def __init__(self,
inp,
hidden_dim,
oup,
kernel_size,
stride,
use_se=False,
use_hs=False,
momentum=0.1):
"""Init InvertedResidualSE."""
super(InvertedResidualSE, self).__init__()
self.identity = stride == 1 and inp == oup
self.ir_block = Sequential(
# pw
ops.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
ops.BatchNorm2d(hidden_dim, momentum=momentum),
ops.Hswish() if use_hs else ops.Relu(inplace=True),
# dw
ops.Conv2d(hidden_dim,
hidden_dim,
kernel_size,
stride, (kernel_size - 1) // 2,
groups=hidden_dim,
bias=False),
ops.BatchNorm2d(hidden_dim, momentum=momentum),
# Squeeze-and-Excite
SELayer(hidden_dim) if use_se else Sequential(),
ops.Hswish() if use_hs else ops.Relu(inplace=True),
# pw-linear
ops.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
ops.BatchNorm2d(oup, momentum=momentum),
)
def transform_model(model):
"""Transform the torch model to Vega model."""
new_model_dict = OrderedDict()
for name, module in model.name_cells().items():
if isinstance(module, atom_op):
if isinstance(module, nn.Flatten):
new_model_dict["mean"] = ops.AdaptiveAvgPool2d()
new_model_dict[name] = _transform_op(module)
sub_modules = OrderedDict()
if isinstance(module, nn.SequentialCell):
for sub_name, sub_module in module.name_cells().items():
if isinstance(sub_module, atom_block):
sub_modules[sub_name] = _transform_block(sub_module)
sub_modules[sub_name].update_parameters_name(name + "." +
sub_name +
".")
if isinstance(sub_module, atom_op):
sub_modules[sub_name] = _transform_op(sub_module)
sub_modules[sub_name].update_parameters_name(name + "." +
sub_name +
".")
new_model_dict[name] = Sequential(sub_modules)
model = Sequential(new_model_dict)
desc = model.to_desc()
vega_model = NetworkDesc(desc).to_model()
return vega_model
def __init__(self,
cfgs,
mode='small',
input_channel=3,
feat_channels=16,
special_stride=1,
num_classes=10,
width_mult=1.,
block=InvertedResidualSE,
momentum=0.1,
is_prune_mode=False,
**kwargs):
"""Init MobileNetV3.
:params cfgs: cfgs for mobilenetv3
:type cfgs: list
:params special_stride: the stride of the first InvertedResidualSE block.
:type special_stride: int (1 for cifar10, 2 for imagenet)
"""
super(MobileNetV3, self).__init__()
self.cfgs = cfgs
# building first layer
if not is_prune_mode:
feat_channels = _make_divisible(feat_channels * width_mult, 8)
else:
feat_channels = int(feat_channels * width_mult)
layers = [
ConvBnAct(input_channel,
feat_channels,
kernel_size=3,
momentum=momentum,
stride=special_stride,
padding=1,
activation='hswish')
]
# buidling blocks
# kernel_size, expand_ratio, output_channels, use_se, use_hs, stride
for k, t, c, use_se, use_hs, s in self.cfgs:
output_channel = _make_divisible(
c * width_mult, 8) if not is_prune_mode else int(c *
width_mult)
hidden_dim = _make_divisible(t, 8) if not is_prune_mode else t
layers.append(
block(feat_channels, hidden_dim, output_channel, k, s, use_se,
use_hs, momentum))
feat_channels = output_channel
self.features = Sequential(*layers)
# building last linear layer
self.avgpool = ops.AdaptiveAvgPool2d((1, 1))
chn = 1280 if mode == 'large' else 1024
self.classifier = Sequential(ops.View(),
ops.Linear(feat_channels, chn),
ops.Hswish(), ops.Dropout(0.2),
ops.Linear(chn, num_classes))
self._initialize_weights()
def __init__(self, encoding, n_class=1000):
super(DNet, self).__init__()
op_names = ["conv3", "conv1", "conv3_grp2", "conv3_grp4", "conv3_base1", "conv3_base32", "conv3_sep"]
block_str, num_channel, macro_str = encoding.split('_')
curr_channel, index = int(num_channel), 0
_big_model = "*" in block_str
