def __init__(self, cfg, pretrain=False):
super().__init__()
self.backbone = None
self.backbone_layers = None
input_channel = 720
self.aspp = ASPP(dim_in=input_channel,
dim_out=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
bn_mom=cfg.TRAIN_BN_MOM)
self.dropout1 = nn.Dropout(0.5)
self.upsample4 = nn.UpsamplingBilinear2d(scale_factor=4)
self.cls_conv = nn.Sequential(
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_ASPP_OUTDIM,
kernel_size=1,
stride=1,
padding=0),
SynchronizedBatchNorm2d(cfg.MODEL_ASPP_OUTDIM, momentum=cfg.TRAIN_BN_MOM),
nn.ReLU(inplace=False),
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_NUM_CLASSES,
kernel_size=1,
stride=1,
padding=0)
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, SynchronizedBatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
self.backbone = get_hrnet_basebone(cfg.MODEL_NUM_CLASSES, pretrain=pretrain)
def __init__(self, cfg, pretrain=False):
from model.deeplabv3plus.deeplabv3plusEnc import Encoding, Mean
super().__init__()
self.backbone = None
self.backbone_layers = None
input_channel = 720
cfg.MODEL_ASPP_OUTDIM = 512
se_channel = 48
self.pspaspp = PSPASPP(in_channels=input_channel,
out_channels=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
norm_layer=SynchronizedBatchNorm2d,
norm_kwargs={'momentum': cfg.TRAIN_BN_MOM})
self.dropout1 = nn.Dropout(0.5)
self.se_loss = nn.Sequential(
# nn.AvgPool2d(kernel_size=7, stride=4, padding=3),
nn.Conv2d(input_channel, se_channel, 1, bias=False),
SynchronizedBatchNorm2d(se_channel, momentum=cfg.TRAIN_BN_MOM),
nn.ReLU(True),
Encoding(D=se_channel, K=32),
nn.BatchNorm1d(32),
nn.ReLU(True),
Mean(dim=1),
nn.Linear(se_channel, cfg.MODEL_NUM_CLASSES)
)
self.cls_conv = nn.Sequential(
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_ASPP_OUTDIM,
kernel_size=1,
stride=1,
padding=0),
SynchronizedBatchNorm2d(cfg.MODEL_ASPP_OUTDIM, momentum=cfg.TRAIN_BN_MOM),
nn.ReLU(inplace=False),
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_NUM_CLASSES,
kernel_size=1,
stride=1,
padding=0)
)
self.upsample4 = nn.UpsamplingBilinear2d(scale_factor=4)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, SynchronizedBatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
self.backbone = get_hrnet_basebone(cfg.MODEL_NUM_CLASSES, pretrain=pretrain, combine=False)
def __init__(self, cfg, pretrain=False):
super().__init__()
self.backbone = None
self.backbone_layers = None
input_channel = [48, 96, 192, 384]
cfg.MODEL_ASPP_OUTDIM = 512
self.pspaspp1 = PSPASPP(in_channels=input_channel[0],
out_channels=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
norm_layer=SynchronizedBatchNorm2d,
norm_kwargs={'momentum': cfg.TRAIN_BN_MOM})
self.pspaspp2 = PSPASPP(in_channels=input_channel[1],
out_channels=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
norm_layer=SynchronizedBatchNorm2d,
norm_kwargs={'momentum': cfg.TRAIN_BN_MOM})
self.pspaspp3 = PSPASPP(in_channels=input_channel[2],
out_channels=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
norm_layer=SynchronizedBatchNorm2d,
norm_kwargs={'momentum': cfg.TRAIN_BN_MOM})
self.pspaspp4 = PSPASPP(in_channels=input_channel[3],
out_channels=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
norm_layer=SynchronizedBatchNorm2d,
norm_kwargs={'momentum': cfg.TRAIN_BN_MOM})
self.dropout1 = nn.Dropout(0.5)
self.upsample4 = nn.UpsamplingBilinear2d(scale_factor=4)
self.cls_conv = nn.Sequential(
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_ASPP_OUTDIM,
kernel_size=1,
stride=1,
padding=0),
SynchronizedBatchNorm2d(cfg.MODEL_ASPP_OUTDIM, momentum=cfg.TRAIN_BN_MOM),
nn.ReLU(inplace=False),
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_NUM_CLASSES,
kernel_size=1,
stride=1,
padding=0)
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, SynchronizedBatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
self.backbone = get_hrnet_basebone(cfg.MODEL_NUM_CLASSES, pretrain=pretrain, combine=False)
def __init__(self, cfg, pretrain=False):
super().__init__()
from model.deeplabv3plus.deeplabv3plusEnc import Encoding, Mean
self.backbone = None
self.backbone_layers = None
input_channel = [48, 96, 192, 384]
self.aspp1 = ASPP(dim_in=input_channel[0],
dim_out=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
bn_mom=cfg.TRAIN_BN_MOM)
self.aspp2 = ASPP(dim_in=input_channel[1],
dim_out=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
bn_mom=cfg.TRAIN_BN_MOM)
self.aspp3 = ASPP(dim_in=input_channel[2],
dim_out=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
bn_mom=cfg.TRAIN_BN_MOM)
self.aspp4 = ASPP(dim_in=input_channel[3],
dim_out=cfg.MODEL_ASPP_OUTDIM,
rate=16 // cfg.MODEL_OUTPUT_STRIDE,
bn_mom=cfg.TRAIN_BN_MOM)
self.dropout1 = nn.Dropout(0.5)
self.upsample4 = nn.UpsamplingBilinear2d(scale_factor=4)
se_channel = input_channel[-1] // 4
self.se_loss = nn.Sequential(
nn.Conv2d(input_channel[-1], se_channel, 1, bias=False),
SynchronizedBatchNorm2d(se_channel, momentum=cfg.TRAIN_BN_MOM),
nn.ReLU(True),
Encoding(D=se_channel, K=32),
nn.BatchNorm1d(32),
nn.ReLU(True),
Mean(dim=1),
nn.Linear(se_channel, cfg.MODEL_NUM_CLASSES)
)
self.cls_conv = nn.Sequential(
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_ASPP_OUTDIM,
kernel_size=1,
stride=1,
padding=0),
SynchronizedBatchNorm2d(cfg.MODEL_ASPP_OUTDIM, momentum=cfg.TRAIN_BN_MOM),
nn.ReLU(inplace=False),
nn.Conv2d(
in_channels=cfg.MODEL_ASPP_OUTDIM,
out_channels=cfg.MODEL_NUM_CLASSES,
kernel_size=1,
stride=1,
padding=0)
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, SynchronizedBatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
self.backbone = get_hrnet_basebone(cfg.MODEL_NUM_CLASSES, pretrain=pretrain, combine=False)