
# add extra conv layers (e.g., RetinaNet)
extra_levels = num_outs - self.backbone_end_level + self.start_level
if self.add_extra_convs and extra_levels >= 1:
for i in range(extra_levels):
if i == 0 and self.add_extra_convs == 'on_input':
# 2048
in_channels = self.in_channels[self.backbone_end_level - 1]
else:
# 256
in_channels = out_channels
extra_fpn_conv = ConvModule(in_channels,
out_channels,
3,
stride=2,
padding=1,
inplace=False)
self.fpn_convs.append(extra_fpn_conv)
def __init__(self,
in_channels,
out_channels,
num_outs,
start_level=0,
end_level=-1,
add_extra_convs=False,
extra_convs_on_inputs=True,
relu_before_extra_convs=False,
no_norm_on_lateral=False,
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
upsample_cfg=dict(mode='nearest')):
super().__init__()
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels)
self.num_outs = num_outs
self.relu_before_extra_convs = relu_before_extra_convs
self.no_norm_on_lateral = no_norm_on_lateral
self.fp16_enabled = False
self.upsample_cfg = upsample_cfg.copy()
if end_level == -1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level < inputs, no extra level is allowed
self.backbone_end_level = end_level
assert end_level <= len(in_channels)
assert num_outs == end_level - start_level
self.start_level = start_level
self.end_level = end_level
self.add_extra_convs = add_extra_convs
assert isinstance(add_extra_convs, (str, bool))
if isinstance(add_extra_convs, str):
# Extra_convs_source choices: 'on_input', 'on_lateral', 'on_output'
assert add_extra_convs in ('on_input', 'on_lateral', 'on_output')
elif add_extra_convs: # True
if extra_convs_on_inputs:
# TODO: deprecate `extra_convs_on_inputs`
warnings.simplefilter('once')
warnings.warn(
'"extra_convs_on_inputs" will be deprecated in v2.9.0,'
'Please use "add_extra_convs"', DeprecationWarning)
self.add_extra_convs = 'on_input'
else:
self.add_extra_convs = 'on_output'
self.lateral_convs = nn.ModuleList()
self.fpn_convs = nn.ModuleList()
concat_in_channels = []
for i in range(self.start_level, self.backbone_end_level):
if i == self.backbone_end_level - 1:
concat_in_channels.append(in_channels[i])
else:
concat_in_channels.append(in_channels[i] + out_channels[i + 1])
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(
concat_in_channels[i],
out_channels[i],
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
fpn_conv = ConvModule(out_channels[i],
out_channels[i],
3,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.lateral_convs.append(l_conv)
self.fpn_convs.append(fpn_conv)
# add extra conv layers (e.g., RetinaNet)
extra_levels = num_outs - self.backbone_end_level + self.start_level
if self.add_extra_convs and extra_levels >= 1:
for i in range(extra_levels):
if i == 0 and self.add_extra_convs == 'on_input':
in_channels = self.in_channels[self.backbone_end_level - 1]
else:
in_channels = out_channels
extra_fpn_conv = ConvModule(in_channels,
out_channels,
3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.fpn_convs.append(extra_fpn_conv)
def __init__(
self,
in_channels=[64, 128, 256, 512],
out_channels=128,
num_outs=4,
start_level=0,
end_level=-1,
add_extra_convs=False,
relu_before_extra_convs=False,
no_norm_on_lateral=False,
conv_cfg=None,
norm_cfg=None,
act_cfg=None,
num_heads=[2, 2, 2, 2],
mlp_ratios=[4, 4, 4, 4],
sr_ratios=[8, 4, 2, 1],
qkv_bias=True,
qk_scale=None,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.,
transformer_norm_cfg=dict(type='LN'),
use_sr_conv=False,
):
super().__init__()
assert isinstance(in_channels, list)
self.in_channels = in_channels
self.out_channels = out_channels
self.num_ins = len(in_channels)
self.num_outs = num_outs
self.no_norm_on_lateral = no_norm_on_lateral
self.fp16_enabled = False
self.norm_cfg = norm_cfg
self.conv_cfg = conv_cfg
self.act_cfg = act_cfg
self.mlp_ratios = mlp_ratios
if end_level == -1 or end_level == self.num_ins - 1:
self.backbone_end_level = self.num_ins
assert num_outs >= self.num_ins - start_level
else:
# if end_level is not the last level, no extra level is allowed
self.backbone_end_level = end_level + 1
assert end_level < self.num_ins
assert num_outs == end_level - start_level + 1
self.start_level = start_level
self.end_level = end_level
self.lateral_convs = nn.ModuleList()
self.merge_blocks = nn.ModuleList()
for i in range(self.start_level, self.backbone_end_level):
l_conv = ConvModule(
in_channels[i],
out_channels,
1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg if not self.no_norm_on_lateral else None,
act_cfg=act_cfg,
inplace=False)
self.lateral_convs.append(l_conv)
for i in range(self.start_level, self.backbone_end_level - 1):
merge_block = TCFormerDynamicBlock(
dim=out_channels,
num_heads=num_heads[i],
mlp_ratio=mlp_ratios[i],
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=drop_path_rate,