if _big_model:
block_encoding_list = block_str.split('*')
# stem
self.layers = Sequential(
create_op('conv3', 3, curr_channel // 2, stride=2),
ops.Relu(),
create_op('conv3', curr_channel // 2, curr_channel // 2),
ops.Relu(),
create_op('conv3', curr_channel // 2, curr_channel, stride=2),
ops.Relu()
)
# body
if not _big_model:
while index < len(macro_str):
stride = 1
if macro_str[index] == '-':
stride = 2
index += 1
channel_increase = int(macro_str[index])
block = EncodedBlock(block_str, curr_channel, op_names, stride, channel_increase)
self.layers.append(block)
curr_channel *= channel_increase
index += 1
else:
block_encoding_index = 0
while index < len(macro_str):
stride = 1
if macro_str[index] == '-':
stride = 2
index += 1
block_encoding_index += 1
channel_increase = int(macro_str[index])
block_encoding = block_encoding_list[block_encoding_index]
block = EncodedBlock(block_encoding, curr_channel, op_names, stride, channel_increase)
self.layers.append(block)
curr_channel *= channel_increase
index += 1
self.layers.append(ops.AdaptiveAvgPool2d((1, 1)))
self.view = ops.View()
self.fc = ops.Linear(in_features=curr_channel, out_features=n_class)
def __init__(self, in_chnls, cardinality, group_depth, stride):
super(ResNeXt_Block, self).__init__()
self.group_chnls = cardinality * group_depth
self.conv1 = BN_Conv2d(in_chnls,
self.group_chnls,
1,
stride=1,
padding=0)
self.conv2 = BN_Conv2d(self.group_chnls,
self.group_chnls,
3,
stride=stride,
padding=1,
groups=cardinality)
self.conv3 = ops.Conv2d(self.group_chnls,
self.group_chnls * 2,
1,
stride=1,
padding=0)
self.bn = ops.BatchNorm2d(self.group_chnls * 2)
if stride != 1 or in_chnls != self.group_chnls * 2:
self.short_cut = Sequential(
ops.Conv2d(in_chnls,
self.group_chnls * 2,
1,
stride,
bias=False), ops.BatchNorm2d(self.group_chnls * 2))
else:
self.short_cut = None
def make_res_layer_from_code(block,
inplanes,
planes,
blocks,
stride=1,
dilation=1,
style='pytorch',
with_cp=False,
code=None):
"""Make res layer from code."""
if code is None:
return make_res_layer(block, inplanes, planes, blocks, stride,
dilation, style, with_cp)
strides = map(int, code)
layers = []
for stride in strides:
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=stride,
dilation=dilation,
style=style,
with_cp=with_cp))
inplanes = planes * block.expansion
return Sequential(*layers)
def __init__(self,
code='111-2111-211111-211',
block='BottleneckBlock',
in_channels=64,
weight_file=None,
out_layers=None):
"""Init SerialBackbone."""
super(SerialBackbone, self).__init__()
self.inplanes = in_channels
self.planes = self.inplanes
self.weight_file = weight_file
self.channels = [3]
self.code = code.split('-')
self.block = ClassFactory.get_cls(ClassType.NETWORK, block)
self._make_stem_layer()
self.layers = Sequential() if out_layers == -1 else OutDictSequential()
self.make_cells()
def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False):
super(BN_Conv2d, self).__init__()
self.seq = Sequential(
ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation, groups=groups, bias=bias),
ops.BatchNorm2d(out_channels),
ops.Relu()
)
def __init__(self, channel, reduction=4):
"""Init SELayer."""
super(SELayer, self).__init__()
self.avg_pool = ops.AdaptiveAvgPool2d(1)
hidden_dim = _make_divisible(channel // reduction, 8)
self.fc = Sequential(ops.Linear(channel, hidden_dim, use_bias=False),
ops.Relu(inplace=True),
ops.Linear(hidden_dim, channel, use_bias=False),
ops.Hsigmoid())
def _make_layers(self, inplanes, d, blocks, stride, code):
"""Make layer."""
