def rgb_tf_efficientnet_b3_ns(num_classes=4,
pretrained=True,
dropout=0.1,
need_embedding=False):
encoder = efficientnet.tf_efficientnet_b3_ns(pretrained=pretrained,
drop_path_rate=0.2)
del encoder.classifier
return TimmRgbModel(encoder,
num_classes=num_classes,
dropout=dropout,
need_embedding=need_embedding)
def hpf_b3_fixed_gap(num_classes, dropout=0, pretrained=False):
from timm.models import efficientnet
encoder = efficientnet.tf_efficientnet_b3_ns(pretrained=True, drop_path_rate=0.1)
encoder.conv_stem = nn.Sequential(HPF3(trainable_hpf=False, stride=2), nn.Conv2d(30, 40, kernel_size=1))
del encoder.classifier
return HPFNetGAP(
encoder,
num_classes=num_classes,
dropout=dropout,
mean=encoder.default_cfg["mean"],
std=encoder.default_cfg["std"],
)
def __init__(self, pretrained=True, layers=[1, 2, 3, 4], act_layer=Swish, no_stride=False):
from timm.models.efficientnet import tf_efficientnet_b3_ns
encoder = tf_efficientnet_b3_ns(
pretrained=pretrained, features_only=True, act_layer=act_layer, drop_path_rate=0.1
)
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__([24, 32, 48, 136, 384], strides, layers)
self.encoder = encoder