norm_cfg=transformer_norm_cfg,
sr_ratio=sr_ratios[i],
use_sr_conv=use_sr_conv)
self.merge_blocks.append(merge_block)
# add extra conv layers (e.g., RetinaNet)
self.relu_before_extra_convs = relu_before_extra_convs
self.add_extra_convs = add_extra_convs
assert isinstance(add_extra_convs, (str, bool))
if isinstance(add_extra_convs, str):
# Extra_convs_source choices: 'on_input', 'on_output'
assert add_extra_convs in ('on_input', 'on_output')
elif add_extra_convs: # True
self.add_extra_convs = 'on_input'
self.extra_convs = nn.ModuleList()
extra_levels = num_outs - (self.end_level + 1 - self.start_level)
if self.add_extra_convs and extra_levels >= 1:
for i in range(extra_levels):
if i == 0 and self.add_extra_convs == 'on_input':
in_channels = self.in_channels[self.end_level]
else:
in_channels = out_channels
extra_fpn_conv = ConvModule(
in_channels,
out_channels,
3,
stride=2,
padding=1,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg,
inplace=False)
self.extra_convs.append(extra_fpn_conv)
def __init__(self,
num_point: int,
radii: List[float],
sample_nums: List[int],
mlp_channels: List[List[int]],
fps_mod: List[str] = ['D-FPS'],
fps_sample_range_list: List[int] = [-1],
dilated_group: bool = False,
norm_cfg: dict = dict(type='BN2d'),
use_xyz: bool = True,
pool_mod='max',
normalize_xyz: bool = False,
bias='auto'):
super().__init__()
assert len(radii) == len(sample_nums) == len(mlp_channels)
assert pool_mod in ['max', 'avg']
assert isinstance(fps_mod, list) or isinstance(fps_mod, tuple)
assert isinstance(fps_sample_range_list, list) or isinstance(
fps_sample_range_list, tuple)
assert len(fps_mod) == len(fps_sample_range_list)
if isinstance(mlp_channels, tuple):
mlp_channels = list(map(list, mlp_channels))
if isinstance(num_point, int):
self.num_point = [num_point]
elif isinstance(num_point, list) or isinstance(num_point, tuple):
self.num_point = num_point
else:
raise NotImplementedError('Error type of num_point!')
self.pool_mod = pool_mod
self.groupers = nn.ModuleList()
self.mlps = nn.ModuleList()
self.fps_mod_list = fps_mod
self.fps_sample_range_list = fps_sample_range_list
self.points_sampler = Points_Sampler(self.num_point, self.fps_mod_list,
self.fps_sample_range_list)
for i in range(len(radii)):
radius = radii[i]
sample_num = sample_nums[i]
if num_point is not None:
if dilated_group and i != 0:
min_radius = radii[i - 1]
else:
min_radius = 0
grouper = QueryAndGroup(
radius,
sample_num,
min_radius=min_radius,
use_xyz=use_xyz,
normalize_xyz=normalize_xyz)
else:
grouper = GroupAll(use_xyz)
self.groupers.append(grouper)
mlp_spec = mlp_channels[i]
if use_xyz:
mlp_spec[0] += 3
mlp = nn.Sequential()
for i in range(len(mlp_spec) - 1):
mlp.add_module(
f'layer{i}',
ConvModule(
mlp_spec[i],
mlp_spec[i + 1],
kernel_size=(1, 1),
stride=(1, 1),
conv_cfg=dict(type='Conv2d'),
norm_cfg=norm_cfg,
bias=bias))
self.mlps.append(mlp)
def __init__(self,
in_channels,
out_channels,
base_channels=64,
norm_cfg=dict(type='IN'),
use_dropout=False,
num_blocks=9,
padding_mode='reflect',
init_cfg=dict(type='normal', gain=0.02)):
super().__init__()
assert num_blocks >= 0, ('Number of residual blocks must be '
f'non-negative, but got {num_blocks}.')
assert isinstance(norm_cfg, dict), ("'norm_cfg' should be dict, but"
f'got {type(norm_cfg)}')
assert 'type' in norm_cfg, "'norm_cfg' must have key 'type'"
# We use norm layers in the resnet generator.
# Only for IN, use bias since it does not have affine parameters.
use_bias = norm_cfg['type'] == 'IN'
model = []
model += [
ConvModule(in_channels=in_channels,
out_channels=base_channels,
kernel_size=7,
padding=3,
bias=use_bias,
norm_cfg=norm_cfg,
padding_mode=padding_mode)
]
num_down = 2
# add downsampling layers
for i in range(num_down):
multiple = 2**i
model += [
ConvModule(in_channels=base_channels * multiple,
out_channels=base_channels * multiple * 2,
kernel_size=3,
stride=2,
padding=1,
bias=use_bias,
norm_cfg=norm_cfg)
]
# add residual blocks
multiple = 2**num_down
for i in range(num_blocks):
model += [
ResidualBlockWithDropout(base_channels * multiple,
padding_mode=padding_mode,
norm_cfg=norm_cfg,
use_dropout=use_dropout)
]
# add upsampling layers
for i in range(num_down):
multiple = 2**(num_down - i)
model += [
ConvModule(in_channels=base_channels * multiple,
out_channels=base_channels * multiple // 2,
kernel_size=3,
stride=2,
padding=1,
bias=use_bias,
conv_cfg=dict(type='Deconv', output_padding=1),
norm_cfg=norm_cfg)
]
model += [
ConvModule(in_channels=base_channels,
out_channels=out_channels,
kernel_size=7,
padding=3,
bias=True,
norm_cfg=None,
act_cfg=dict(type='Tanh'),
padding_mode=padding_mode)
]
self.model = nn.Sequential(*model)
self.init_type = 'normal' if init_cfg is None else init_cfg.get(
'type', 'normal')
self.init_gain = 0.02 if init_cfg is None else init_cfg.get(
'gain', 0.02)