strides = map(int, code)
layers = []
for stride in strides:
layers.append(ResNeXt_Block(
inplanes, self.cardinality, d, stride))
inplanes = self.cardinality * d * 2
return Sequential(*layers), inplanes
def make_layers(self, block, inplanes, planes, code=None):
"""Make ResNet layers."""
strides = list(map(int, code))
layers = []
layers.append(block(inplanes, planes, stride=strides[0]))
inplanes = planes * block.expansion
for stride in strides[1:]:
layers.append(block(inplanes, planes, stride=stride))
inplanes = planes * block.expansion
return Sequential(*layers), inplanes
def make_resnext_layer_from_code(block, inplanes, planes, cardinality=32, dilation=1,
with_cp=False, code=None):
"""Make resnext layer from code."""
strides = list(map(int, code))
layers = []
layers.append(block(in_chnls=inplanes, cardinality=cardinality,
group_depth=planes, stride=strides[0]))
inplanes = planes * cardinality * 2
for stride in strides[1:]:
layers.append(block(in_chnls=inplanes, cardinality=cardinality,
group_depth=planes, stride=stride))
return Sequential(*layers), inplanes
def make_resnet_layer_from_code(block, inplanes, planes, dilation=1,
with_cp=False, code=None):
"""Make resnet layer from code."""
strides = list(map(int, code))
layers = []
layers.append(block(inplanes=inplanes, planes=planes, stride=strides[0], dilation=dilation,
with_cp=with_cp, downsample=True))
inplanes = planes * block.expansion
for stride in strides[1:]:
layers.append(block(inplanes=inplanes, planes=planes, stride=stride, dilation=dilation,
with_cp=with_cp))
inplanes = planes * block.expansion
return Sequential(*layers), inplanes
def __init__(self, encoding):
super(DNetBackbone, self).__init__()
op_names = ["conv3", "conv1", "conv3_grp2", "conv3_grp4", "conv3_base1", "conv3_base32", "conv3_sep"]
# code with kangning
block_str, num_channel, macro_str = encoding.split('_')
curr_channel, index = int(num_channel), 0
_big_model = "*" in block_str
if _big_model:
block_encoding_list = block_str.split('*')
# stem
layers = [
create_op('conv3', 3, curr_channel // 2, stride=2),
ops.Relu(),
create_op('conv3', curr_channel // 2, curr_channel // 2),
ops.Relu(),
create_op('conv3', curr_channel // 2, curr_channel, stride=2),
ops.Relu()
]
# body
if not _big_model:
while index < len(macro_str):
stride = 1
if macro_str[index] == '-':
stride = 2
index += 1
channel_increase = int(macro_str[index])
block = EncodedBlock(block_str, curr_channel, op_names, stride, channel_increase)
layers.append(block)
curr_channel *= channel_increase
index += 1
else:
block_encoding_index = 0
while index < len(macro_str):
stride = 1
if macro_str[index] == '-':
stride = 2
index += 1
block_encoding_index += 1
channel_increase = int(macro_str[index])
block_encoding = block_encoding_list[block_encoding_index]
block = EncodedBlock(block_encoding, curr_channel, op_names, stride, channel_increase)
layers.append(block)
curr_channel *= channel_increase
index += 1
layers.append(ops.AdaptiveAvgPool2d((1, 1)))
self.layers = Sequential(*layers)
def make_res_layer(block,
inplanes,
planes,
blocks,
stride=1,
dilation=1,
style='pytorch',
with_cp=False):
"""Build resnet layer."""
