def __init__(self,
pretrained=True,
layers=[1, 2, 3, 4],
act_layer: Union[str, nn.Module] = Swish,
no_stride=False):
from timm.models.efficientnet import tf_efficientnet_b0_ns
if isinstance(act_layer, str):
act_layer = get_activation_block(act_layer)
encoder = tf_efficientnet_b0_ns(pretrained=pretrained,
features_only=True,
act_layer=act_layer,
drop_path_rate=0.05)
strides = [2, 4, 8, 16, 32]
if no_stride:
encoder.blocks[5][0].conv_dw.stride = (1, 1)
encoder.blocks[5][0].conv_dw.dilation = (2, 2)
encoder.blocks[3][0].conv_dw.stride = (1, 1)
encoder.blocks[3][0].conv_dw.dilation = (2, 2)
strides[3] = 8
strides[4] = 8
super().__init__([16, 24, 40, 112, 320], strides, layers)
self.encoder = encoder
def __init__(self,
pretrained=True,
layers=[1, 2, 3, 4],
act_layer: Union[str, nn.Module] = Swish):
from timm.models.efficientnet import mixnet_xl
if isinstance(act_layer, str):
act_layer = get_activation_block(act_layer)
encoder = mixnet_xl(pretrained=pretrained,
features_only=True,
act_layer=act_layer,
drop_path_rate=0.2)
super().__init__([40, 48, 64, 192, 320], [2, 4, 8, 16, 32], layers)
self.encoder = encoder
def __init__(
self,
input_channels: int = 3,
stack_level: int = 8,
depth: int = 4,
features: int = 256,
activation=ACT_RELU,
repeats=1,
pooling_block=nn.MaxPool2d,
):
super().__init__(
channels=[features] + [features] * stack_level,
strides=[4] + [4] * stack_level,
layers=list(range(0, stack_level + 1)),
)
self.stack_level = stack_level
self.depth_level = depth
self.num_features = features
act = get_activation_block(activation)
self.stem = HGStemBlock(input_channels, features, activation=act)
input_features = features
modules = []
for _ in range(stack_level):
modules.append(
HGBlock(
depth,
input_features,
features,
increase=0,
activation=act,
repeats=repeats,
pooling_block=pooling_block,
))
input_features = features
self.num_blocks = len(modules)
self.blocks = nn.ModuleList(modules)
self.features = nn.ModuleList([
HGFeaturesBlock(features, blocks=4, activation=act)
for _ in range(stack_level)
])
self.merge_features = nn.ModuleList([
nn.Conv2d(features, features, kernel_size=1)
for _ in range(stack_level - 1)
])