def get_backbone(self):
if self.backbone_type == 'MobileFaceNet':
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = MobileFaceNet(feat_dim, out_h, out_w)
elif self.backbone_type == 'ResNet':
depth = self.backbone_param[
'depth'] # depth of the ResNet, e.g. 50, 100, 152.
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
net_mode = self.backbone_param[
'net_mode'] # 'ir' for improved by resnt, 'ir_se' for SE-ResNet.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = Resnet(depth, drop_ratio, net_mode, feat_dim, out_h,
out_w)
else:
pass
return backbone
def get_backbone(self):
if self.backbone_type == 'MobileFaceNet':
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = MobileFaceNet(feat_dim, out_h, out_w)
elif self.backbone_type == 'ResNet':
depth = self.backbone_param[
'depth'] # depth of the ResNet, e.g. 50, 100, 152.
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
net_mode = self.backbone_param[
'net_mode'] # 'ir' for improved by resnt, 'ir_se' for SE-ResNet.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = Resnet(depth, drop_ratio, net_mode, feat_dim, out_h,
out_w)
elif self.backbone_type == 'EfficientNet':
width = self.backbone_param[
'width'] # width for EfficientNet, e.g. 1.0, 1.2, 1.4, ...
depth = self.backbone_param[
'depth'] # depth for EfficientNet, e.g. 1.0, 1.2, 1.4, ...
image_size = self.backbone_param[
'image_size'] # input image size, e.g. 112.
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
blocks_args, global_params = efficientnet(width_coefficient=width,
depth_coefficient=depth,
dropout_rate=drop_ratio,
image_size=image_size)
backbone = EfficientNet(out_h, out_w, feat_dim, blocks_args,
global_params)
elif self.backbone_type == 'HRNet':
config = {}
config['MODEL'] = self.backbone_param
backbone = HighResolutionNet(config)
elif self.backbone_type == 'GhostNet':
width = self.backbone_param['width']
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final feature
backbone = GhostNet(width, drop_ratio, feat_dim, out_h, out_w)
elif self.backbone_type == 'AttentionNet':
stage1_modules = self.backbone_param[
'stage1_modules'] # the number of attention modules in stage1.
stage2_modules = self.backbone_param[
'stage2_modules'] # the number of attention modules in stage2.
stage3_modules = self.backbone_param[
'stage3_modules'] # the number of attention modules in stage3.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = ResidualAttentionNet(stage1_modules, stage2_modules,
stage3_modules, feat_dim, out_h,
out_w)
elif self.backbone_type == 'TF-NAS':
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
backbone = TF_NAS_A(out_h, out_w, feat_dim, drop_ratio)
else:
pass
return backbone
def get_backbone(self):
if self.backbone_type == 'MobileFaceNet':
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = MobileFaceNet(feat_dim, out_h, out_w)
elif self.backbone_type == 'ResNet':
depth = self.backbone_param[
'depth'] # depth of the ResNet, e.g. 50, 100, 152.
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
net_mode = self.backbone_param[
'net_mode'] # 'ir' for improved by resnt, 'ir_se' for SE-ResNet.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = Resnet(depth, drop_ratio, net_mode, feat_dim, out_h,
out_w)
elif self.backbone_type == 'EfficientNet':
width = self.backbone_param[
'width'] # width for EfficientNet, e.g. 1.0, 1.2, 1.4, ...
depth = self.backbone_param[
'depth'] # depth for EfficientNet, e.g. 1.0, 1.2, 1.4, ...
image_size = self.backbone_param[
'image_size'] # input image size, e.g. 112.