downsample = None
if stride != 1 or inplanes != planes * block.expansion:
conv_layer = ops.Conv2d(inplanes,
planes * block.expansion,
kernel_size=1,
stride=stride,
bias=False)
norm_layer = ops.BatchNorm2d(planes * block.expansion)
downsample = Sequential(conv_layer, norm_layer)
layers = []
layers.append(
block(inplanes=inplanes,
planes=planes,
stride=stride,
dilation=dilation,
downsample=downsample,
style=style,
with_cp=with_cp))
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))
return Sequential(*layers)
def _blocks(self, out_channels, desc_blocks):
blocks = ModuleList()
in_channels = 32
for i in range(desc_blocks):
blocks.append(
Sequential(
ops.Conv2d(in_channels,
out_channels,
padding=1,
kernel_size=3),
ops.BatchNorm2d(out_channels),
ops.Relu(inplace=True),
))
in_channels = out_channels
return blocks
def __init__(self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style='pytorch',
with_cp=False):
"""Init Bottleneck."""
super(Bottleneck, self).__init__()
assert style in ['pytorch', 'caffe']
self.inplanes = inplanes
self.planes = planes
self.stride = stride
self.dilation = dilation
self.style = style
self.with_cp = with_cp
self.norm1 = ops.BatchNorm2d(planes)
self.norm2 = ops.BatchNorm2d(planes)
self.norm3 = ops.BatchNorm2d(planes * self.expansion)
self.conv1 = ops.Conv2d(inplanes, planes, kernel_size=1, bias=False)
self.with_modulated_dcn = False
self.conv2 = ops.Conv2d(
planes,
planes,
kernel_size=3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False,
)
self.conv3 = ops.Conv2d(planes,
planes * self.expansion,
kernel_size=1,
bias=False)
self.relu = ops.Relu(inplace=True)
if stride > 1 or downsample is not None:
conv_layer = ops.Conv2d(inplanes,
planes * self.expansion,
kernel_size=1,
stride=stride,
bias=False)
norm_layer = ops.BatchNorm2d(planes * self.expansion)
self.downsample = Sequential(conv_layer, norm_layer)
else:
self.downsample = None
def _transfowm_model(model):
"""Transform the torch model to Vega model."""
new_model_dict = OrderedDict()
for name, module in model.named_children():
# print("name:", name, "module:", module, "type:", type(module))
if isinstance(module, atom_op):
new_model_dict[name] = _transsorm_op(module)
sub_modules = OrderedDict()
if isinstance(module, nn.Sequential):
for sub_name, sub_module in module.named_children():
if isinstance(sub_module, atom_block):
sub_modules[sub_name] = _transform_block(sub_module)
if isinstance(sub_module, atom_op):
sub_modules[sub_name] = _transsorm_op(sub_module)
new_model_dict[name] = Sequential(sub_modules)
return new_model_dict
def __init__(self, num_classes, backbone='ResNetBackbone', neck='FPN', **kwargs):
"""Create layers.
:param num_class: number of class
:type num_class: int
"""
backbone_cls = ClassFactory.get_instance(ClassType.NETWORK, backbone)
neck_cls = ClassFactory.get_instance(ClassType.NETWORK, neck, in_channels=backbone_cls.out_channels)
backbone_neck = Sequential()
backbone_neck.append(backbone_cls, 'body')
backbone_neck.append(neck_cls, 'fpn')
super(FasterRCNN, self).__init__(backbone_neck, num_classes, **kwargs)
def __init__(self, inp, oup, stride, kernel=3, expand_ratio=1):
"""Construct InvertedResidual class.
:param inp: input channel
:param oup: output channel
:param stride: stride
:param kernel: kernel
:param expand_ratio: channel increase multiplier
"""
super(InvertedConv, self).__init__()
hidden_dim = round(inp * expand_ratio)
conv = []
if expand_ratio > 1:
conv = [
ops.Conv2d(in_channels=inp,
out_channels=hidden_dim,
kernel_size=1,
stride=1,
padding=0,
bias=False),
ops.BatchNorm2d(num_features=hidden_dim),
ops.Relu6(inplace=True)
]
conv = conv + [
ops.Conv2d(in_channels=hidden_dim,
out_channels=hidden_dim,
kernel_size=kernel,
stride=stride,
padding=kernel // 2,
groups=hidden_dim,
bias=False,
depthwise=True),
ops.BatchNorm2d(num_features=hidden_dim),
ops.Relu6(inplace=True),
ops.Conv2d(in_channels=hidden_dim,
out_channels=oup,
kernel_size=1,
stride=1,
padding=0,
bias=False),
ops.BatchNorm2d(num_features=oup)
]
self.models = Sequential(*conv)
def __init__(self,
inplanes,
planes,
stride=1,
dilation=1,
downsample=None,
style='pytorch',
with_cp=False):
"""Init BasicBlock."""
