def build_resnet_fpn_p3p7_backbone(cfg): #构建论文中backbone部分以及特征金字塔的P3-P7部分网络
body = resnet.ResNet(cfg) #构建body backbone中的C3-C5
#获取fpn所需要的channels参数
in_channels_stage2 = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS #256
out_channels = cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS #256*4
#cfg.MODEL.RETINANET.USE_C5 = True
# in_channels_p6p7 = 256*8
in_channels_p6p7 = in_channels_stage2 * 8 if cfg.MODEL.RETINANET.USE_C5 \
else out_channels
#构建fpn部分网络
fpn = fpn_module.FPN(
in_channels_list=[
0, # 因为从C3起才有P3，所以stage2跳过，设置为0
in_channels_stage2 * 2, #c3
in_channels_stage2 * 4, #c4
in_channels_stage2 * 8, #c5
],
out_channels=out_channels, #256*4
conv_block=
conv_with_kaiming_uniform( # 这个conv如果stride=1的话就不变size,返回的是nn.Conv2d
cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU #false false
),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7,
out_channels), #256 256
)
# 通过有序字典将body和fpn送入nn.Sequential构造模型
model = nn.Sequential(OrderedDict([("body", body),
("fpn", fpn)])) # 写成一个,body的输出作为fpn的输入
# 这个是为了之后有用，再赋一次值
model.out_channels = out_channels #256*4
return model
def __init__(self, cfg):
super(GrowR50FPN, self).__init__()
self.stem_conv1 = nn.Conv2d(3,
64,
kernel_size=(7, 7),
stride=(2, 2),
padding=(3, 3),
bias=False)
self.stem_bn1 = nn.BatchNorm2d(64)
self.stage_1_1 = GrowR50_BottleNeck(64, 256, 256, down_stride=2)
# self.stage_1_2 = GrowR50_BottleNeck(256, 256, 256)
# self.stage_2_1 = GrowR50_BottleNeck(256, 256, 512, down_stride=2)
self.stage_2_1 = GrowR50_BottleNeck(256, 512, 512, down_stride=2)
self.stage_3_1 = GrowR50_BottleNeck(512, 1024, 1024, down_stride=2)
self.stage_4_1 = GrowR50_BottleNeck(1024, 2048, 2048, down_stride=2)
self.fpn = fpn_module.FPN(
in_channels_list=[0, 512, 1024, 2048],
out_channels=256,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels=256,
out_channels=256),
)
def build_resnet_fpn_p3p7_backbone(cfg):
body = resnet.ResNet(cfg)
# 获取 fpn 所需的channels参数
in_channels_stage2 = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS # 256
out_channels = cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS # 256
in_channels_p6p7 = in_channels_stage2 * 8 if cfg.MODEL.RETINANET.USE_C5 \
else out_channels
fpn = fpn_module.FPN(
in_channels_list=[
0,
in_channels_stage2 * 2, # 因为从C3起才有P3，所以stage2跳过，设置为0
in_channels_stage2 * 4,
in_channels_stage2 * 8,
],
out_channels=out_channels, # cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS
conv_block=
conv_with_kaiming_uniform( # 这个conv如果stride=1的话就不变size,返回的是一个函数
cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU # eg: False, False
),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7,
out_channels), # eg: 256, 256
)
# 通过有序字典将body和fpn送入nn.Sequential构造模型
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
# 这个是为了之后有用，再赋一次值
model.out_channels = out_channels
return model
def build_mnv2_fpn_backbone(cfg, basemodel=None):
body = mobilenet.MobileNetV2(cfg) if basemodel is None else basemodel
in_channels_stage2 = body.return_features_num_channels
out_channels = cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS
fpn = fpn_module.FPN(
in_channels_list=[
0,
in_channels_stage2[1],
in_channels_stage2[2],
in_channels_stage2[3],
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(out_channels, out_channels),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_resnet_fpn_backbone(cfg, basemodel=None):
