def get_primitives(cls, **primitive_kwargs) -> [CNNPrimitive]:
act = dict(act_fun='relu', order='act_w_bn')
df = dict(act_inplace=False, bn_affine=True, use_bn=True)
return [
CNNPrimitive(cls=SepConvLayer,
kwargs=dict(k_size=3, dilation=1, **act, **df),
stacked=2),
CNNPrimitive(cls=SepConvLayer,
kwargs=dict(k_size=5, dilation=1, **act, **df),
stacked=2),
CNNPrimitive(cls=SepConvLayer,
kwargs=dict(k_size=3, dilation=2, **act, **df)),
CNNPrimitive(cls=SepConvLayer,
kwargs=dict(k_size=5, dilation=2, **act, **df)),
CNNPrimitive(PoolingLayer,
kwargs=dict(k_size=3,
pool_type='max',
act_fun=None,
order='w_bn',
**df)),
CNNPrimitive(PoolingLayer,
kwargs=dict(k_size=3,
pool_type='avg',
act_fun=None,
order='w_bn',
**df)),
StrideChoiceCNNPrimitive([
CNNPrimitive(cls=SkipLayer, kwargs=dict()),
CNNPrimitive(cls=FactorizedReductionLayer,
kwargs=dict(**act, **df))
]),
CNNPrimitive(cls=ZeroLayer, kwargs=dict()),
]
def get_primitives(cls, **primitive_kwargs) -> [CNNPrimitive]:
df = dict(dilation=1,
bn_affine=True,
act_fun='relu',
act_inplace=False,
expansion=1.0,
att_dict=None)
return [
CNNPrimitive(cls=ShuffleNetV2Layer, kwargs=dict(k_size=3, **df)),
CNNPrimitive(cls=ShuffleNetV2Layer, kwargs=dict(k_size=5, **df)),
CNNPrimitive(cls=ShuffleNetV2Layer, kwargs=dict(k_size=7, **df)),
CNNPrimitive(cls=ShuffleNetV2XceptionLayer,
kwargs=dict(k_size=3, **df)),
]
def get_primitives(cls, stride=1, **primitive_kwargs) -> [CNNPrimitive]:
primitives = MobileNetV2Primitives.get_primitives(stride=stride,
**primitive_kwargs)
if stride == 1:
primitives.append(
CNNPrimitive(cls=LinearTransformerLayer, kwargs=dict())),
return primitives
def get_primitives(cls, stride=1, **primitive_kwargs) -> [CNNPrimitive]:
df = dict(dilation=1,
act_inplace=True,
bn_affine=True,
act_fun='swish')
df['att_dict'] = dict(att_cls='EfficientChannelAttentionModule',
use_c_substitute=True)
primitives = [
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=3, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=5, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=7, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=3, expansion=6.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=5, expansion=6.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=7, expansion=6.0, **df)),
]
if stride == 1:
primitives.append(
CNNPrimitive(cls=LinearTransformerLayer, kwargs=dict()))
return primitives
def get_primitives(cls, stride=1, **primitive_kwargs) -> [CNNPrimitive]:
df = dict(dilation=1,
act_inplace=True,
bn_affine=True,
act_fun='swish')
df['att_dict'] = dict(att_cls='SqueezeExcitationChannelModule',
use_c_substitute=True,
c_mul=0.25,
squeeze_act='swish',
excite_act='sigmoid',
squeeze_bias=True,
excite_bias=True,
squeeze_bn=False)
primitives = [
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=3, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=5, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=7, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=3, expansion=6.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=5, expansion=6.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=7, expansion=6.0, **df)),
]
if stride == 1:
primitives.append(
CNNPrimitive(cls=LinearTransformerLayer, kwargs=dict()))
return primitives
def get_primitives(cls, **primitive_kwargs) -> [CNNPrimitive]:
act = dict(act_fun='relu',
order='w_bn_act',
act_inplace=False,
bn_affine=False,
use_bn=True)
return [
CNNPrimitive(cls=ConvLayer,
kwargs=dict(k_size=3, dilation=1, **act)),
CNNPrimitive(cls=ConvLayer,
kwargs=dict(k_size=1, dilation=1, **act)),
CNNPrimitive(PoolingLayer,
