def prepare_model():
# input image channels: BGR--3, gray--1
config_dict['num_input_channels'] = 3
classifiation_loss = FocalLoss(use_sigmoid=True,
gamma=2.0,
alpha=0.25,
reduction='mean',
loss_weight=1.0)
regression_loss = torch.nn.SmoothL1Loss(reduction='mean')
# number of classes
config_dict['num_classes'] = 1
config_dict[
'backbone_init_param_file_path'] = None # if no pretrained weights, set to None
lfd_backbone = LFDResNet(
block_mode='fast', # affect block type
stem_mode='fast', # affect stem type
body_mode='fast', # affect body architecture
input_channels=config_dict['num_input_channels'],
stem_channels=64,
body_architecture=None,
body_channels=None,
out_indices=((0, 3), (1, 1), (2, 1), (3, 0), (4, 0)),
frozen_stages=-1,
activation_cfg=dict(type='ReLU', inplace=True),
norm_cfg=dict(type='BatchNorm2d'),
init_with_weight_file=config_dict['backbone_init_param_file_path'],
norm_eval=False)
lfd_neck = SimpleNeck(
num_neck_channels=128,
num_input_channels_list=lfd_backbone.num_output_channels_list,
num_input_strides_list=lfd_backbone.num_output_strides_list,
norm_cfg=dict(type='BatchNorm2d'),
activation_cfg=dict(type='ReLU', inplace=True))
lfd_head = LFDHead(
num_classes=config_dict['num_classes'],
num_heads=len(lfd_neck.num_output_strides_list),
num_input_channels=128,
num_head_channels=128,
num_conv_layers=2,
activation_cfg=dict(type='ReLU', inplace=True),
norm_cfg=dict(type='BatchNorm2d'),
share_head_flag=False,
merge_path_flag=False,
classification_loss_type=type(classifiation_loss).__name__)
config_dict['model'] = LFD(
backbone=lfd_backbone,
neck=lfd_neck,
head=lfd_head,
num_classes=config_dict['num_classes'],
regression_ranges=((0, 16), (16, 32), (32, 64), (64, 128), (128, 256)),
gray_range_factors=(0.9, 1.1),
point_strides=lfd_neck.num_output_strides_list,
classification_loss_func=classifiation_loss,
regression_loss_func=regression_loss,
distance_to_bbox_mode='exp',
classification_threshold=0.05,
nms_threshold=0.5,
pre_nms_bbox_limit=1000,
post_nms_bbox_limit=100,
)
# init param weights file
# when set, the executor will init the whole net using this file
config_dict['weight_path'] = None
# resume training path
# when set, the 'weight_path' will be ignored. The executor will init the whole net and training parameters using this file
config_dict['resume_path'] = None
# evaluator
# the evaluator should match the dataset
config_dict['evaluator'] = None
def prepare_model():
# input image channels: BGR--3, gray--1
config_dict['num_input_channels'] = 3
classification_loss = CrossEntropyLoss(reduction='mean', loss_weight=1.0)
regression_loss = IoULoss(eps=1e-6, reduction='mean', loss_weight=1.0)
# number of classes
config_dict['num_classes'] = 1
config_dict[
'backbone_init_param_file_path'] = None # if no pretrained weights, set to None
lfd_backbone = LFDResNet(
block_mode='faster', # affect block type
stem_mode='faster', # affect stem type
body_mode=None, # affect body architecture
input_channels=config_dict['num_input_channels'],
stem_channels=64,
body_architecture=[4, 2, 2, 3],
body_channels=[64, 64, 64, 128],
out_indices=((0, 3), (1, 1), (2, 1), (3, 0), (3, 2)),
frozen_stages=-1,
activation_cfg=dict(type='ReLU', inplace=True),
norm_cfg=dict(type='BatchNorm2d'),
init_with_weight_file=config_dict['backbone_init_param_file_path'],
norm_eval=False)
lfd_neck = SimpleNeck(
num_neck_channels=128,
num_input_channels_list=lfd_backbone.num_output_channels_list,
num_input_strides_list=lfd_backbone.num_output_strides_list,
norm_cfg=dict(type='BatchNorm2d'),
activation_cfg=dict(type='ReLU', inplace=True))
lfd_head = LFDHead(
num_classes=config_dict['num_classes'],
num_heads=len(lfd_neck.num_output_strides_list),
