def __init__(self, out_planes, is_training,
criterion, ohem_criterion, pretrained_model=None,
norm_layer=nn.BatchNorm2d):
super(BiSeNet, self).__init__()
self.context_path = xception39(pretrained_model, norm_layer=norm_layer)
self.business_layer = []
self.is_training = is_training
self.spatial_path = SpatialPath(3, 128, norm_layer)
conv_channel = 128
self.global_context = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnRelu(256, conv_channel, 1, 1, 0,
has_bn=True,
has_relu=True, has_bias=False, norm_layer=norm_layer)
)
# stage = [256, 128, 64]
arms = [AttentionRefinement(256, conv_channel, norm_layer),
AttentionRefinement(128, conv_channel, norm_layer)]
refines = [ConvBnRelu(conv_channel, conv_channel, 3, 1, 1,
has_bn=True, norm_layer=norm_layer,
has_relu=True, has_bias=False),
ConvBnRelu(conv_channel, conv_channel, 3, 1, 1,
has_bn=True, norm_layer=norm_layer,
has_relu=True, has_bias=False)]
if is_training:
heads = [BiSeNetHead(conv_channel, out_planes, 2,
True, norm_layer),
BiSeNetHead(conv_channel, out_planes, 1,
True, norm_layer),
BiSeNetHead(conv_channel * 2, out_planes, 1,
False, norm_layer)]
else:
heads = [None, None,
BiSeNetHead(conv_channel * 2, out_planes, 8,
False, norm_layer)]
self.ffm = FeatureFusion(conv_channel * 2, conv_channel * 2,
1, norm_layer)
self.arms = nn.ModuleList(arms)
self.refines = nn.ModuleList(refines)
self.heads = nn.ModuleList(heads)
self.business_layer.append(self.spatial_path)
self.business_layer.append(self.global_context)
self.business_layer.append(self.arms)
self.business_layer.append(self.refines)
self.business_layer.append(self.heads)
self.business_layer.append(self.ffm)
if is_training:
self.criterion = criterion
self.ohem_criterion = ohem_criterion
def __init__(self, out_planes, is_training,
criterion, pretrained_model=None,
norm_layer=nn.BatchNorm2d):
super(BiSeNet, self).__init__()
self.context_path = resnet101(pretrained_model, norm_layer=norm_layer,
bn_eps=config.bn_eps,
bn_momentum=config.bn_momentum,
deep_stem=True, stem_width=64)
self.business_layer = []
self.is_training = is_training
self.spatial_path = SpatialPath(3, 128, norm_layer)
conv_channel = 128
self.global_context = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnRelu(2048, conv_channel, 1, 1, 0,
has_bn=True,
has_relu=True, has_bias=False, norm_layer=norm_layer)
)
# stage = [512, 256, 128, 64]
arms = [AttentionRefinement(2048, conv_channel, norm_layer),
AttentionRefinement(1024, conv_channel, norm_layer)]
refines = [ConvBnRelu(conv_channel, conv_channel, 3, 1, 1,
has_bn=True, norm_layer=norm_layer,
has_relu=True, has_bias=False),
ConvBnRelu(conv_channel, conv_channel, 3, 1, 1,
has_bn=True, norm_layer=norm_layer,
has_relu=True, has_bias=False)]
heads = [BiSeNetHead(conv_channel, out_planes, 16,
True, norm_layer),
BiSeNetHead(conv_channel, out_planes, 8,
True, norm_layer),
BiSeNetHead(conv_channel * 2, out_planes, 8,
False, norm_layer)]
self.ffm = FeatureFusion(conv_channel * 2, conv_channel * 2,
1, norm_layer)
self.arms = nn.ModuleList(arms)
self.refines = nn.ModuleList(refines)
self.heads = nn.ModuleList(heads)
self.business_layer.append(self.spatial_path)
self.business_layer.append(self.global_context)
self.business_layer.append(self.arms)
self.business_layer.append(self.refines)
self.business_layer.append(self.heads)
self.business_layer.append(self.ffm)
if is_training:
self.criterion = criterion
def __init__(self, out_planes, is_training, BN2D = BatchNorm2d):
super(Network_Res101, self).__init__()
self.layers = []
self.is_training = is_training
conv_channel = 128
# use base model of resnet 101 from resnet.py
self.context = resnet101(pretrained_model=None, norm_layer=BN2D, bn_eps=config.bn_eps, bn_momentum=config.bn_momentum, deep_stem=False, stem_width=64)
self.context_refine = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
ConvBnRelu(2048,conv_channel, 1, 1, 0, has_bn=True, has_relu=True, has_bias=False, norm_layer=BN2D)
)
# ARM for ResBlock 2,3,4 of resnet output
arms = [AttentionRefinement(2048, conv_channel, norm_layer=BN2D),
AttentionRefinement(1024, conv_channel, norm_layer=BN2D),
AttentionRefinement(512, conv_channel, norm_layer=BN2D)]
# Refinement of corresponding output
refines = [
ConvBnRelu(conv_channel, conv_channel, 3, 1, 1, has_bn=True, norm_layer=BN2D, has_relu=True, has_bias=False),
ConvBnRelu(conv_channel, conv_channel, 3, 1, 1, has_bn=True, norm_layer=BN2D, has_relu=True, has_bias=False),
ConvBnRelu(conv_channel, conv_channel, 3, 1, 1, has_bn=True, norm_layer=BN2D, has_relu=True, has_bias=False)]
self.arms = nn.ModuleList(arms)
self.refines = nn.ModuleList(refines)
# Refinement on first layer of resnet output
self.res_top_refine = ConvBnRelu(256, conv_channel, 3, 1, 1, has_bn=True, norm_layer=BN2D, has_relu=True, has_bias=False)
self.ffm = FeatureFusion(conv_channel*2, conv_channel, 1, BN2D)
# classifier for final output
self.class_refine = nn.Sequential(
ConvBnRelu(conv_channel, conv_channel//2, 3, 1, 1, has_bn=True, has_relu=True, has_bias=False, norm_layer=BN2D),
nn.Conv2d(conv_channel//2, out_planes, kernel_size=1, stride=1, padding=0)
)
self.layers.append(self.context)
self.layers.append(self.class_refine)
self.layers.append(self.context_refine)
self.layers.append(self.arms)
self.layers.append(self.ffm)
self.layers.append(self.refines)
self.layers.append(self.res_top_refine)
self.loss = nn.CrossEntropyLoss(reduction='mean',ignore_index=255)