body = resnet.ResNet(cfg) if basemodel is None else basemodel
in_channels_stage2 = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS
out_channels = cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS
fpn = fpn_module.FPN(
in_channels_list=[
in_channels_stage2,
in_channels_stage2 * 2,
in_channels_stage2 * 4,
in_channels_stage2 * 8,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelMaxPool(),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_efficientnet_b3_fpn_p3p7_backbone(cfg):
body = efficientnet_model.EfficientNet.efficientnet_b3_fpn()
out_channels = 128
in_channels_p6p7 = 128
fpn = fpn_module.FPN(
in_channels_list=[
0,
48,
136,
384,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7, out_channels),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_dla34_fpn_p3p7_backbone(cfg):
body = dla_lite.dla34_fpn()
out_channels = 128
in_channels_p6p7 = 128
fpn = fpn_module.FPN(
in_channels_list=[
0,
128,
256,
512,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7, out_channels),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_fbnet_fpn_p3p7_backbone(cfg):
body = fbnet.add_conv_body_fpn(cfg)
out_channels = 64
in_channels_p6p7 = 64
fpn = fpn_module.FPN(
in_channels_list=[
0,
64,
88,
104,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7, out_channels),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_vovnet_fpn_backbone(cfg):
body = vovnet.VoVNet(cfg)
in_channels_stage = cfg.MODEL.VOVNET.OUT_CHANNELS
out_channels = cfg.MODEL.VOVNET.BACKBONE_OUT_CHANNELS
fpn = fpn_module.FPN(
in_channels_list=[
in_channels_stage,
in_channels_stage * 2,
in_channels_stage * 3,
in_channels_stage * 4,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelMaxPool(),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_resnet_fpn_p3p7_backbone(cfg):
body = resnet.ResNet(cfg)
in_channels_stage2 = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS
out_channels = cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS
in_channels_p6p7 = in_channels_stage2 * 8 if cfg.MODEL.RETINANET.USE_C5 \
else out_channels
fpn = fpn_module.FPN(
in_channels_list=[
0,
in_channels_stage2 * 2,
in_channels_stage2 * 4,
in_channels_stage2 * 8,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7, out_channels),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def build_vgg_fpn_backbone(cfg):
body = VGG(depth=16, with_last_pool=True, frozen_stages=2)
in_channels_stage2 = 128 # default: cfg.MODEL.RESNETS.RES2_OUT_CHANNELS (256)
out_channels = 256 # default: cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS (256)
in_channels_p6p7 = in_channels_stage2 * 4 if cfg.MODEL.RETINANET.USE_C5 \
else out_channels
fpn = fpn_module.FPN(
in_channels_list=[
0,
0,
in_channels_stage2 * 2,
in_channels_stage2 * 4,
in_channels_stage2 * 4, # in_channels_stage2 * 8
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7, out_channels),
)
model = nn.Sequential(OrderedDict([("body", body), ("fpn", fpn)]))
model.out_channels = out_channels
return model
def __init__(self, cfg):
super(GeneralizedRCNN, self).__init__()
self.extra = True
self.backbone = build_backbone(cfg)
self.cls_backbone = self.backbone[0]
self.reg_backbone = self.backbone[1]
if self.extra == False:
self.rpn = build_rpn(cfg, self.cls_backbone.out_channels,
self.extra)
if self.extra == True:
self.rpn, self.extra_rpn = build_rpn(
cfg, self.cls_backbone.out_channels, self.extra)