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
blocks_args, global_params = efficientnet(width_coefficient=width,
depth_coefficient=depth,
dropout_rate=drop_ratio,
image_size=image_size)
backbone = EfficientNet(out_h, out_w, feat_dim, blocks_args,
global_params)
elif self.backbone_type == 'HRNet':
config = {}
config['MODEL'] = self.backbone_param
backbone = HighResolutionNet(config)
elif self.backbone_type == 'GhostNet':
width = self.backbone_param['width']
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final feature
backbone = GhostNet(width, drop_ratio, feat_dim, out_h, out_w)
elif self.backbone_type == 'AttentionNet':
stage1_modules = self.backbone_param[
'stage1_modules'] # the number of attention modules in stage1.
stage2_modules = self.backbone_param[
'stage2_modules'] # the number of attention modules in stage2.
stage3_modules = self.backbone_param[
'stage3_modules'] # the number of attention modules in stage3.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = ResidualAttentionNet(stage1_modules, stage2_modules,
stage3_modules, feat_dim, out_h,
out_w)
elif self.backbone_type == 'TF-NAS':
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
backbone = TF_NAS_A(out_h, out_w, feat_dim, drop_ratio)
elif self.backbone_type == 'ResNeSt':
depth = self.backbone_param[
'depth'] # depth of the ResNet, e.g. 50, 100, 152.
drop_ratio = self.backbone_param['drop_ratio'] # drop out ratio.
feat_dim = self.backbone_param[
'feat_dim'] # dimension of the output features, e.g. 512.
out_h = self.backbone_param[
'out_h'] # height of the feature map before the final features.
out_w = self.backbone_param[
'out_w'] # width of the feature map before the final features.
backbone = ResNeSt(depth, drop_ratio, feat_dim, out_h, out_w)
elif self.backbone_type == 'ReXNet':
input_ch = self.backbone_param['input_ch']
final_ch = self.backbone_param['final_ch']
width_mult = self.backbone_param['width_mult']
depth_mult = self.backbone_param['depth_mult']
use_se = True if self.backbone_param['use_se'] == 1 else False
se_ratio = self.backbone_param['se_ratio']
out_h = self.backbone_param['out_h']
out_w = self.backbone_param['out_w']
feat_dim = self.backbone_param['feat_dim']
dropout_ratio = self.backbone_param['dropout_ratio']
backbone = ReXNetV1(input_ch, final_ch, width_mult, depth_mult,
use_se, se_ratio, out_h, out_w, feat_dim,
dropout_ratio)
elif self.backbone_type == 'LightCNN':
depth = self.backbone_param['depth']
out_h = self.backbone_param['out_h']
out_w = self.backbone_param['out_w']
feat_dim = self.backbone_param['feat_dim']
drop_ratio = self.backbone_param['dropout_ratio']
backbone = LightCNN(depth, drop_ratio, out_h, out_w, feat_dim)
elif self.backbone_type == 'RepVGG':
blocks1 = self.backbone_param['blocks1']
blocks2 = self.backbone_param['blocks2']
blocks3 = self.backbone_param['blocks3']
blocks4 = self.backbone_param['blocks4']
width1 = self.backbone_param['width1']
width2 = self.backbone_param['width2']
width3 = self.backbone_param['width3']
width4 = self.backbone_param['width4']
out_h = self.backbone_param['out_h']
out_w = self.backbone_param['out_w']
feat_dim = self.backbone_param['feat_dim']
backbone = RepVGG([blocks1, blocks2, blocks3, blocks4],
[width1, width2, width3, width4], feat_dim,
out_h, out_w)
elif self.backbone_type == 'SwinTransformer':
img_size = self.backbone_param['img_size']
patch_size = self.backbone_param['patch_size']
in_chans = self.backbone_param['in_chans']
embed_dim = self.backbone_param['embed_dim']
depths = self.backbone_param['depths']
num_heads = self.backbone_param['num_heads']
window_size = self.backbone_param['window_size']
mlp_ratio = self.backbone_param['mlp_ratio']
drop_rate = self.backbone_param['drop_rate']
drop_path_rate = self.backbone_param['drop_path_rate']
backbone = SwinTransformer(img_size=img_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
depths=depths,
num_heads=num_heads,
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=True,
qk_scale=None,
drop_rate=drop_rate,
drop_path_rate=drop_path_rate,
ape=False,
patch_norm=True,
use_checkpoint=False)
else:
pass
return backbone