num_input_channels=128,
num_head_channels=128,
num_conv_layers=2,
activation_cfg=dict(type='ReLU', inplace=True),
norm_cfg=dict(type='GroupNorm', num_groups=16),
share_head_flag=True,
merge_path_flag=True,
classification_loss_type=type(classification_loss).__name__,
regression_loss_type=type(regression_loss).__name__)
config_dict['detection_scales'] = ((4, 20), (20, 40), (40, 80), (80, 160),
(160, 320))
config_dict['model'] = LFD(
backbone=lfd_backbone,
neck=lfd_neck,
head=lfd_head,
num_classes=config_dict['num_classes'],
regression_ranges=config_dict['detection_scales'],
gray_range_factors=(0.9, 1.1),
point_strides=lfd_neck.num_output_strides_list,
classification_loss_func=classification_loss,
regression_loss_func=regression_loss,
distance_to_bbox_mode='sigmoid')
# init param weights file
# when set, the executor will init the whole net using this file
config_dict['weight_path'] = None
# resume training path
# when set, the 'weight_path' will be ignored. The executor will init the whole net and training parameters using this file
config_dict['resume_path'] = None
# evaluator
# the evaluator should match the dataset
config_dict['evaluator'] = None
def prepare_model():
# input image channels: BGR--3, gray--1
config_dict['num_input_channels'] = 3
classification_loss = QualityFocalLoss(use_sigmoid=True,
beta=2.0,
reduction='mean',
loss_weight=2.0)
regression_loss = IoULoss(eps=1e-6, reduction='mean', loss_weight=1.0)
# number of classes
config_dict['num_classes'] = 1
config_dict[
'backbone_init_param_file_path'] = './pre-trained_backbones/lfd_resnet_TL_L_imagenet_20210707_234720/epoch_60.pth'
lfd_backbone = LFDResNet(
block_mode='faster', # affect block type
stem_mode='fast', # affect stem type
body_mode=None, # affect body architecture
input_channels=config_dict['num_input_channels'],
stem_channels=64,
body_architecture=[5, 3, 2, 2, 2],
body_channels=[64, 64, 128, 128, 128],
out_indices=((0, 4), (1, 2), (2, 1), (3, 1), (4, 1)),
frozen_stages=-1,
activation_cfg=dict(type='ReLU', inplace=True),
norm_cfg=dict(type='BatchNorm2d'),
init_with_weight_file=config_dict['backbone_init_param_file_path'],
norm_eval=False)
lfd_neck = SimpleNeck(
num_neck_channels=128,
num_input_channels_list=lfd_backbone.num_output_channels_list,
num_input_strides_list=lfd_backbone.num_output_strides_list,
norm_cfg=dict(type='BatchNorm2d'),
activation_cfg=dict(type='ReLU', inplace=True))
lfd_head = LFDHead(
num_classes=config_dict['num_classes'],
num_heads=len(lfd_neck.num_output_strides_list),
num_input_channels=128,
num_head_channels=128,
num_conv_layers=2,
activation_cfg=dict(type='ReLU', inplace=True),
norm_cfg=None, # dict(type='GroupNorm', num_groups=16),
share_head_flag=True,
merge_path_flag=True,
classification_loss_type=type(classification_loss).__name__,
regression_loss_type=type(regression_loss).__name__)
config_dict['detection_scales'] = ((4, 32), (32, 64), (64, 128),
(128, 256), (256, 512))
config_dict['model'] = LFD(
backbone=lfd_backbone,
neck=lfd_neck,
head=lfd_head,
num_classes=config_dict['num_classes'],
regression_ranges=config_dict['detection_scales'],
gray_range_factors=(0.9, 1.1),
range_assign_mode='dist',
point_strides=lfd_neck.num_output_strides_list,
classification_loss_func=classification_loss,
regression_loss_func=regression_loss,
distance_to_bbox_mode='sigmoid',
enable_regression_weight=False,
enable_classification_weight=False)
# init param weights file
# when set, the executor will init the whole net using this file
config_dict['weight_path'] = None
# resume training path
# when set, the 'weight_path' will be ignored. The executor will init the whole net and training parameters using this file
config_dict['resume_path'] = None