self.roi_heads = build_roi_heads(cfg, self.cls_backbone.out_channels)
self.cfg = cfg
self.deform_conv_cls = DFConv2d(256,
256,
with_modulated_dcn=False,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
deformable_groups=1,
bias=False)
self.deform_conv_reg = DFConv2d(256,
256,
with_modulated_dcn=False,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
deformable_groups=1,
bias=False)
self.cls_head = build_cls_head(True, False, False)
conv_wo_relu = conv_with_kaiming_uniform(False, True)
self.cls_weight_conv = conv_wo_relu(256, 1, 1)
self.reg_weight_conv = conv_wo_relu(256, 1, 1)
def build_resnet_fpn_p3p7_backbone(cfg):
in_channels_stage2 = cfg.MODEL.RESNETS.RES2_OUT_CHANNELS
cls_in_channels_stage2 = cfg.MODEL.RESNETS.CLS_RES2_OUT_CHANNELS
out_channels = cfg.MODEL.RESNETS.BACKBONE_OUT_CHANNELS
in_channels_p6p7 = in_channels_stage2 * 8 if cfg.MODEL.RETINANET.USE_C5 \
else out_channels
cls_in_channels_p6p7 = cls_in_channels_stage2 * 8 if cfg.MODEL.RETINANET.USE_C5 \
else out_channels
cls_resnet = resnet.ResNet(cfg, 'CLS')
cls_fpn = fpn_module.FPN(
in_channels_list=[
0,
cls_in_channels_stage2 * 2,
cls_in_channels_stage2 * 4,
cls_in_channels_stage2 * 8,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(cls_in_channels_p6p7,
out_channels),
last_backbone=True,
last_fpn=False)
reg_resnet = resnet.ResNet(cfg, 'REG')
reg_fpn = fpn_module.FPN(
in_channels_list=[
0,
in_channels_stage2 * 2,
in_channels_stage2 * 4,
in_channels_stage2 * 8,
],
out_channels=out_channels,
conv_block=conv_with_kaiming_uniform(cfg.MODEL.FPN.USE_GN,
cfg.MODEL.FPN.USE_RELU),
top_blocks=fpn_module.LastLevelP6P7(in_channels_p6p7, out_channels),
last_backbone=False,
last_fpn=False)
#if cfg.MODEL.BACKBONE.REG_CONV_BODY == "HRNET-W18":
# width = 18
#elif cfg.MODEL.BACKBONE.REG_CONV_BODY == "HRNET-W32":
# width = 32
#elif cfg.MODEL.BACKBONE.REG_CONV_BODY == "HRNET-W40":
# width = 40
#else:
# raise NotImplementedError
#hrnet_args = dict(
# stage1=dict(
# num_modules=1,
# num_branches=1,
# block='BOTTLENECK',
# num_blocks=(4,),
# num_channels=(64,),
# fuse_method='SUM'),
# stage2=dict(
# num_modules=1,
# num_branches=2,
# block='BASIC',
# num_blocks=(4, 4),
# num_channels=(width, width * 2),
# fuse_method='SUM'),
# stage3=dict(
# num_modules=4,
# num_branches=3,
# block='BASIC',
# num_blocks=(4, 4, 4),
# num_channels=(width, width * 2, width * 4),
# fuse_method='SUM'),
# stage4=dict(
# num_modules=3,
# num_branches=4,
# block='BASIC',
# num_blocks=(4, 4, 4, 4),
# num_channels=(width, width * 2, width * 4, width * 8),
# fuse_method='SUM')
#)
#fpn_in_channels = [width, width * 2, width * 4, width * 8]
#hrnet.BatchNorm2d = nn.SyncBatchNorm if cfg.MODEL.SYNCBN else nn.BatchNorm2d
#reg_hrnet = hrnet.HighResolutionNet(extra=hrnet_args)
#reg_fpn = getattr(hrfpn_module, cfg.MODEL.HRNET.FPN.TYPE)(
# in_channels=fpn_in_channels,
# out_channels=cfg.MODEL.HRNET.FPN.OUT_CHANNEL,
# conv_stride=cfg.MODEL.HRNET.FPN.CONV_STRIDE,
# num_level=len(cfg.MODEL.FCOS.FPN_STRIDES),
#)
cls_backbone = nn.Sequential(
OrderedDict([("cls_backbone", cls_resnet), ("cls_fpn", cls_fpn)]))
cls_backbone.out_channels = out_channels
reg_backbone = nn.Sequential(
OrderedDict([("reg_backbone", reg_resnet), ("reg_fpn", reg_fpn)]))
#reg_backbone = nn.Sequential(OrderedDict([("reg_backbone", reg_hrnet), ("reg_fpn", reg_fpn)]))
reg_backbone.out_channels = out_channels
return cls_backbone, reg_backbone
