def features(self):
resnet101 = resnet.resnet101(pretrained=self._pretrained)
# list(resnet101.children()) consists of following modules
# [0] = Conv2d, [1] = BatchNorm2d, [2] = ReLU, [3] = MaxPool2d,
# [4] = Sequential(Bottleneck...), [5] = Sequential(Bottleneck...),
# [6] = Sequential(Bottleneck...), [7] = Sequential(Bottleneck...),
# [8] = AvgPool2d, [9] = Linear
features_base = list(
resnet101.children())[0:7] # give base features till res4
features_top = list(resnet101.children())[7:8] # give res5 to top
for parameters in [
feature.parameters() for i, feature in enumerate(features_base)
if i <= 4
]: # Fix features till res2
for parameter in parameters:
parameter.requires_grad = False
# features.append(nn.ConvTranspose2d(in_channels=2048, out_channels=512, kernel_size=3, stride=2, padding=1))
# features.append(nn.ReLU())
features_base = nn.Sequential(*features_base)
features_top = nn.Sequential(*features_top)
return features_base, features_top
def __init__(self,
num_classes,
fpn_features=256,
ratios=None,
scales=None,
backbone='resnet50'):
super(RetinaNet, self).__init__()
self.anchor_ratios = [0.5, 1, 2] if ratios == None else ratios
self.anchor_scales = [2**0, 2**(1.0 / 3.0), 2**(2.0 / 3.0)
] if scales == None else scales
num_anchors = len(self.anchor_ratios) * len(self.anchor_scales)
self.num_classes = num_classes
if backbone == 'resnet50':
self.backbone = resnet50()
if backbone == 'resnet101':
self.backbone = resnet101()
if backbone == 'resnet50_cbam':
self.backbone = resnet50_cbam()
if backbone == 'resnet101_cbam':
self.backbone = resnet101_cbam()
self.fpn = FPN(features=fpn_features)
self.classifier = Classifier(in_channels=fpn_features,
num_anchors=num_anchors,
num_classes=num_classes)
self.regressor = Regressor(in_channels=fpn_features,
num_anchors=num_anchors)
self.anchors = Anchor()
def __init__(self, num_classes=21, output_stride=16, pretrained=True):
super(PSPNet, self).__init__()
self.backbone = resnet101(output_stride=output_stride)
self.layer5a = PyramidPool(2048, 512, 1)
self.layer5b = PyramidPool(2048, 512, 2)
self.layer5c = PyramidPool(2048, 512, 3)
self.layer5d = PyramidPool(2048, 512, 6)
self.final_conv = nn.Sequential(
nn.Conv(4096, 512, 3, padding=1, bias=False),
nn.BatchNorm(512),
nn.ReLU(),
nn.Dropout(.3),
nn.Conv(512, num_classes, 1),
)
def __init__(self, num_classes=21, output_stride=16, pretrained=True):
super(ANNNet, self).__init__()
self.backbone = resnet101(output_stride=output_stride)
self.layer5a = PyramidPool(1)
self.layer5b = PyramidPool(3)
self.layer5c = PyramidPool(6)
self.layer5d = PyramidPool(8)
self.conv_0 = nn.Conv(2048, 2048, kernel_size=1)
self.conv_1 = nn.Conv(2048, 2048, kernel_size=1)
self.conv_2 = nn.Conv(2048, 2048, kernel_size=1)
self.final_conv = nn.Sequential(
nn.Conv(2048, 512, 3, padding=1, bias=False),
nn.BatchNorm(512),
nn.ReLU(),
nn.Dropout(.3),
nn.Conv(512, num_classes, 1),
)
def __init__(self, num_classes=21, output_stride=16):
super(OCRNet, self).__init__()
self.num_classes = num_classes
in_channels = [1024, 2048]
self.backbone = resnet101(output_stride)
self.head = OCRHead(512, num_classes)
self.get_context = nn.Sequential(
nn.Conv(in_channels[0], 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm(512), nn.ReLU(), nn.Dropout(0.05),
nn.Conv(512,
self.num_classes,
kernel_size=1,
stride=1,
padding=0,
bias=True))
self.conv_3x3 = nn.Sequential(
nn.Conv(in_channels[1], 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm(512),
nn.ReLU(),
)
def __init__(self, num_classes=21, output_stride=16):
super(DANet, self).__init__()
self.backbone = resnet101(output_stride)
self.head = DANetHead(2048, num_classes)
def __init__(self, output_stride=16, num_classes=21):
super(DeepLab, self).__init__()
self.backbone = resnet101(output_stride=output_stride)
#self.backbone = mobilenet(output_stride=output_stride)
self.aspp = ASPP(output_stride)
self.decoder = Decoder(num_classes)
