def ela_tf_efficientnet_b2_ns(num_classes=4, pretrained=True, dropout=0):
encoder = efficientnet.tf_efficientnet_b2_ns(in_chans=9,
pretrained=False,
drop_path_rate=0.1)
del encoder.classifier
if pretrained:
donor = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained)
transfer_weights(encoder, donor.state_dict())
return TimmElaOnlyRichModel(encoder,
num_classes=num_classes,
dropout=dropout)
def rgb_res_tf_efficientnet_b2_ns(num_classes=4, pretrained=True, dropout=0):
encoder = efficientnet.tf_efficientnet_b2_ns(in_chans=6,
pretrained=False,
drop_path_rate=0.1)
del encoder.classifier
if pretrained:
donor = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained)
transfer_weights(encoder, donor.state_dict())
return ImageAndResidualModel(
encoder,
num_classes=num_classes,
dropout=dropout,
mean=encoder.default_cfg["mean"],
std=encoder.default_cfg["std"],
)
def rgb_res_sms_v2_tf_efficientnet_b2_ns(num_classes=4,
pretrained=True,
dropout=0):
rgb_encoder = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained,
drop_path_rate=0.1)
del rgb_encoder.classifier
res_encoder = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained,
drop_path_rate=0.1)
del res_encoder.classifier
return SiameseImageAndResidualModelV2(
rgb_encoder,
res_encoder,
num_classes=num_classes,
dropout=dropout,
mean=rgb_encoder.default_cfg["mean"],
std=rgb_encoder.default_cfg["std"],
)
def rgb_tf_efficientnet_b2_ns(num_classes=4,
pretrained=True,
dropout=0.1,
need_embedding=False):
encoder = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained,
drop_path_rate=0.1)
del encoder.classifier
return TimmRgbModel(encoder,
num_classes=num_classes,
dropout=dropout,
need_embedding=need_embedding)
def rgb_tf_efficientnet_b2_ns_avgmax(num_classes=4,
pretrained=True,
dropout=0):
encoder = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained)
del encoder.classifier
return TimmRgbModelAvgMax(
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_b2_ns
encoder = tf_efficientnet_b2_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__([16, 24, 48, 120, 352], strides, layers)
self.encoder = encoder
def rgb_qf_tf_efficientnet_b2_ns(num_classes=4, pretrained=True, dropout=0):
encoder = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained)
del encoder.classifier
return ImageAndQFModel(encoder, num_classes=num_classes, dropout=dropout)
def res_tf_efficientnet_b2_ns(num_classes=4, pretrained=True, dropout=0):
encoder = efficientnet.tf_efficientnet_b2_ns(pretrained=pretrained,
drop_path_rate=0.1)
del encoder.classifier
return ResidualOnlyModel(encoder, num_classes=num_classes, dropout=dropout)