super(BasicBlock, self).__init__()
self.expansion = 1
self.norm1 = ops.BatchNorm2d(planes)
self.norm2 = ops.BatchNorm2d(planes)
self.conv1 = ops.Conv2d(inplanes,
planes,
3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False)
self.conv2 = ops.Conv2d(planes, planes, 3, padding=1, bias=False)
self.relu = ops.Relu(inplace=True)
if stride > 1 or downsample is not None:
conv_layer = ops.Conv2d(inplanes,
planes * self.expansion,
kernel_size=1,
stride=stride,
bias=False)
norm_layer = ops.BatchNorm2d(planes)
self.downsample = Sequential(conv_layer, norm_layer)
else:
self.downsample = None
self.inplanes = inplanes
self.planes = planes
self.stride = stride
self.dilation = dilation
self.style = style
assert not with_cp
def conv33_sep(in_channel, out_channel, stride):
"""Conv 3*3 sep."""
return Sequential(
conv33(in_channel, in_channel, stride, groups=in_channel),
conv11(in_channel, out_channel))
class SerialBackbone(Module):
"""Serial Net for spnas."""
def __init__(self,
code='111-2111-211111-211',
block='BottleneckBlock',
in_channels=64,
weight_file=None,
out_layers=None):
"""Init SerialBackbone."""
super(SerialBackbone, self).__init__()
self.inplanes = in_channels
self.planes = self.inplanes
self.weight_file = weight_file
self.channels = [3]
self.code = code.split('-')
self.block = ClassFactory.get_cls(ClassType.NETWORK, block)
self._make_stem_layer()
self.layers = Sequential() if out_layers == -1 else OutDictSequential()
self.make_cells()
@property
def out_channels(self):
"""Output Channel for Module."""
return self.layers.out_channels
def load_state_dict(self, state_dict=None, strict=None):
"""Load and freeze backbone state."""
if isinstance(self.layers, Sequential):
return super().load_state_dict(state_dict, strict or False)
state_dict = {
k.replace('backbone.', ''): v
for k, v in state_dict.items()
}
state_dict = {k.replace('body.', ''): v for k, v in state_dict.items()}
not_swap_keys = super().load_state_dict(state_dict, strict or False)
if not_swap_keys:
need_freeze_layers = [
name for name, parameter in self.named_parameters()
if name not in not_swap_keys
]
self.freeze(need_freeze_layers)
else:
self.freeze(['layers'])
def _make_stem_layer(self):
"""Make stem layer."""
self.conv1 = ops.Conv2d(3,
self.inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = ops.BatchNorm2d(self.inplanes)
self.relu = ops.Relu(inplace=True)
self.maxpool = ops.MaxPool2d(kernel_size=3, stride=2, padding=1)
def make_cells(self):
"""Make ResNet Cell."""
for i, code in enumerate(self.code):
layer, planes = self.make_layers(self.block,
self.inplanes,
self.planes,
code=code)
self.channels.append(planes)
self.inplanes = planes
self.layers.append(layer)
self.planes = self.planes * 2
def make_layers(self, block, inplanes, planes, code=None):
"""Make ResNet layers."""
strides = list(map(int, code))
layers = []
layers.append(block(inplanes, planes, stride=strides[0]))
inplanes = planes * block.expansion
for stride in strides[1:]:
layers.append(block(inplanes, planes, stride=stride))
inplanes = planes * block.expansion
return Sequential(*layers), inplanes
def create_op(opt_name, in_channel, out_channel, stride=1):
"""Create op."""
layer = OPS[opt_name](in_channel, out_channel, stride)
bn = ops.BatchNorm2d(out_channel)
return Sequential(layer, bn)