kwargs=dict(k_size=3,
pool_type='max',
act_fun=None,
order='w',
use_bn=False)),
]
def get_primitives(cls, **primitive_kwargs) -> [CNNPrimitive]:
act = dict(act_fun='relu', order='act_w_bn')
df = dict(act_inplace=False, bn_affine=True, use_bn=True)
dfnb = df.copy()
dfnb['use_bn'] = False
return [
CNNPrimitive(cls=SepConvLayer, kwargs=dict(k_size=3, dilation=1, **act, **df), stacked=2),
CNNPrimitive(cls=SepConvLayer, kwargs=dict(k_size=5, dilation=1, **act, **df), stacked=2),
CNNPrimitive(cls=SepConvLayer, kwargs=dict(k_size=3, dilation=2, **act, **df)),
CNNPrimitive(cls=SepConvLayer, kwargs=dict(k_size=5, dilation=2, **act, **df)),
DifferentConfigPrimitive(
CNNPrimitive(PoolingLayer, kwargs=dict(k_size=3, pool_type='max', act_fun=None, order='w_bn', **df)),
CNNPrimitive(PoolingLayer, kwargs=dict(k_size=3, pool_type='max', act_fun=None, order='w', **dfnb))),
DifferentConfigPrimitive(
CNNPrimitive(PoolingLayer, kwargs=dict(k_size=3, pool_type='avg', act_fun=None, order='w_bn', **df)),
CNNPrimitive(PoolingLayer, kwargs=dict(k_size=3, pool_type='avg', act_fun=None, order='w', **dfnb))),
StrideChoiceCNNPrimitive([
CNNPrimitive(cls=LinearTransformerLayer, kwargs=dict()),
CNNPrimitive(cls=FactorizedReductionLayer, kwargs=dict(**act, **df))
]),
]
def get_primitives(cls, **primitive_kwargs) -> [CNNPrimitive]:
act = dict(act_fun='relu',
order='act_w_bn',
act_inplace=False,
bn_affine=False,
use_bn=True)
return [
CNNPrimitive(cls=ZeroLayer, kwargs=dict()),
StrideChoiceCNNPrimitive([
CNNPrimitive(cls=LinearTransformerLayer, kwargs=dict()),
CNNPrimitive(cls=FactorizedReductionLayer, kwargs=dict(**act))
]),
CNNPrimitive(cls=ConvLayer,
kwargs=dict(k_size=1, dilation=1, **act)),
CNNPrimitive(cls=ConvLayer,
kwargs=dict(k_size=3, dilation=1, **act)),
CNNPrimitive(PoolingConvLayer,
kwargs=dict(k_size=3,
pool_type='avg',
act_fun=None,
order='w',
bn_affine=False,
use_bn=False,
bias=False)),
]
def get_primitives(cls, stride=1, **primitive_kwargs) -> [CNNPrimitive]:
df = dict(dilation=1,
act_inplace=True,
bn_affine=True,
act_fun='hswish')
att_dict = dict(att_cls='EfficientChannelAttentionModule',
use_c_substitute=False,
k_size=-1,
gamma=2,
b=1,
excite_act='sigmoid')
primitives = [
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(expansion=3,
att_dict=att_dict,
k_size=(3, ),
**df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(expansion=3,
att_dict=att_dict,
k_size=(3, 5),
**df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(expansion=3,
att_dict=att_dict,
k_size=(3, 5, 7),
**df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(expansion=6,
att_dict=att_dict,
k_size=(3, ),
**df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(expansion=6,
att_dict=att_dict,
k_size=(3, 5),
**df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(expansion=6,
att_dict=att_dict,
k_size=(3, 5, 7),
**df)),
]
if stride == 1:
primitives.append(
CNNPrimitive(cls=LinearTransformerLayer, kwargs=dict()))
return primitives
def get_primitives(cls, stride=1, **primitive_kwargs) -> [CNNPrimitive]:
df = dict(dilation=1,
act_inplace=True,
bn_affine=True,
act_fun='relu6',
att_dict=None,
stride=stride)
return [
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=3, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=5, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=7, expansion=3.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=3, expansion=6.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=5, expansion=6.0, **df)),
CNNPrimitive(cls=MobileInvertedConvLayer,
kwargs=dict(k_size=7, expansion=6.0, **df)),
]